A Multidimensional Media Adaptation Framework for Live Holographic Communication

1 A Multidimensional Media Adaptation Framework for Live Holographic Communication Carl Udora, Peng Qian, Jingxuan Men and Ning Wang, Senior Member, IEEE Abstract—The rapidly increasing popularity of immersive1 multimedia services such as live holographic communication2 represents the future trend of extended reality (XR) applications.3 However, the realization of such immersive and interactive expe-4 riences is limited by the lack of fundamental understanding of5 how different user behaviours and environmental factors jointly6 affect the overall quality of experience (QoE). In particular,7 compared with the media adaptation mechanisms applied in8 conventional video applications, considerably more independent9 factors may influence user QoE in these applications, including10 both humanand network-related factors. In this paper, we11 investigate the fundamental design principles of dynamic media12 adaptation methods for live holographic communication by holis-13 tically considering these factors. Specifically, a machine learning-14 based scheme is introduced to facilitate intelligent adaptation15 of both the frame quality (resolution level) and frame rate16 according to specific contexts, such as the user intent/behaviour17 (including object motion patterns and user movements) and18 real-time network conditions. Extensive real-world experiments19 are conducted to assess the feasibility and performance of the20 proposed method, and comparisons with state-of-the-art methods21 are performed. The results indicate that the proposed approach22 can effectively satisfy user intent, with increased user QoE.23 Index Terms—User Intent, Holographic Communication, Im-24 mersive Media, Intelligent Adaptation, Quality of Experience.25 I. INTRODUCTION26 THE increased demand for immersive and interactive27 user-driven multimedia experiences has been addressed28 through the use of consumer-friendly multimedia devices29 such as head-mounted displays (HMDs) and immersive media30 applications. One such application, projected to be critical for31 multimedia services [1], [2], is holographic communication32 [3]. This application provides immersive user-driven experi-33 ences through photorealistic object representations. These rep-34 resentations are accessed through unrestricted user movements35 and object views across six degrees of freedom (6DoF) in36 extended reality (XR) space.37 The functionalities of holographic communication systems38 can be realized through existing 5G networks and edge com-39 puting facilities; however, existing systems are considered low40 level, with limited functions [4]. In particular, widespread41 utilization is limited owing to strict computational, interactive42 and network requirements [5].43 Carl Udora, Peng Qian and Jingxuan Men are with the Institute for Communication Systems, University of Surrey, GU2 7XH Guildford, United Kingdom, (E-mails: [cu00029, peng.qian, j.men]@surrey.ac.uk) Ning Wang is with the University of Bristol, BS8 1QU, Bristol, United Kingdom, (Email: [email protected]) This manuscript was edited for English language by American Journal Experts (AJE). Immersive multimedia applications provide a sense of 44 presence, copresence, and interaction, thus facilitating more 45 immersive user experiences [6]–[8]. On the other hand, it 46 can be inferred that a variety of independent factors may 47 influence user quality of experience (QoE) in a complex 48 fashion. These factors can be broadly categorized into three 49 influence factor (IF) classes: human IFs, system IFs and 50 context IFs [8], [9]. Here, environmental IFs collectively refer 51 to system and context IFs due to contextual overlaps [8]. 52 Human-oriented factors include influential human properties 53 [8], which, in holographic communication, include source 54 object behaviour and user behaviour and intent. In this con55 text, human locomotion and navigation are highly important 56 factors, as they enable more immersive experiences [10] and 57 significantly impact the QoE [11], [12]. Furthermore, source 58 object behaviours are important factors, including both static 59 and dynamic behaviours, as they are contextually influential 60 environment-oriented factors [8]. Moreover, in holographic 61 communication, the network environment is critical, as strict 62 latency requirements are implemented in such applications 63 [1], [5]. Thus, network parameters are highly important, as 64 such parameters could be associated with adverse network 65 conditions and substantially impact the user QoE [13]. It is 66 therefore essential to incorporate the aforementioned human67 oriented factors and network-oriented factors in adaptation 68 approaches for live holographic communication to further 69 improve the immersive experience, satisfy user intent and 70 enhance user QoE. 71 For the very first time, this paper presents a multidimen72 sional media adaptation framework for live holographic com73 munication to address the fundamental limitations of existing 74 schemes, which are tailored for traditional media adaptation. 75 Specifically, in the proposed framework, the uncertainties 76 associated with unpredictable user and object behaviours in 77 heterogeneous network environments are considered, as these 78 uncertainties present significant problems for immersive mul79 timedia applications such as live holographic communica80 tion. In this context, artificial intelligence (AI) technologies 81 are leveraged to consider multiple humanand environment82 oriented uncertainties. Furthermore, we introduce an AI-driven 83 user and source object behavioural and network awareness 84 strategy for immersive multimedia applications. Extensive 85 experiments involving representations of various quality, net86 work delays and packet losses are performed to generate 87 machine learning models, which are applied in the proposed 88 framework. Moreover, to our knowledge, we are the first to 89 explore multidimensional adaptation strategies, particularly in 90 relation to QoE performance metrics such as the quality of the 91 This article has been accepted for publication in IEEE Transactions on Multimedia. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TMM.2025.3623505 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: UNIVERSITY OF BRISTOL. Downloaded on December 09,2025 at 13:56:45 UTC from IEEE Xplore. Restrictions apply. 2 representations, or the resolution, and the number of frames92 per second (FPS) for immersive multimedia applications. The93 results of thorough experiments demonstrate the feasibility and94 good performance of the proposed framework compared with95 those of state-of-the-art frameworks, indicating its suitability96 to satisfy user intent while maintaining user QoE.97 Although more immersive multimedia systems are being98 used, extremely immersive and interactive holographic com-99 munication is still considered to be in the early stages of100 development, with more advancements expected in the future.101 The technical contributions of this paper can be summarized102 as follows:103 •To our best knowledge, this is the first piece of work104 in the literature to systematically elucidate how different105 human-oriented and environment-oriented factors affect106 specific QoE performance metrics such as resolution and107 frame rates. Furthermore, we explore physical user loco-108 motion and static and dynamic source object behaviours109 and evaluate their impact on various performance metrics.110 Moreover, we explore the impact of adverse network con-111 ditions on adaptive holographic communication strategies112 and show that network awareness should be considered in113 adaptation frameworks to maintain satisfactory QoE for114 the user.115 •We propose an AI-driven adaptation strategy to in-116 telligently optimize the user QoE in immersive holo-117 graphic communication applications. We introduce novel118 behavioural and network awareness facets to account119 for uncertainties in user and source object behaviour120 and adverse network conditions. Furthermore, owing to121 the agnostic design approach, the proposed method has122 widespread applicability and can be employed in holo-123 graphic communication applications that incorporate sim-124 ilar QoE performance metrics. Moreover, to our knowl-125 edge, we are the first group to propose a multidimensional126 adaptation strategy based on QoE performance metrics,127 particularly for emerging holographic communication.128 •We conducted real-world experiments to demonstrate the129 effectiveness of the proposed approach, accounting for130 user and source object behaviours and, in particular,131 adverse network conditions caused by network delays132 and packet losses. The results of thorough evaluations133 revealed the feasibility and good performance of the134 proposed method. The detailed analysis results indicate135 that the proposed method can effectively satisfy user136 intent while accounting for various uncertainties and137 maintaining the user QoE. Additionally, comparisons138 with relevant state-of-the-art methods indicate that the139 proposed approach achieves the best performance.140 II. LITERATURE REVIEW141 In immersive and volumetric multimedia applications such142 as holographic communication, crucial factors related to user143 behaviour, environmental factors and the multimedia presenta-144 tion strategy are critical for ensuring an immersive user expe-145 rience. The influence of such factors has drawn considerable146 interest, with research dedicated to elucidating the impact of147 factors such as user navigation strategies, object/environment 148 complexity, network limitations, compression performance, 149 and interactivity on the immersive experience of the user and, 150 more broadly, the QoE performance metrics of immersive and 151 holographic multimedia systems [1], [2], [14]–[17]. 152 For example, Park et al. [18] introduced an optimized rate153 utility scheme for volumetric media streaming. Moreover, 154 Van Der Hooft et al. [19] proposed an adaptive rate scheme 155 for streaming compressed point cloud objects. The proposed 156 scheme is a dynamic adaptive streaming over HTTP (DASH)- 157 compliant framework [20]. The study of Gudumasu et al. [21] 158 is similar to that of [19], who proposed a DASH-compliant 159 scheme. The proposed scheme has some unique differences, 160 with the primary content type being visual volumetric video161 based coding (V3C). Other DASH-compliant schemes have 162 been proposed in the literature, with Subramanyam et al. [22] 163 proposing a DASH-compliant user-centred adaptive stream164 ing framework for dynamic point clouds. Similarly, Farrugia 165 et al. [23] proposed a DASH-compliant adaptive streaming 166 framework for web-based 3D content. This framework was 167 tailored for textured mesh-based 3D objects, incorporating 168 compression and camera path navigation strategies and an 169 index for perceivable quality [24]. 170 Multiple studies have aimed to evaluate user behaviours 171 and QoE in adaptive streaming applications. For example, 172 Rossi et al. [12] investigated user behaviour in virtual reality 173 (VR) environments and its subsequent influence on VR-based 174 adaptive streaming systems. Wang et al. [25] proposed a QoE175 driven framework for point cloud-based adaptive streaming. 176 A tile-based adaptive approach, accounting for the user view 177 and distance via a proposed QoE model, is employed in this 178 framework. Furthermore, Minh et al. [11] conducted subjective 179 experiments to evaluate the impact of several factors on mixed 180 reality-based adaptive streaming. In addition, Subramanyam 181 et al. [26] performed subjective assessments of a user-centric 182 tiling-based adaptive framework for dynamic point cloud 183 streaming, and Van Der Hooft et al. [14] focused on objective 184 and subjective quality evaluations for adaptive point cloud 185 streaming applications. 186 Although various studies have presented adaptive streaming 187 approaches for immersive and volumetric media systems, four 188 important matters must still be addressed. First, few works 189 have considered adverse network conditions. Second, although 190 several works have considered user behaviour/motion via 191 predetermined camera paths, few studies have considered the 192 uncertainties associated with both user and object behaviours. 193 Third, few works have focused primarily on live streaming 194 scenarios. Finally, few works have explored the impact of mul195 tidimensional adaptation strategies, with most works focusing 196 only on the quality of the representations. 197 We aim to address these knowledge gaps through a holis198 tic intelligent adaptation approach for immersive multime199 dia applications such as live holographic communication, 200 accounting for the influence of user and object behavioural 201 uncertainties and adverse network conditions. Furthermore, the 202 proposed framework facilitates multidimensional adaptation. 203 Importantly, the focus of this work is not compression; rather, 204 we primarily consider user and object behaviours and adverse 205 This article has been accepted for publication in IEEE Transactions on Multimedia. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TMM.2025.3623505 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: UNIVERSITY OF BRISTOL. Downloaded on December 09,2025 at 13:56:45 UTC from IEEE Xplore. Restrictions apply. 3 network conditions.206 III. PROPOSED FRAMEWORK207 A. Motivation and Problem Formulation208 To address the knowledge gaps in the literature, we explore209 the impact of various behaviours and their associated uncer-210 tainties on user QoE. Here, we considered the user, source211 object, and dynamic network behaviours.212 Previous studies have shown that user behaviour and the213 resulting changes, which could be considered the user intent214 [27], influence user experience. This conclusion was reached215 considering the characterization of the influence of user be-216 haviour in previous works. For example, in [11], subjective217 studies were performed to evaluate the impact of user motion218 and the associated changes on the object viewing distance.219 The authors revealed that the perceived user experience is220 better at larger distances with reduced object quality. However,221 when closer objects are viewed, the object quality must be222 increased to achieve a similar perceived quality [11]. Thus,223 it is necessary to account for such behaviours to effectively224 satisfy user intent while maintaining optimal QoE.225 To account for changes in user behaviour, which in this226 scenario is based on user motion, the relative user motion227 or distance must be quantified. To achieve this, a viewer228 threshold relative to the source object position can be utilized,229 as represented by Equation (1):230 Vt=Vp−Op(1) where Vtis the viewer threshold, Vpis the viewer position231 and Opis the object position. Here, the viewer threshold is232 utilized to confirm whether the user intent has been satisfied233 according to Ict and Ift, which are the close and far proximity234 intent thresholds, respectively, as shown in Algorithm III-F1.235 On the basis of the viewer and intent thresholds, adaptations236 could be made so that the source object has higher quality237 when the viewer is closer and lower quality when the viewer238 is farther away. Thus, the user intent can be satisfied, and the239 optimal QoE can be maintained.240 Furthermore, previous studies have revealed the influence241 of the characterization of the visual content on the QoE.242 Specifically, in the case of visual media, the behaviour of243 the media object, particularly its motion, influences the QoE244 [28]. Furthermore, when considering factors beyond object245 motion, the associated frame rate also impacts perceptive246 quality. For example, [29], [30] revealed that an increased247 frame rate increases the perceptual quality for visual media248 objects/content with fast camera and background motions.249 Similarly, the techniques with reduced frame rates are suitable250 for use in the case of visual media objects or content with slow251 motion [31]. Thus, the motion of the visual media content252 must be considered to ensure an optimal user experience.253 Building on this, we explore such behaviours in the context254 of holographic communication. The association between the255 frame rate and media object motion can be reflected through256 a simplified FPS threshold, as shown in Equation (2)257 Fo=a.expk.Ob(t)(2) where Fois the FPS threshold for an object, aand kare 258 constants and Ob(t)is the object behaviour at time t. Here, 259 the object behaviour or object motion information is utilized 260 to determine the FPS threshold. In particular, faster motion 261 and a larger FPS threshold would result in a higher FPS for 262 the object, whereas a slow or static object motion would result 263 in a lower FPS for the object. 264 In addition, the network parameters and dynamic network 265 conditions notably impact user QoE [1], [2], [13]. Specifically, 266 the dynamic nature of network environments could lead to 267 adverse and erroneous network conditions, and these adverse 268 network conditions can significantly degrade user QoE [32]. 269 Moreover, the impact of such deteriorative behaviour on key 270 performance metrics such as quality and frame rate must be 271 considered. In this context, it is necessary to consider dynamic 272 network behaviours to facilitate an optimal user experience. 273 For example, a previous study [32] indicated that in the case 274 of adverse network conditions, a stable FPS is more important 275 than the object quality level. 276 To account for dynamic network conditions, we represent 277 the dynamic behaviour using a qualifier, as displayed in 278 Equation (3) 279 Ci(t) = f(Ni(t)) (3) where Ci(t)is the network impairment categorization at 280 time tand Ni(t)is the current network state for a given set of 281 parameters iat time t. Using the impairment categorization, 282 which is based on the network parameter information at a 283 given time, the changes in both QRd(t)and Fd(t), which are 284 the adaptive changes in the quality of the representations and 285 FPS at time t, can be represented using Equations (4) and (5), 286 respectively. 287 QRd(t) = f(Vt, Ci(t)) (4) Fd(t) = f(Fo, Ci(t)) (5) The aim is to optimally satisfy user intent considering the 288 various behaviours and changes expressed in Equations (2) to 289 (5). We mathematically represent this problem as shown in 290 Equation (6) [33]: 291 P=optimal(V I =f(QRd(t), Fd(t))) (6) where V I is the user intent, which depends on the 292 changes in the resolution and FPS QRd(t)and Fd(t).293 Owing to limitations in effectively obtaining a suitable 294 solution using standard numerical optimization methods 295 [33], the formulation Ppresents an issue. Similar issues 296 are encountered for Equations (2) to (5). To adequately 297 address these issues, machine learning is employed within 298 the proposed framework, enabling various behaviours and 299 their associated uncertainties to be effectively considered. 300 In addition, analyses and quantitative assessments of the 301 proposed framework are conducted using key performance 302 indicators, the FPS and resolution, as both are quantitative 303 performance metrics relevant to the user experience that also 304 reflect the capability of the proposed system to satisfy the 305 This article has been accepted for publication in IEEE Transactions on Multimedia. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TMM.2025.3623505 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: UNIVERSITY OF BRISTOL. Downloaded on December 09,2025 at 13:56:45 UTC from IEEE Xplore. Restrictions apply. 4 user intent.306 307 We aimed to explore the influence of different human-308 oriented and environment-oriented factors on the QoE perfor-309 mance metrics of the proposed framework. Fig. 1 shows the310 dependence of various performance metrics on these factors.311 The factors considered are the object motion, user behaviour312 and the dynamic nature of the network. Each factor clearly313 influences the adaptable parameters. The object behaviour314 directly influences the change in the FPS, with a lower315 FPS utilized for static objects and a higher FPS utilized for316 dynamic objects. The user behaviour is correlated with the user317 intent and directly impacts the quality of the representations,318 represented by the resolution. In this context, when the user319 has a specific intent, for example, when the user is close320 to or far from the hologram, the quality is increased or321 decreased appropriately. The change made corresponds to the322 literature, which provides information on the influence of the323 user distance and physical locomotion on the QoE. Finally,324 the dynamic network properties significantly influence the325 QoE because these properties impact both the quality of the326 representations and the number of FPS. Thus, the dynamic327 network properties are prioritized, which ensures that the net-328 work conditions that might impact other adaptation decisions329 are sufficiently considered. Specifically, this approach ensures330 that the user intent can be accurately satisfied. Furthermore,331 the proposed approach is flexible, as it considers performance332 trade-offs between the number of FPS and the quality of333 the representations, namely, the resolution. This enables more334 diverse and nuanced changes to be made, making the proposed335 approach more robust.336 Adaptation decision engine Object movement Viewer behaviour Human-oriented factors Environmentoriented factors Network dynamicity Frame rate Resolution Performance Trade-off Fig. 1. Dependency of the performance metrics on external factors B. System Overview337 We aimed to design an effective framework, and the system338 configuration depicted in Fig. 2 was utilized. This setup339 displays the configuration for live holographic communication340 over a network, which occurs via the internet. The source341 object is placed within the range of a single (RGB-D) camera342 sensor, which is utilized to capture both the colour and depth343 information of the object. The colour and depth information of344 the source object is processed in conjunction with additional345 tracking data collated at the client to prepare the frames of the346 captured object in a 3D format. Additionally, the translation347 and orientation information of the camera are available from348 its calibration process, which is used to determine the relative349 position of the object in the environment. The client generates350 frames with a point cloud format. The point cloud frames 351 consisting of both the spatial and colour information of each 352 point are then prepared for transmission and sent to the 353 server, at which point the necessary rendering and adaptation 354 processes are performed. The server also maintains frame 355 requests throughout the session to enable live holographic 356 communication. The complete 3D photorealistic hologram of 357 the object is then transmitted to the viewer equipment, such 358 as an HMD, allowing the viewer to have unrestricted 6DoF 359 views in the XR space. 360 Source object Server Public Internet RGB-D cameras Client Viewer with HMD Local processing of raw data •Frame rendering •Live adaptation Frame request Frame transmission Fig. 2. Holographic communication application platform C. Handling Viewer Behaviours 361 In live holographic communication, the user can view the 362 environment with 6DoF. This creates a problem with respect 363 to uncertainties in user behaviour, as this behaviour cannot be 364 adequately predicted because of the spontaneous nature of the 365 user. With the proposed framework, we attempt to address this 366 issue by incorporating the user intention in the design of our 367 adaptive framework. We apply an agnostic coding approach to 368 account for user behaviour. Various behaviours or intentions 369 might have different realization thresholds, which could then 370 be used to determine whether a specific behaviour or intention 371 has been carried out. This agnostic approach is implemented 372 in the proposed framework via the introduction of intention 373 thresholds that can be met on the basis of user behaviour. 374 For simplicity, we explain the proposed approach in the case 375 of one user and one source object. The user intention and 376 the physical motion of the user are considered for two main 377 reasons. First, freedom of movement is a key requirement for 378 immersive 6DoF environments, thus increasing the likelihood 379 of such behaviour being performed by users. Second, user 380 movement is important because of its impact on immersive 381 media adaptive streaming and, in particular, the QoE [11], 382 [22], [34]. Thus, user movement must be one of the fundamen383 tal behaviours considered when accounting for user intention. 384 Fig. 3 depicts the user intention based on the user movement. 385 To realize this type of user intention, we exploit the avail386 ability of real-time user location information facilitated by 387 6DoF tracking and spatial mapping, as well as the intention 388 tracking systems deployed in HMDs such as HoloLens 2 [27]. 389 This information is utilized to calculate the user threshold rel390 ative to the intention threshold. When the intention threshold 391 is met, the server and the intelligent adaptation framework are 392 informed that the user intention has been triggered, and user 393 behaviour-based adaptation is performed to satisfy the user 394 intent. 395 This article has been accepted for publication in IEEE Transactions on Multimedia. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TMM.2025.3623505 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: UNIVERSITY OF BRISTOL. Downloaded on December 09,2025 at 13:56:45 UTC from IEEE Xplore. Restrictions apply. 5 VIEWER HOLOGRAPHIC OBJECT UNRESTRICTED VIEWER MOVEMENT INTENTION THRESHOLD INTENTION THRESHOLD Fig. 3. User intention based on the user movement D. Handling Source Object Movement Patterns396 To account for the object behaviour in immersive multi-397 media applications such as live holographic communication,398 which is the focus of this paper, a machine learning approach399 is adopted. Furthermore, we strictly consider animated ob-400 jects or participants, as such objects are primarily used in401 live holographic communication applications. We attempt to402 consider object awareness and focus on two primary object403 motion modes1. These two object motion modes are static and404 dynamic object behaviours, both of which are considered in405 the design of the neural network.406 1) Design of the Object Behaviour Recognition Model:407 To adequately consider animated object behaviours, object408 behaviour recognition is assumed to be a human activity recog-409 nition (HAR) problem in our live holographic communication410 application. We utilize recurrent neural networks (RNNs) in411 our object motion mode recognition approach. Specifically, we412 use long short-term memory (LSTM) [35] neural networks to413 address this issue. We selected this type of network due to the414 suitability of LSTM networks for HAR [36]–[40].415 We design a stacked LSTM neural network for object416 activity recognition. The topology of the network is illustrated417 in Fig. 4. Here, the network is made up of two connected418 LSTM layers and an additional hidden layer. To stream-419 line the proposed framework, we use object tracking data420 rather than video data. This is facilitated via the utilization421 of Microsoft’s Azure Kinect DK camera sensor and Azure422 Kinect Body Tracking SDK, allowing for the extraction of423 skeletal tracking data. The extracted skeletal data are used to424 form “skeletal frames” that consist of key points and their425 corresponding X, Y, and Z coordinate data, similar to data426 obtainable with other body tracking sensors. Offline model427 training utilizing frame data is conducted. The class label428 vector is fed into a label encoder (LE), with reshaping and429 normalization processes then performed. The skeletal frames430 that serve as inputs are fed into the neural network, which431 contains 3 layers and 318 neurons. The ReLU activation432 function [41] is utilized, and the tanh activation function [42]433 is also employed within certain layers. The softmax activation434 function [43] is utilized in the output layer, with the Adam435 [44] optimizer and sparse categorical cross−entropy loss436 function applied. The outputs are limited to two classes at437 1In this paper, we consider only two simple modes, namely, static and dynamic modes, for simplicity and clarity. However, the proposed framework can be easily extended to support more fine-grained motion modes. Object Behaviour Recognition Recognized Behaviour LSTM Neural Network LSTM Object Tracking Data LSTM LSTM LSTM LSTM LSTM LSTM LSTM LSTM Static Object Behaviour Dynamic Object Behaviour Fig. 4. Topology of the Object activity recognition neural network this stage; static and dynamic object behaviours. The object 438 behaviours identified with the network depicted in Fig. 4 439 provide additional information for the framework, thereby 440 enhancing the performance of the multidimensional adaptation 441 strategy. 442 E. Handling Dynamic Network Conditions 443 To account for the dynamic network conditions in the pro444 posed framework, a machine learning-based approach is em445 ployed. Error-induced transmission facilitated via the Netem 446 network emulator [45] is conducted. Furthermore, various 447 quality representation metrics, FPS values, which are the 448 target values for the camera sensor, and delay and packet loss 449 values are considered in the erroneous streaming conditions 450 to adequately account for adverse network conditions and 451 their impact on user-related metrics. Moreover, this approach 452 enables effective handling of the dynamic network properties, 453 with the information obtained via dedicated network monitor454 ing [46] utilized for adaptation-related decision-making. 455 1) Design of the Network-Aware Adaptation Model: To 456 effectively perform network aware adaptation, we aimed to 457 develop an intelligent network aware adaptation model that 458 uses the aforementioned key parameters to provide multiple 459 quality representation recommendations. As such, we consider 460 this problem to be a multiclass recommendation problem. 461 To solve this problem, offline model training is performed. 462 The output vector, which contains each class value, is reshaped 463 by feeding the vector into an LE, after which the subsequent 464 output is then processed via a one-hot encoding approach. This 465 step is performed to convert the text-based label data, which 466 cannot be directly used in the neural network, to numeric 467 data. Furthermore, one-hot encoding is applied for effective 468 training [47]. The input parameters are fed into a multilayer 469 perceptron (MLP) neural network with multiple hidden layers 470 for training. The fully connected network contains six hidden 471 layers, with 280 neurons, with the four input parameters 472 projected to 128 neurons. In each hidden layer, the ReLU 473 activation function [41] is used. The activation function was 474 chosen because of its computational simplicity and suitability 475 for training feedforward networks [41], [48]. The output layer 476 has three output values, each corresponding to an output 477 parameter. The softmax activation function [43] is utilized in 478 the output layer, facilitating prediction probabilities, and the 479 output parameter with the largest probability is recommended. 480 This article has been accepted for publication in IEEE Transactions on Multimedia. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TMM.2025.3623505 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: UNIVERSITY OF BRISTOL. Downloaded on December 09,2025 at 13:56:45 UTC from IEEE Xplore. Restrictions apply. 6 Network Awareness Throughput FPS Delay Packet Loss Recommended Quality Representation Multi-Layer Perceptron Network Q3 Q1 Q2 Fig. 5. Basic topology of the network-aware model The loss function used in the network is the logarithmic loss481 function categorical cross −entropy, with the Adam [44]482 optimizer utilized for gradient-based optimization owing to its483 efficiency and simplicity. The basic topology of the neural484 network is depicted in Fig. 5.485 Fig. 5 and Table I clearly show that the trained model fo-486 cuses primarily on providing network-aware recommendations487 for adaptations based on the quality of the representations. The488 output parameters Q1, Q2 and Q3 represent varying quality489 levels or point densities. In relation to the framework and the490 overall system, these quality levels directly correspond to vary-491 ing resolution levels, with Q1, Q2 and Q3 corresponding to492 the high definition (HD), full high definition (FHD) and wide493 quad high definition (WQHD) levels, respectively. Additional494 details related to the resolution/quality levels are presented in495 Table III [4]. To facilitate this multidimensional adaptation496 process, the network aware adaptation model is implemented497 in conjunction with the impairment categorization network498 described in subsection III-E2.499 TABLE I NETWORK AWARE MODEL PARAMETERS Input Features Throughput FPS Delay Packet Loss Output Parameters Q1 Q2 Q3 2) Design of the Impairment Categorization Model :500 To facilitate the effective performance of the network aware501 adaptation model, an impairment categorization model is de-502 signed. This model is used to enable adaptation of different503 output parameters. The model proposed in Subsection III-E1504 strictly focuses on quality representation-based adaptation,505 limiting the ability of the framework to perform multidimen-506 sional adaptation, which we refer to as simultaneous quality507 representation and FPS adaptation. To address this issue,508 the impairment categorization model informs the framework509 of the top-level impairment categorization, allowing for bet-510 ter adaptation with the introduction of a second parameter,511 namely, the number of FPS. This model is very similar to512 the base network, with slight differences in the topology513 and output parameters. The model described in Subsection514 III-E1 outputs a recommended quality representation on the515 basis of the network awareness, but the model described in516 this subsection outputs an impairment category, as listed in517 Table II. The impairment level indicates the degree to which518 the normal or optimal performance of the application layer519 Network Awareness (Application Layer Impairment Categorization) Throughput FPS Delay Packet Loss Multi-Layer Perceptron Network LOW MEDIUM HIGH BAD Application Layer Impairment Level Fig. 6. Basic topology of the impairment categorization neural network is impacted when considering the severity of the erroneous 520 network conditions. For example, a low categorization implies 521 that the system performance is stable and normal, whereas 522 a bad categorization indicates that the erroneous network 523 conditions have severe impacts on performance. Fig. 6 de524 picts the basic topology of the impairment categorization 525 neural network. With the impairment categorization model 526 and quality recommendation model, the proposed framework 527 can handle user behaviours and dynamic network properties 528 effectively in the network awareness-based adaptation process. 529 The network-aware recommended quality ensures that the 530 proposed framework can effectively make optimal adaptation 531 decisions, particularly when the influence of user behaviour 532 is considered. Specifically, this approach ensures that the 533 proposed framework provides the most effective adaptation 534 decisions while attempting to satisfy user intent. Furthermore, 535 the impairment level provides additional context for the pro536 posed framework, increasing its robustness and flexibility in 537 making more fine-grained adaptation decisions to account for 538 and address unexpected and atypical scenarios. 539 TABLE II IMPAIRMENT CATEGORIZATION MODEL PARAMETERS Input Features Throughput FPS Delay Packet Loss Output Categorization Low Medium High Bad F. Proposed Intelligent Adaptation Framework 540 Building on the approaches described in Subsections III-C, 541 III-E and III-D, a novel user intention-driven network-aware 542 intelligent framework is proposed. In the proposed framework, 543 machine learning strategies are leveraged, a novel multidimen544 sional adaptation approach for holographic communication is 545 introduced, and trade-offs are optimized to effectively satisfy 546 user intentions. The shrewd handling of the various behaviours 547 ensures that the proposed framework appropriately accounts 548 for the uncertainties associated with the user and object 549 behaviours and the network environment. Furthermore, the 550 optimization of the trade-offs considering the adaptable pa551 rameters ensures that the optimal user experience is prioritized. 552 The design of the proposed framework is depicted in Fig. 7. 553 The figure displays the multifaceted awareness approach 554 used within the proposed framework. The framework accounts 555 This article has been accepted for publication in IEEE Transactions on Multimedia. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TMM.2025.3623505 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: UNIVERSITY OF BRISTOL. Downloaded on December 09,2025 at 13:56:45 UTC from IEEE Xplore. Restrictions apply. 7 Viewer behaviour monitoring Source object movement detecon AI-driven adaptaon decision-making Network condion monitoring Adapve frame request (rate and resoluon) Network Source object Viewer with HMD Fig. 7. Proposed framework for human-oriented and environment-oriented uncertainties,556 leveraging AI for intelligent adaptation-related decision-557 making. Furthermore, the proposed framework is quite robust,558 as the various awareness-related strategies are independent,559 as observed with the source object movement detection and560 user behaviour monitoring processes. Additionally, because of561 its flexibility, the proposed framework enables simultaneous562 changes in the frame rate and resolution, which facilitates563 complex decision-making. To prevent logical conflicts in de-564 cision making that may arise due to the independence of565 the awareness-related strategies, an additional paradigm, the566 awareness dependency strategy, is incorporated in the proposed567 framework.568 1) Awareness Dependency: One potential issue with the569 multidimensional awareness approach for the proposed adapta-570 tion framework is the possibility of logical conflicts in decision571 making due to inconsistent adaptation decisions. To prevent572 this issue, an awareness dependency strategy is utilized in573 the proposed framework. Here, the output of one awareness-574 related network influences the output of the other networks,575 such as the network aware and user intention aware strategies576 shown in Fig. 7. When the physical movement of the user577 is considered, adaptation decisions might be made on the578 basis of the proximity of the user to the object. However,579 the network might not be able to support significant increases580 in the quality of the object representation. To overcome581 this issue, the network-aware strategy provides the highest582 quality representations that the network can support, ensuring583 that the user intent can always be satisfied. Moreover, the584 impairment categorization model allows for greater flexibility585 in responding to adverse network conditions, thereby ensuring586 a stable user experience.587 Building on the proposed framework design shown in Fig.588 7, we leveraged the awareness dependencies to implement the589 algorithm outlined in Algorithm III-F1. Here, two conditions590 must be met. The first condition, Vp−Op>0, is necessary,591 as it is used to confirm that user movement-based behaviours592 are present. This step also enables the user threshold Vtto be593 calculated. The second condition, SessionActive == true,594 is required to ensure that the necessary inputs can be acquired.595 An animated object must be captured by the camera sensor to596 generate object tracking data Otd, and the server must have an597 active connection with both the client and the HMD. Provided598 that adverse network conditions are absent when Ci(t)== 599 Low, the algorithm satisfies the user intention by providing 600 the appropriate quality representation that can be supported by 601 the network. In the case of adverse network conditions, with 602 Ci(t)=Low, the framework remains flexible owing to the 603 impairment awareness model, allowing for changes in both the 604 quality of the representation QRd(t) = QRCand the number 605 of FPS Fd(t) = FCon the basis of the impairment level Ci(t).606 The priority of the network-aware adaptation decisions under 607 such conditions is determined on the basis of the awareness 608 dependency strategy and network-aware qualifier NAQ, which 609 is a generic parameter utilized primarily within network-driven 610 adaptation applications. With this approach, the framework at611 tempts to provide a satisfactory experience for the user. In this 612 context, a QoE-based model for holographic communication 613 [32] is employed. Furthermore, the object behaviour awareness 614 component based on the recognized activity Ob(t)facilitates 615 effective interactions and resource utilization through changes 616 in the number of FPS. 617 IV. EXPERIMENTAL ENVIRONMENT 618 To assess the feasibility and performance of the proposed 619 approach accurately, the configuration in Fig. 8 is employed. 620 The proposed framework is implemented at the server, and 621 extensive experiments are conducted. The figure shows the 622 complete experimental configuration, and an extensively mod623 ified version of the LiveScan3D platform [49] was utilized 624 for the experiments. A camera sensor placed in front of the 625 animated source object, which in this scenario is a human, is 626 used to capture relevant data required for the generation of 627 a 3D hologram. The data are captured in point cloud format, 628 with each point having unique spatial information (X, Y, and 629 Z coordinates) and colour information (RGB value ranging 630 from 0-255). The camera sensor used in the experiments was 631 Microsoft’s Azure Kinect DK. This camera was chosen as 632 it can capture the relevant depth and colour information of 633 the source object. Additionally, the camera sensor supports 634 multiple quality representations and frame rates, as detailed 635 in Table III [4]. The camera sensor, which is connected to a 636 computer, referred to as the client, facilitates tracking of the 637 source object. Additional processing is carried out at the client 638 level to generate the relevant skeletal data and point cloud 639 frames, which are continuously transmitted to the server to 640 resolve the frame request by the server. The server leverages 641 its powerful computing resources to perform the relevant man642 agement processes, and the point cloud data are processed via 643 transformation and rendering operations, with the output sent 644 to the HMD. The server plays an additional important role, 645 as it houses the proposed framework, incorporating the user, 646 source object and various network-aware intelligent adaptation 647 models. In this setup, network links with a 1 Gbps bandwidth 648 capacity are utilized to facilitate connectivity. Frame requests 649 are continuously sent to maintain the live streaming session. 650 Furthermore, a middle box is introduced in the experimental 651 configuration to enable controlled experiments to be performed 652 under adverse network conditions. The 6DoF views of the 653 3D holographic object allow unrestricted user movement; 654 This article has been accepted for publication in IEEE Transactions on Multimedia. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TMM.2025.3623505 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: UNIVERSITY OF BRISTOL. Downloaded on December 09,2025 at 13:56:45 UTC from IEEE Xplore. Restrictions apply. 8 Algorithm 1 Algorithm of the Proposed Framework Vt= Viewer Threshold Ict = Close Proximity Intent Threshold Ift = Far Proximity Intent Threshold Vp= Viewer Position Op= Object Position Ni(t)= Network aware input parameters at time t Otd = Object Tracking Data Ob(t)= Object behaviour at time t Ci(t)= Application layer impairment categorization at time t QR = Quality Representation F= FPS QRN= Network aware recommended quality representation QRC= Impairment categorization recommended quality representation FC= Impairment categorization recommended FPS QRd(t)= Quality representation adaptation decision at time t Fd(t)= FPS adaptation decision at time t NAQ= Network aware qualifier NAQth = Network aware qualifier threshold Input: Vt,Ni(t),Otd Output: QRd(t),Fd(t) Require: Vp−Op>0 Ensure: Session Active 1: while SessionActive == true do 2: if Ci(t)== Low then 3: if Vt≤Ict then 4: for Ni(t),i= 1,2, ...N do 5: QRd(t) = max(QRN) 6: end for 7: else if Ift ≤Vtthen 8: QRd(t) = min(QR) 9: end if 10: else if Ci(t)=Low then 11: if Vt≤Ict then 12: for Ni(t),i= 1,2, ...N do 13: QRd(t) = QRC 14: Fd(t) = FC 15: end for 16: else if Ift ≤Vtthen 17: QRd(t) = min(QR) 18: Fd(t) = FC 19: end if 20: end if 21: while NAQ≥NAQth do 22: for Otd at time t do 23: if Ob(t) == static then 24: Fd(t) = min(F) 25: else if Ob(t) == dynamic then 26: Fd(t) = max(F) 27: end if 28: end for 29: end while 30: end while thus, the uncertainties associated with the user and object 655 behaviours and the impact of adverse network conditions are 656 both considered in the experiments. 657 TABLE III RESOLUTION LEVEL DETAILS Resolution Level Display Pixels FPS Point Density HD 1280 x 720 5,15,30 418000 FHD 1920 x 1080 5,15,30 940000 WQHD 2560 x 1440 5,15,30 1671000 V. RESULTS 658 A. Framework Performance Analysis 659 To assess the effectiveness of the framework for supporting 660 increases in the user QoE, experiments are conducted in which 661 the user and object behaviours are considered. The network 662 condition is controlled according to specific parameter settings 663 so that the performance evaluations can be conducted under 664 different network conditions. 665 The proposed framework enables effective and rapid adapta666 tions, as it sufficiently accounts for the uncertainties associated 667 with the user and object behaviours. Fig. 9 depicts the number 668 of FPS, throughput and quality representation metrics of the 669 proposed framework for diverse viewer and object behaviours. 670 For clarity, we use resolution to represent the quality of the 671 representation, with QR1, QR2 and QR3 corresponding to HD, 672 FHD and WQHD, respectively. The resolution levels correlate 673 with the number of display pixels, with HD corresponding to 674 the worst quality, FHD corresponding to a higher quality level, 675 and WQHD corresponding to the best quality. Furthermore, 676 these resolutions indicate the object point density. Table III 677 [4] lists the various resolutions or quality representations and 678 their associated number of display pixels and relevant point 679 densities. The FPS plot in the figure displays the flexibility of 680 the framework in responding to different object behaviours. 681 The results demonstrate that the framework can adequately 682 identify multiple object behaviours and appropriately adjust 683 in response to such behaviours. This is inferred from the 684 algorithm-controlled reduction in the number of FPS, which 685 occurs when a static object behaviour is detected, and the 686 subsequent increase in the number of FPS, with the maximum 687 FPS chosen when the object displays dynamic movements. 688 The resolution results demonstrate the ability of the framework 689 to satisfy user intentions. Here, when the user approaches the 690 holographic object, the conditions associated with the user 691 intention are satisfied, and the quality or resolution of the 692 object is increased to the highest possible level in accordance 693 with the user behaviour. This adaptation is facilitated by the 694 stability of the network, with the resolution decreasing to the 695 lowest level when the user is far from the holographic object. 696 Furthermore, the FPS and resolution results highlight the 697 feasibility of the multioutput parameter adaptation approach, 698 as both the number of FPS and the resolution are adapted in 699 response to the user and object behaviours. This is evident 700 from the FPS results, which indicate that when the object 701 is static and the user is close, the lowest number of FPS 702 and highest resolution levels are chosen. In contrast, when 703 This article has been accepted for publication in IEEE Transactions on Multimedia. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TMM.2025.3623505 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: UNIVERSITY OF BRISTOL. Downloaded on December 09,2025 at 13:56:45 UTC from IEEE Xplore. Restrictions apply. 9 Server Content Production Continuous Capture Frame Generation Processing Delivery Object Tracking Server Handling Aggregation Management Delivery Object Behaviour Recognition User Intent Recognition Network aware adaptation Adaptation Request Animated Hologram User Equipment Live Viewing Interactivity Unrestricted viewer movement Client Middlebox Controlled Delay and Packet loss introduction Animated Hologram UNRESTRICTED VIEWER MOVEMENT Database Animated Object Data Handling Storage Fig. 8. Experimental configuration for the assessment of the proposed framework the user is close and the object has dynamic movements,704 the highest number of FPS and resolution are chosen. The705 same response is observed when the user is far from the706 object, with the lowest resolution chosen and the number707 of FPS dependent on the object behaviour. Furthermore, the708 throughput plot confirms the effectiveness of the proposed709 framework. In particular, the throughput metric increases when710 the user is close to the object, with the throughput levels711 also dependent on the object behaviour. Moreover, the results712 demonstrate the effectiveness of the resources utilized by the713 framework, with less bandwidth utilization for static object714 behaviours and more bandwidth utilization for dynamic object715 behaviours. Additionally, the results support the feasibility716 of the framework for multioutput adaptation with no logical717 conflicts in satisfying the user intention while considering the718 object behaviour.719 1) Analysis of the Effectiveness of the Facets Integrated720 in the Proposed Framework: To further evaluate the effec-721 tiveness of the proposed framework, additional experiments722 are conducted to investigate the importance of the integrated723 behaviour handling strategies. Fig. 10 shows the performance724 of the proposed framework without the source object be-725 haviour handling and user behaviour handling strategies. The726 figure highlights the necessity of considering the source object727 behaviour. This is reflected in the inability of the proposed728 framework without the source object behaviour handling strat-729 egy to make suitable adaptations in response to the object730 behaviour, thus reducing its performance. In particular, the731 absence of adaptive FPS changes required for the proposed732 framework without the source object behaviour handling strat-733 egy indicates that this framework cannot adequately address734 changes in object behaviour, which in this case is static.735 This conclusion is further supported by the throughput results,736 which show significant increases in the throughput in the737 static object case when the source object behaviour handling738 algorithm is not applied in the adaptation process. Thus, the739 exclusion of the source object behaviour handling algorithm740 limits the flexibility and adaptability of the proposed frame-741 work, resulting in higher throughput for the same static object742 compared with that achieved with the complete framework.743 Although user intention can be somewhat accounted for by744 including the user behaviour and network handling algorithms,745 as shown by the resolution results, the omission of the source 746 object behaviour handling component considerably degrades 747 framework performance. 748 The second set of plots in Fig. 10 illustrates the performance 749 of the framework without the user behaviour handling algo750 rithm. The results demonstrate that the inclusion of this com751 ponent is critical for ensuring the performance of the proposed 752 framework. This conclusion is reached on the basis of the 753 resolution results, which show that without this component, the 754 proposed framework cannot adapt to effectively satisfy the user 755 intent. This conclusion is further supported by the throughput 756 results, which show no increases in throughput in response to 757 the changes in the resolution. Moreover, although the inclusion 758 of the source object behaviour handling algorithm in this 759 scenario allows the proposed framework to account for the 760 source object behaviour, the performance of the proposed 761 framework remains significantly degraded. 762 Fig. 11 shows the effective performance of the proposed 763 network when the dynamic network conditions are not consid764 ered. The figure shows that the framework can still account for 765 the source object behaviour, which is dynamic in this case, and 766 provides relevant adaptations to satisfy the user intent. This is 767 inferred from the change in the number of FPS to 30 FPS, 768 as well as the increases in resolution. However, under such 769 conditions, the impact of the erroneous network conditions 770 and the requirement for the inclusion of the dynamic network 771 condition adaptation strategy is more apparent, with significant 772 decreases in the number of FPS observed. In this particular 773 scenario, the absence of the dynamic network condition adap774 tation strategy hinders the ability of the framework to make 775 suitable and effective trade-offs in the adaptation process, lim776 iting its ability to provide stable performance while satisfying 777 user intent. This is further confirmed by the resolution and 778 throughput results, which indicate that although the resolution 779 increases to the WQHD level, the throughput remains fairly 780 unstable. Thus, the source object behaviour, user behaviour 781 and dynamic network condition adaptation strategies must all 782 be included to achieve optimal framework performance and 783 effectively satisfy user intent. 784 2) Visualization of the Framework Functionality: Fig. 12 785 depicts the rendered source object before and after the res786 olution adaptation strategy based on user intent is applied. 787 This article has been accepted for publication in IEEE Transactions on Multimedia. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TMM.2025.3623505 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: UNIVERSITY OF BRISTOL. Downloaded on December 09,2025 at 13:56:45 UTC from IEEE Xplore. Restrictions apply. 16 [12] S. Rossi, C. Ozcinar, A. Smolic, and L. Toni, “Do users behave1119 similarly in vr? investigation of the user influence on the system design,”1120 ACM Transactions on Multimedia Computing, Communications, and1121 Applications (TOMM), vol. 16, no. 2, pp. 1–26, 2020.1122 [13] S. Anmulwar, N. Wang, V. S. H. Huynh, S. Bryant, J. Yang, and R. Tafa-1123 zolli, “Holosync: Frame synchronisation for multi-source holographic1124 teleportation applications,” IEEE Transactions on Multimedia, pp. 1–14,1125 2022.1126 [14] J. van der Hooft, M. T. Vega, C. Timmerer, A. C. Begen, F. De Turck,1127 and R. Schatz, “Objective and subjective qoe evaluation for adaptive1128 point cloud streaming,” in 2020 Twelfth International Conference on1129 Quality of Multimedia Experience (QoMEX), pp. 1–6, 2020.1130 [15] Y. Gao, X. Wei, J. Chen, and L. Zhou, “Toward immersive experience:1131 Evaluation for interactive network services,” IEEE Network, vol. 36,1132 no. 1, pp. 144–150, 2022.1133 [16] G. Kougioumtzidis, V. Poulkov, Z. Zaharis, and P. Lazaridis, “Qoe1134 assessment aspects for virtual reality and holographic telepresence1135 applications,” in Future Access Enablers for Ubiquitous and Intelli-1136 gent Infrastructures (D. Perakovic and L. Knapcikova, eds.), (Cham),1137 pp. 171–180, Springer International Publishing, 2022.1138 [17] I. Selinis, N. Wang, B. Da, D. Yu, and R. Tafazolli, “On the internet-1139 scale streaming of holographic-type content with assured user quality1140 of experiences,” in 2020 IFIP Networking Conference (Networking),1141 pp. 136–144, 2020.1142 [18] J. Park, P. A. Chou, and J.-N. Hwang, “Rate-utility optimized streaming1143 of volumetric media for augmented reality,” IEEE Journal on Emerging1144 and Selected Topics in Circuits and Systems, vol. 9, no. 1, pp. 149–162,1145 2019.1146 [19] J. van der Hooft, T. Wauters, F. De Turck, C. Timmerer, and H. Hellwag-1147 ner, “Towards 6dof http adaptive streaming through point cloud com-1148 pression,” in Proceedings of the 27th ACM International Conference on1149 Multimedia, MM ’19, (New York, NY, USA), p. 2405–2413, Association1150 for Computing Machinery, 2019.1151 [20] I. Sodagar, “The mpeg-dash standard for multimedia streaming over the1152 internet,” IEEE MultiMedia, vol. 18, no. 4, pp. 62–67, 2011.1153 [21] S. Gudumasu, G. Maury, A. Glasroth, and A. Hamza, “Adaptive stream-1154 ing of visual volumetric video-based coding media,” in Proceedings1155 of the 15th International Workshop on Immersive Mixed and Virtual1156 Environment Systems, MMVE ’23, (New York, NY, USA), p. 30–33,1157 Association for Computing Machinery, 2023.1158 [22] S. Subramanyam, I. Viola, A. Hanjalic, and P. Cesar, “User centered1159 adaptive streaming of dynamic point clouds with low complexity tiling,”1160 in Proceedings of the 28th ACM International Conference on Multime-1161 dia, MM ’20, (New York, NY, USA), p. 3669–3677, Association for1162 Computing Machinery, 2020.1163 [23] J.-P. Farrugia, L. Billaud, and G. Lavoue, “Adaptive streaming of 3d1164 content for web-based virtual reality: an open-source prototype including1165 several metrics and strategies,” in Proceedings of the 14th Conference1166 on ACM Multimedia Systems, MMSys ’23, (New York, NY, USA),1167 p. 430–436, Association for Computing Machinery, 2023.1168 [24] R. Mantiuk, K. J. Kim, A. G. Rempel, and W. Heidrich, “Hdr-vdp-2:1169 A calibrated visual metric for visibility and quality predictions in all1170 luminance conditions,” ACM Transactions on graphics (TOG), vol. 30,1171 no. 4, pp. 1–14, 2011.1172 [25] L. Wang, C. Li, W. Dai, S. Li, J. Zou, and H. Xiong, “Qoe-driven1173 adaptive streaming for point clouds,” IEEE Transactions on Multimedia,1174 pp. 1–1, 2022.1175 [26] S. Subramanyam, I. Viola, J. Jansen, E. Alexiou, A. Hanjalic, and1176 P. Cesar, “Subjective qoe evaluation of user-centered adaptive streaming1177 of dynamic point clouds,” in 2022 14th International Conference on1178 Quality of Multimedia Experience (QoMEX), pp. 1–6, 2022.1179 [27] P. Qian, N. Wang, C. H. Foh, C. Udora, and R. Tafazolli, “Enabling1180 user intent-based network path adaptation for live volumetric streaming,”1181 in 2024 International Federation for Information Processing (IFIP)1182 Networking Conference, 2024.1183 [28] T. Lyko, Y. Elkhatib, M. Sparks, R. Ramdhany, and N. Race, “Qoe1184 assessment for multi-video object based media,” in 2022 14th Inter-1185 national Conference on Quality of Multimedia Experience (QoMEX),1186 pp. 1–4, 2022.1187 [29] N. Van den Ende, H. de Hesselle, and L. Meesters, “Towards content-1188 aware coding: User study,” in Interactive TV: a Shared Experience1189 (P. Cesar, K. Chorianopoulos, and J. F. Jensen, eds.), (Berlin, Heidel-1190 berg), pp. 185–194, Springer Berlin Heidelberg, 2007.1191 [30] Y. Shen, Y. Liu, N. Qiao, L. Sang, and D. Yang, “Qoe-based evaluation1192 model on video streaming service quality,” in 2012 IEEE Globecom1193 Workshops, pp. 1314–1318, 2012.1194 [31] G.-M. Su, X. Su, Y. Bai, M. Wang, A. V. Vasilakos, and H. Wang, “Qoe 1195 in video streaming over wireless networks: perspectives and research 1196 challenges,” Wireless networks, vol. 22, pp. 1571–1593, 2016. 1197 [32] C. Udora, P. Qian, S. Anmulwar, A. Fernando, and N. Wang, “Quality of 1198 experience modelling and analysis for live holographic teleportation,” in 1199 2024 International Conference on Computing, Networking and Commu1200 nications (ICNC): Multimedia Computing and Communications, 2024. 1201 [33] W. Wen, J. Yan, Y. Zhang, Z. Huang, L. Liang, and Y. Jia, 1202 “Adaptive cooperative streaming of holographic video over wireless 1203 networks: A proximal policy optimization solution,” arXiv preprint 1204 arXiv:2406.08806, 2024. 1205 [34] S. Van Damme, I. Mahdi, H. K. Ravuri, J. van der Hooft, F. De Turck, 1206 and M. T. Vega, “Immersive and interactive subjective quality assess1207 ment of dynamic volumetric meshes,” 2023. 1208 [35] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural 1209 Computation, vol. 9, no. 8, pp. 1735–1780, 1997. 1210 [36] J. D. Domingo, J. G´ omez-Garc´ ıa-Bermejo, and E. Zalama, “Improving 1211 human activity recognition integrating lstm with different data sources: 1212 Features, object detection and skeleton tracking,” IEEE Access, vol. 10, 1213 pp. 68213–68230, 2022. 1214 [37] M. Ullah, H. Ullah, S. D. Khan, and F. A. Cheikh, “Stacked lstm 1215 network for human activity recognition using smartphone data,” in 2019 1216 8th European Workshop on Visual Information Processing (EUVIP),1217 pp. 175–180, 2019. 1218 [38] S. W. Pienaar and R. Malekian, “Human activity recognition using lstm1219 rnn deep neural network architecture,” in 2019 IEEE 2nd Wireless Africa 1220 Conference (WAC), pp. 1–5, 2019. 1221 [39] Y. Chen, K. Zhong, J. Zhang, Q. Sun, and X. Zhao, “Lstm networks 1222 for mobile human activity recognition,” in Proceedings of the 2016 1223 International Conference on Artificial Intelligence: Technologies and 1224 Applications, pp. 50–53, Atlantis Press, 2016/01. 1225 [40] A. A. Aljarrah and A. H. Ali, “Human activity recognition using pca and 1226 bilstm recurrent neural networks,” in 2019 2nd International Conference 1227 on Engineering Technology and its Applications (IICETA), pp. 156–160, 1228 2019. 1229 [41] I. Goodfellow, Y. Bengio, and A. Courville, Deep learning. MIT press, 1230 2016. 1231 [42] M. H. Essai Ali, A. B. Abdel-Raman, and E. A. Badry, “Developing 1232 novel activation functions based deep learning lstm for classification,” 1233 IEEE Access, vol. 10, pp. 97259–97275, 2022. 1234 [43] D. Zhu, S. Lu, M. Wang, J. Lin, and Z. Wang, “Efficient precision1235 adjustable architecture for softmax function in deep learning,” IEEE 1236 Transactions on Circuits and Systems II: Express Briefs, vol. 67, no. 12, 1237 pp. 3382–3386, 2020. 1238 [44] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” 1239 arXiv preprint arXiv:1412.6980, 2014. 1240 [45] S. Hemminger et al., “Network emulation with netem,” in Linux conf 1241 au, vol. 5, p. 2005, Citeseer, 2005. 1242 [46] G. Miranda, D. F. Macedo, and J. M. Marquez-Barja, “A qoe inference 1243 method for dash video using icmp probing,” in 2020 16th International 1244 Conference on Network and Service Management (CNSM), pp. 1–5, 1245 2020. 1246 [47] A. Y. Hussein, P. Falcarin, and A. T. Sadiq, “Enhancement performance 1247 of random forest algorithm via one hot encoding for iot ids,” Periodicals 1248 of Engineering and Natural Sciences, vol. 9, no. 3, pp. 579–591, 2021. 1249 [48] X. Glorot, A. Bordes, and Y. Bengio, “Deep sparse rectifier neural 1250 networks,” in Proceedings of the Fourteenth International Conference 1251 on Artificial Intelligence and Statistics (G. Gordon, D. Dunson, and 1252 M. Dud´ ık, eds.), vol. 15 of Proceedings of Machine Learning Research,1253 (Fort Lauderdale, FL, USA), pp. 315–323, PMLR, 11–13 Apr 2011. 1254 [49] M. Kowalski, J. Naruniec, and M. Daniluk, “Livescan3d: A fast and 1255 inexpensive 3d data acquisition system for multiple kinect v2 sensors,” 1256 in 2015 International Conference on 3D Vision, pp. 318–325, 2015. 1257 This article has been accepted for publication in IEEE Transactions on Multimedia. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TMM.2025.3623505 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: UNIVERSITY OF BRISTOL. Downloaded on December 09,2025 at 13:56:45 UTC from IEEE Xplore. Restrictions apply. 17 APPENDIX1258 A. Results1259 1) Framework Performance with Static Object Behaviour1260 Under Stable Network Conditions: Fig. 17 shows the perfor-1261 mance metrics of the proposed framework with static object1262 behaviour under stable network conditions with a 0 ms delay1263 and 0% packet loss. The figure clearly shows that the proposed1264 framework, which considers user and object behaviours, can1265 effectively perform intelligent adaptations to satisfy user in-1266 tent. Moreover, the FPS results demonstrate the effectiveness1267 of the proposed framework. Upon the detection of static object1268 behaviour, the necessary adaptation, which in this scenario1269 is reducing the number of FPS to 5, is made. Furthermore,1270 the resolution results highlight the suitability of the proposed1271 framework in satisfying user intent. In this context, the frame-1272 work satisfies user intent while also accounting for object1273 behaviour. The throughput results provide additional insight,1274 with increases in throughput corresponding to forwards user1275 movement, indicating that more data were utilized to provide1276 higher-quality representations of the object.1277 We next compared the performance of the proposed frame-1278 work with that of the anchor implementation strategy, in which1279 only user movement is considered, under the same stable net-1280 work conditions. We can infer that the anchor implementation1281 strategy can also satisfy user intent under such conditions,1282 as shown by the resolution results. The FPS results for the1283 anchor implementation strategy indicate that this method does1284 not account for the object behaviour, since a higher FPS of1285 ≈30 is utilized for the same static object. This conclusion was1286 confirmed by the throughput results, showing the increased1287 bandwidth under these conditions. Thus, the proposed method1288 provides more effective resource utilization than the anchor1289 implementation method.1290 2) Framework Performance with Dynamic Object Be-1291 haviour Under Stable Network Conditions: Fig. 18 shows the1292 performance of the proposed framework in the case of dynamic1293 object behaviour under stable network conditions with a 0 ms1294 delay and 0% packet loss. The figure results indicate that1295 the proposed framework can effectively satisfy user intent1296 via the intelligent adaptation strategies while accounting for1297 user and object behaviours under such ideal conditions. The1298 FPS results provide further insight. After the object behaviour1299 is determined, the maximum FPS supported for the corre-1300 sponding resolution is provided. Furthermore, the resolution1301 results indicate the suitability and flexibility of the proposed1302 framework in attempting to satisfy user intent. The flexibility1303 of the proposed framework is displayed by its ability to freely1304 choose between the WQHD resolution, which allows for the1305 highest quality but not the highest FPS of ≥15 FPS, and1306 the FHD resolution, which provides higher quality than the1307 HD level but lower quality than the WQHD level with a1308 higher FPS of ≈25+ in an effort to satisfy user intent. The1309 throughput results validate these findings, showing increases1310 in throughput, indicating that the proposed framework can1311 effectively satisfy user intent under such stable conditions.1312 A comparison of the performance of the proposed frame-1313 work with that of the anchor implementation strategy, which1314 accounts for only user behaviour, indicates that the anchor 1315 implementation strategy can also satisfy user intentions under 1316 such stable conditions. However, the anchor implementation 1317 strategy does not account for object behaviour. Furthermore, 1318 the FPS and resolution results indicate that the anchor imple1319 mentation strategy is not flexible, thus limiting its ability to 1320 satisfy user intent. 1321 3) Comparative Analysis: Figs. 19 and 20 display a com1322 parison of the performance of the proposed framework with 1323 that of the state-of-the-art methods described in [27] and [22] 1324 in the case of static object behaviour under stable network 1325 conditions with a 0 ms delay and 0% packet loss. The figure 1326 shows that, similar to the proposed framework under such 1327 stable network conditions, the approaches described in [27] 1328 and [22] can provide the necessary adaptations to satisfy 1329 user intent. This is evident from the resolution results, which 1330 indicate that the quality level or resolution is increased on the 1331 basis of the user motion and proximity to the object, providing 1332 a satisfactory response. This conclusion is confirmed with 1333 increases in the throughput, corresponding to the adaptation 1334 decisions made. However, although the methods described in 1335 [27] and [22] display suitable performance in satisfying user 1336 intent, both adaptation schemes fail to effectively account 1337 for the source object behaviour. This is observed from the 1338 FPS results, which show that no adaptations based on the 1339 source object behaviour are made. The changes in the FPS 1340 for the methods described in [27] and [22] correspond with 1341 the maximum FPS of 15 for the WQHD level supported by 1342 the hardware. Thus, although the methods in [27] and [22] 1343 can satisfy user intent, the proposed framework provides more 1344 intelligent decisions and increased flexibility in addressing un1345 certainties associated with source object behaviour, as reflected 1346 in its multidimensional adaptability capacity. 1347 This article has been accepted for publication in IEEE Transactions on Multimedia. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TMM.2025.3623505 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: UNIVERSITY OF BRISTOL. Downloaded on December 09,2025 at 13:56:45 UTC from IEEE Xplore. Restrictions apply. 18 0 50 100 150 200 250 300 350 Time duration (s) 0 5 10 15 20 25 30 FPS Viewer Intent with the complete proposed framework (Static , Delay=0ms, PLR=0 %) Forward Movement Backward Movement Static Motion 0 50 100 150 200 250 300 350 Time duration (s) HD FHD WQHD Resolution Viewer Intent with the complete proposed framework (Static , Delay=0ms, PLR=0 %) Forward Movement Backward Movement 0 50 100 150 200 250 300 350 Time duration (s) 0 50 100 150 200 Throughput (Mbps) Viewer Intent with the complete proposed framework (Static , Delay=0ms, PLR=0 %) Forward Movement Backward Movement 0 50 100 150 200 250 300 350 Time duration (s) 0 5 10 15 20 25 30 FPS Viewer Intent without the complete proposed framework (Static, Delay=0ms, PLR=0 %) Forward Movement Backward Movement 0 50 100 150 200 250 300 350 Time duration (s) HD WQHD Resolution Viewer Intent without the complete proposed framework (Static, Delay=0ms, PLR=0 %) Forward Movement Backward Movement 0 50 100 150 200 250 300 350 Time duration (s) 0 25 50 75 100 125 150 175 200 Throughput (Mbps) Viewer Intent without the complete proposed framework (Static, Delay=0ms, PLR=0 %) Forward Movement Backward Movement Fig. 17. Experimental results: Framework performance with static object behaviour under stable network conditions (0 ms, 0% PLR) 0 50 100 150 200 250 300 350 400 Time duration (s) 0 5 10 15 20 25 30 FPS Viewer Intent with the complete proposed framework (Dynamic , Delay=0ms, PLR=0 %) Forward Movement Backward Movement 0 50 100 150 200 250 300 350 400 Time duration (s) HD FHD WQHD Resolution Viewer Intent with the complete proposed framework (Dynamic , Delay=0ms, PLR=0%) Forward Movement Backward Movement 0 50 100 150 200 250 300 350 400 Time duration (s) 0 50 100 150 200 250 300 Throughput (Mbps) Viewer Intent with the complete proposed framework (Dynamic , Delay=0ms, PLR=0 %) Forward Movement Backward Movement 0 50 100 150 200 250 300 Time duration (s) 0 5 10 15 20 25 30 FPS Viewer Intent without the complete proposed framework (Dynamic, Delay=0ms, PLR=0 %) Forward Movement Backward Movement 0 50 100 150 200 250 300 Time duration (s) HD WQHD Resolution Viewer Intent without the complete proposed framework (Dynamic, Delay=0ms, PLR=0%) Forward Movement Backward Movement 0 50 100 150 200 250 300 Time duration (s) 0 50 100 150 200 Throughput (Mbps) Viewer Intent without the complete proposed framework (Dynamic, Delay=0ms, PLR=0 %) Forward Movement Backward Movement Fig. 18. Experimental results: Framework performance with dynamic object behaviour under stable network conditions (0 ms, 0% PLR) This article has been accepted for publication in IEEE Transactions on Multimedia. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TMM.2025.3623505 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: UNIVERSITY OF BRISTOL. Downloaded on December 09,2025 at 13:56:45 UTC from IEEE Xplore. Restrictions apply. 19 0 50 100 150 200 250 Time duration (s) 0 5 10 15 20 25 30 FPS Viewer Intent with the complete proposed framework (Static , Delay=0ms, PLR=0 %) Forward Movement Backward Movement Static Motion 0 50 100 150 200 250 Time duration (s) HD WQHD Resolution Viewer Intent with the complete proposed framework (Static , Delay=0ms, PLR=0 %) Forward Movement Backward Movement 0 50 100 150 200 250 Time duration (s) 0 25 50 75 100 125 150 175 200 Throughput(Mbps) Viewer Intent with the complete proposed framework (Static , Delay=0ms, PLR=0 %) Forward Movement Backward Movement 0 50 100 150 200 250 300 350 Time duration (s) 0 5 10 15 20 25 30 FPS Adaptation with reference [27] (Static , Delay=0ms, PLR=0 %) Forward Movement Backward Movement 0 50 100 150 200 250 300 350 Time duration (s ) HD WQHD Resolution Adaptation with reference [27] (Static , Delay=0ms, PLR=0% ) Forward Movement Backward Movement 0 50 100 150 200 250 300 350 Time duration (s ) 0 50 100 150 200 250 Throughput(Mbps) Adaptation with reference [27] (Static , Delay=0ms, PLR=0%) Forward Movement Backward Movement Fig. 19. Experimental results: Comparison of the performance of the proposed framework with that of state-of-the-art methods (Ref [27]) in the case of static object behaviour (0 ms, 0% PLR) 0 50 100 150 200 250 Time duration (s) 0 5 10 15 20 25 30 FPS Viewer Intent with the complete proposed framework (Static , Delay=0ms, PLR=0 %) Forward Movement Backward Movement Static Motion 0 50 100 150 200 250 Time duration (s) HD WQHD Resolution Viewer Intent with the complete proposed framework (Static , Delay=0ms, PLR=0 %) Forward Movement Backward Movement 0 50 100 150 200 250 Time duration (s) 0 25 50 75 100 125 150 175 200 Throughput(Mbps) Viewer Intent with the complete proposed framework (Static , Delay=0ms, PLR=0 %) Forward Movement Backward Movement 0 50 100 150 Time duration (s) 0 5 10 15 20 25 30 FPS Adaptation with reference [22] (Static , Delay=0ms, PLR=0 %) Forward Movement Backward Movement 0 50 100 150 Time duration (s) HD WQHD Resolution Adaptation with reference [22] (Static , Delay=0ms, PLR=0 %) Forward Movement Backward Movement 0 50 100 150 Time duration (s) 0 50 100 150 200 Throughput(Mbps) Adaptation with reference [22] (Static , Delay=0ms, PLR=0 %) Forward Movement Backward Movement Fig. 20. Experimental results: Comparison of the performance of the proposed framework with that of state-of-the-art methods (Ref [22]) in the case of static object behaviour (0 ms, 0% PLR) This article has been accepted for publication in IEEE Transactions on Multimedia. This is the author's version which has not been fully edited and content may change prior to final publication. Citation information: DOI 10.1109/TMM.2025.3623505 © 2025 IEEE. All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: UNIVERSITY OF BRISTOL. Downloaded on December 09,2025 at 13:56:45 UTC from IEEE Xplore. Restrictions apply.