Advanced UAV edge computing ML solutions for livestock management

Advanced UAV edge computing ML solutions for livestock management Aristotelis C. Tagarakis1, Costas Davarakis2, Alexander Loos3, Christian Weigel3, Maria Theologou2, Niki Rovatsou2, Oliver Doll3, Dimitrios Kateris1, and Dionysis Bochtis1 1 Institute for Bio-Economy and Agri-Technology (IBO), Centre of Research and Technology-Hellas (CERTH), 6th km Charilaou-Thermi Rd, GR 57001 Thessaloniki, Greece 2 Nydor System Technologies, Kondilaki 7, Psychico, Greece 3 Fraunhofer Institute for Digital Media Technology (IDMT), Ilmenau, Germany Abstract The role of Unmanned Aerial Systems (UAS) in different application areas in the agricultural sector is increasing rapidly. Following the current trends, the focus of SPADE EU Project (HE 101060778) is to investigate the potential benefits of UAS contributing to multiple field operations and processes and promote sustainable digital services in the sectors of agriculture, forestry, and livestock. Part of the developments within the SPADE ecosystem involves deployment of UAV formations (single UAV, collaborating UAVs, UAV swarms) equipped with edge-computing devices (AI/ML) for direct applications in livestock management such as detecting focused risks. These developments are being evaluated in real field conditions through the dedicated pilot activities. The current work presents the preliminary results of SPADE livestock use cases undergone two trials to evaluate the performance of UAS-enabled edge-computing AI tools, for ML modelling & data assimilation. The scope is to enable the flying systems for real time monitoring of sheep in open field grazing environments. Within this framework, the SPADE Livestock setup utilized state-of-the-art object detectors and edge computing devices. The Models being investigated included Faster R-CNN, YOLO, and SSD as backbone object detection in aerial images captured by the UAS. The acquired dataset was used to train selected Tiny Object Detection algorithms. The trials included flights at varying altitudes for capturing objects at different scales and evaluate the motion blur due to high-speed at low-altitude flights. The analysis proposed the utilization of TPH-YOLOv5 model, an enhanced version of YOLOv5, where an additional prediction head is introduced to detect objects at different scales. Furthermore, the original prediction heads are replaced with Transformer Prediction Heads (TPH), which leverage a self-attention mechanism to enhance object detection capabilities. Moreover, the convolutional block attention model (CBAM) was integrated into the model to identify attention regions in scenarios with dense objects. The results showed that TPH-YOLOv5 exhibits excellent performance when applied to drone-captured scenarios and thus outperformed competing methods. Specifically, on the DETtest-challenge dataset, TPH-YOLOv5 achieved an average precision (AP) of 39.18%, surpassing the previous state-of-the-art method by 1.81%. Plans for future work include the implementation of further in-situ pilot trials, to evaluate the real-time response of the system. Keywords: Digital Agriculture, Unmanned Aerial Systems, Edge-computing, Machine learning, Livestock management 1. Introduction Some of the key applications of UAVs in livestock health care include Aerial Surveillance, Environmental Monitoring, Grazing Management, Predator Management, Herd Management and Mustering, Data Collection and Analysis, Remote Access, and Monitoring. This work addresses the utilisation of UAVs with edge computing capabilities as tools for more efficient Livestock management. This concept is enhanced by combining direct sensing on the physical livestock assets (i.e. animals) with remote sensing using UAV edge computing ML tools aiming at implementing digital twin(s) modelling. The latter enables augmenting the traditional role of UAVs, in supporting livestock health care and farming operations, by providing knowledge-full aerial surveillance, grazing management, environmental and healthcare monitoring. Moreover, our ecosystem determines interoperability between two independent DTs that are addressing two independent dataspaces corresponding to the same physical phenomenon (livestock). Thus, we are achieving the implementation of a cognitive continuum among independent dataspaces defined by direct sensing (DT1) AgEng 2024 Proceedings 612 and UAV-based remote sensing (DT2). In the literature there are object detection and tracking algorithms that can be utilised in UAV footage, to be implemented on edge devices. Such modern deep learning technologies can help farmers to optimize livestock management. In this work, a state of the art overview was done based on which such algorithms can be categorised in the following categories based on their applications: a) General Object Detection b) Object Tracking c) Edge Computing d) Tiny Object Detection on UAVs a) General Object Detection: Automatic object detection is one of the most well studied fields of research in the computer vision domain. Especially with the rise of deep learning and publicly available large annotated datasets, automatic detection and classification of objects gained huge popularity. One of the first attempts to utilize CNNs for object detection called Regions with CNN features (R-CNN) was published by Girshick et al. (2014). Several improvements of this basic algorithm such as Fast R-CNN (Girshick, 2015) and Faster R-CNN (Ren et al., 2015) were developed afterwards. As an alternative, a one-stage object detector called YOLO (You Only Look Once) was introduced by Redmon et al (2015). Compared to previous deep-learning frameworks for object detection, YOLO is extremely fast. Since then, several improvements of the original implementation of YOLO were made over the years which results in a whole family of object detection architectures and models pretrained on the COCO dataset. Another one-stage detector that eliminated the need for region proposals called SSD (Single-Shot-Detector) was presented in Liu et al. (2015). More recently, the authors of Tan et al. (2020) presented what is known as EfficientDet, a family of object detectors that focuses on achieving a balance between accuracy and efficiency. It utilizes EfficientNet (Tan et al., 2019) as backbone, a high-performance CNN which has fewer parameters and thus requires less computational resources. Another family of anchorless one-stage object detectors is CenterNet (Duan et al., 2019). Unlike other object detectors that focus on predicting bounding box coordinates or keypoints, CenterNet aims to simplify the detection by focusing on detecting object centers and relative bounding box dimensions rather than bounding box coordinates directly. Another advantage of CenterNet is that it replaces the need for Non-Maximum-Suppression (NMS) as post processing step with a more efficient algorithm which can directly be integrated into the CNN. Consequently, removing NMS as post-processing step enables much faster inference than competing methods. b) Object Tracking: Furthermore, real-time object tracking is a vital component of UAV-based applications. Traditional tracking algorithms such as Kalman Filters and Particle Filters as well as more advanced deep learning-based trackers such as DeepSORT (Wojke et al., 2017), StrongSORT (Du et al., 2023), ByteTrack (Zhang et al., 2022), and CenterTrack (Zhou et al., 2020), enable UAVs to continuously track and follow objects of interest in real-time, even in complex and dynamic environments. Continuous multi-object tracking and object association in consecutive frames builds the basis for subsequent tasks such as motion analysis for instance. By tracking objects over time, computer vision systems can analyse their motion patterns, speed, direction, and trajectories. This information is valuable for tasks such as activity recognition, behaviour analysis, and anomaly detection. c) Edge Computing: With the proliferation of edge computing and onboard processing capabilities in modern UAV systems, there is a growing demand on performing object detection and analysis directly on the UAV itself. Edge computing offers several advantages for deep learning applications: Firstly, edge computing reduces the latency and response time by performing computations directly on edge devices or embedded systems. This enables real-time decision-making which makes it well-suited for time-sensitive applications like real-time surveillance of livestock for instance. Secondly, edge computing minimizes the reliance on cloud connectivity, ensuring that deep learning models can operate even in remote or disconnected environments which is particularly beneficial in scenarios where reliable network connectivity may be limited. Thirdly, edge computing reduces the bandwidth and storage requirements by processing data locally. Deep learning models can be deployed directly on edge devices, eliminating the need to send large amounts of raw data to the cloud for processing. This results in significant cost savings and optimized resource utilization, as only relevant information needs to be transmitted. In recent years, several edge devices for deep learning applications have been developed by the industry. The NVIDIA Jetson family, such as the Jetson Nano, Jetson Xavier, or Jetson TX2, for instance, is one of the most popular edge computing platforms designed specifically for accelerating deep learning applications AgEng 2024 Proceedings 613 on the edge. An alternative to this is Google’s Coral which is a platform of hardware components equipped with Google's Edge TPU (Tensor Processing Unit) and software tools for deep learning inference on the edge. d) Tiny Object Detection on UAVs: With the astonishing capabilities of state-of-the-art object detectors and the rise of edge computing devices, also automatic object detection and tracking in UAVs has seen significant advancements in recent years. Models like Faster R-CNN, YOLO, and SSD have emerged as the backbone of object detection in aerial images, but since UAVs usually fly in high altitudes, a variety of adjustments and advancements had to be invented to accurately detect oriented tiny objects with high accuracy. A comprehensive overview of state-of-the-art object detectors especially designed for small objects and object localization in UAV images were presented by Chen et al. (2023) and Wu et al. (2022). One interesting approach to cope with objects being captured at different scales due to varying altitudes at which drones navigate as well as motion blur due to high-speed and low-altitude flight of drones was presented in Zhu et al. (2021). In the following chapters we describe the ecosystem developed within the framework of SPADE EU project focusing on the enhanced management of Livestock utilising DTs that are based on wearable technologies and ML algorithms for UAVs with edge computing capabilities. 2. Concept and Methodology The SPADE Project Livestock Pilot Addressing the needs of the Livestock sector and its community, SPADE’s (https://spade-horizon.eu/) efforts focused on the Digital Twin (DT) concept that is being standardized (ISO/IEC JTC 1/SC 41) and especially in a DT System of Systems (SoS) Virtual to Physical ecosystem (Figure 1). Figure 1. Digital Twinning of SPADE Livestock platform in line with ISO 2022 ISO/IEC JTC1/SC 27 In this approach Livestock animal-based sensing and Livestock UAV-based monitoring are being simulated by two DT interoperability cases. The DT A represents the Physical to Virtual space monitoring of Livestock data using wearable devices. The DT B represents the Physical to Virtual space monitoring of Livestock data using UAV based equipment. In DT A, physical conditions such as herd position, mobility, and deduced activity can be extracted to the virtual space, while in DT B the physical conditions of the UAV and its flight mission as well as Livestock physical conditions such as herd-flock and animal tracking, grazing landscape status monitoring, environmental and predator risk assessment, can be extracted to the virtual space. Objectives, piloting, and ecosystem prototype: This work develops a prototype platform to pilot sheep breeding and healthcare monitoring (Madesis et al., 2019), which is an essential economic activity offering hundreds of jobs. It also helps to shape the ecosystem and adds significantly to the agricultural, cultural, and gastronomic history of regions. The system evaluation is performed on the Greek island of Lesvos, where sheep breed consists of 350,000 purebred animals distributed in approximately 2.000 flocks. Almost 80% of the animals are kept in the north-western part of the island. The average milk yield is 223±70 kg in a milking period of 163±34 days. The breed’s prolificacy is 1.10 - 1.25 lambs/ewe. Therefore, this work presents a AgEng 2024 Proceedings 614 scenario case where a multipurpose drone is used for enabling livestock grazing and healthcare monitoring, in line to assisting the overall livestock management. The overall objective of this work is the optimization of both animal grazing and healthcare management, to improve milk yield, safeguard the sector’s sustainability and contribute to the authentication, tracing, monitoring and conservation of traditional production systems and sheep rearing. Moreover, the proposed technology aspires to improve the working conditions and income of farmers and entice young farmers to engage in traditional sheep rearing by offering a platform capable of providing evidence-based decision support and minimize the effort required for everyday tasks. 3. Results and discussion 3.1. Physical and health-care conditions in the grazing process (DT A), usage of wireless animal wearable sensors Motivation: For DT A we use animal wearable sensors to trace animal location and correlate animal movements (via x,y,z accelerometers) with activity/grazing and/or health status. This solution attempts to have a first level of automated tracking and detection of activity and potential healthcare issues. Focusing on grazing, we realized that eating habits may imply a first level animal healthcare status. Therefore, DT A implements a digital mapping of the physical situation as per monitoring a number of animals (the optimum would be all) via wearables. The expected outcome, based only on DT A, would be detecting fluctuations in eating patterns (e.g. lower grazing times) and thus identifying whether animals might be unhealthy, lame, or unwell. During the first SPADE project stage lightweight wearable devices were used comprising of a multisensor cellular IoT prototyping platform (Nordic Thingy:91) using LTE-M and GNSS. The cellular communication combined with the GNSS positioning systems and the multitude of sensors for motion and impact are making this solution ideal for grazing and healthcare monitoring. As depicted in Figure 2, a lightweighted device (Thingy:91) attached to sheep collars is used as a ‘wearable’. The device is a dedicated, battery-operated microcontroller equipped with low-power inertial sensors (accelerometer and gps) and wireless communication. Furthermore, the device is built around the nRF9160 System-in-Package (SiP), it is certified for a broad range of LTE bands globally and it supports LTE-M, NB-IoT and GNSS, and a nRF52840 board controller, with Bluetooth Low Energy and NFC. The position of the animal is detected with the GNSS integrated in the nRF9160 SiP while the accelerometers provide the measurements to perform animal motion analysis. The device samples regular accelerometer and tracking measurements at the frequency of 20 Hz. The SPADE Livestock platform is currently under development utilizing the OpenRemote open-source platform (OpenRemote IoT Platform). The prototype version has been customized by NST (http://www.nydorsystem.com/) for position tracking and activity status monitoring. The direct sensing part of the experiment resulted in a successful real time feed of the SPADE Livestock platform from the direct animal sensing collars. The system showed high potential for real time tracking of animals’ status as well as detection of specific animal behaviour patterns (grazing, limping, moving in cycles). Communication issues have been clarified and timestamping has been satisfactorily employed as the main reference element to establish coherence between DTA and DTB datasets. Further development is planned to enhance the combined DTA – DTB system via augmenting the timestamping reference element with animal IDs and location tracking. Preliminary results on using monitoring technologies for direct livestock healthcare and grazing management: SPADE decided to exploit a direct sensing link (DT A) using wearables directly on animals and UAVs as well as a remote sensing link (DT B) utilising remote sensing provided by UAVs with edge computing capabilities. Field trials have been conducted for testing the system in real field conditions and these will be continued throughout the next 24 months. Hitherto, monitoring real time grazing has been evaluated in conjunction with parts of DT B, where remote sensing via UAV flight sensing missions tests advanced computer vision algorithms for animal detection/tracking/identification/counting towards behavioural patterns recognition. AgEng 2024 Proceedings 615 Figure 2. The physical to digital context of SPADE Livestock in DT A: wearable device attached to animal collar (left); the digital representation on the OpenRemote cloud platform, showing both position in the herd and activity based on a three-axis acceleration (right). 3.2. Computer Vision on Edge Devices for Livestock Monitoring (DT B), usage of UAVs and ML edge computing tools Motivation: Unmanned aerial vehicles (UAVs) equipped with automated animal detection systems implemented on edge devices have emerged as promising tools for wildlife monitoring, conservation, and livestock farming. Advanced computer vision algorithms, integrated into aerial platforms, for animal detection and tracking in footage from UAVs provides researchers, conservationists, and farmers with a nonintrusive, cost-effective, and efficient way to monitor livestock, increasing animal welfare and optimising livestock management. Preliminary results on selecting a tiny objects detection ML model and training with Livestock footage acquired in SPADE: The authors propose TPH-YOLOv5, an enhanced version of YOLOv5, where an additional prediction head was introduced to detect object at different scales. Furthermore, the original prediction heads are replaced with Transformer Prediction Heads (TPH), which leverage a self-attention mechanism to enhance object detection capabilities. To identify attention regions in scenarios with dense objects the convolutional block attention model (CBAM) was integrated into the model. The authors showed that TPH-YOLOv5 exhibits excellent performance when applied to drone-captured scenarios and thus outperformed competing methods. Specifically, on the DET-test-challenge dataset, TPH-YOLOv5 achieves an average precision (AP) of 39.18%, surpassing the previous state-of-the-art method by 1.81%. Using State of The Art on Animal Detection in UAV images: Large, annotated datasets are a prerequisite to properly train and evaluate object detectors for animal detection in UAV images. To compare different approaches and their ability to detect sheep in UAV images, two datasets are considered: The first dataset we currently investigate is SheepCounter which consists of 1731 images containing only white sheep mainly on meadows (Nolan G., 2021). The second, more challenging dataset, is called Sheep Detection from Above with 2018 images (it3915masterpreparatoryproject, 2022). Compared to Nolan G. (2021), this dataset not only contains white sheep but also brown, black, and grey sheep, which makes the dataset more diverse. Additionally, terrain and illumination are more challenging. Due to the lack of available video datasets and to be able to benchmark different types of object detectors and tracking paradigms, we recorded a dataset of videos from drones showing herds of sheep on the University Farm of Aristotle University of Thessaloniki (A.U.Th.). The dataset currently contains 17 videos of different lengths, ranging from 30 seconds to over 4 minutes. We are currently annotating the dataset using the open-source Computer Vision Annotation Tool (CVAT) , an open-source web-based tool specifically designed for annotating images and videos to create training datasets for computer vision algorithms. Although, general object detection is getting not only more accurate but also more efficient, detecting animals in UAV images is still a challenging problem. Especially in precision livestock farming or monitoring of endangered species, high accuracy is demanded while smaller and more energy efficient models are needed due to implementation on mini-computers and edge devices. In relevant work (de Lima et al., 2023), the authors compared YOLOv4 and YOLOv5 models to count bovine cattle in images taken at altitudes of 20, 40, 80 and 100 m. All variants of YOLOv5 exceed a precision of 92% with the smallest model reaching a precision of 96% and the largest model 98%. In Wang et al. (2023) the newer YOLOX nano model was used and improved the detection performance for small objects. The obtained mean average precision (mAP) for cattle, sheep and horses is 86.47% at a height of 300m. In case of common cranes, it was shown in Chen et al., (2023) that the use of automatic approaches to count individuals in UAV images can be more AgEng 2024 Proceedings 616 accurate than manual counting of field observers, who underestimated the population. The applied YOLOv3 model reached a precision of 99.91% and a recall of 94.59% for RGB images at daylight. Within SPADE we already trained and compared several state-of-the-art object detectors, achieving promising results on the SheepCounter dataset mentioned above. Figure 4 shows the performance of the SSD object detector with MobileNetV2 backbone on a test image of the SheepCounter dataset. Figure 2. Detection results of SSD with MobileNetV2 backbone on an image of the SheepCounter dataset. The green boxes represent the detected bounding boxes. 4. Conclusions This work produced a first set of tangible results in line with the EU Horizon (HE) project SPADE. On the Tiny Object Detection on UAVs tasks, satisfying results were achieved especially on a self-created hyperspectral image dataset with a F1-score of 0.836. Although the achieved results are promising, the proposed model architecture is relatively complex which makes implementation on edge devices infeasible. Another interesting approach was presented by Fan and Lu (2021). The authors used a simplified AlexNet CNN architecture for landscape classification but introduced a spatial and spectral feature fusion paradigm to improve crop classification robustness, raising the accuracy for their dataset from 86.07% to 92.76%. On the process of modelling Livestock assets via Digital Twinning our first results proved several benefits for Livestock Farming: Animal Health and Welfare: Digital twinning can contribute to monitoring and improving animal health and welfare by creating virtual representations of livestock and integrating data from wearable sensors. In addition, to support the UAV flight mission, the wearables introduced in the architecture of the Livestock platform to provide to the UAV pilot the exact GPS position of the herd. Based on the conclusions from the Livestock trial, the wearables can be used to achieve a) tracking and reporting, in real time, the exact position of the animals by tracking the bell animals of the herd, b) if equipped with extra sensors like accelerometer and thermometer, the wearables can be used for the data collection aiming to train and validate AI models to detect animal’s activity (such as normal non-grazing activity, normal grazing activity, abnormal activity (i.e., illness, kneeling, moving in cycles, fever, etc.), c) testing the virtual fencing of the pasture area as well as to validate the UAV’s calculations of animals trespassing. Overall, this work presented findings from the SPADE Livestock pilot trials on the automatic detection of sheep from aerial images. This work is being improved by comparing different state-of-the-art object detectors. Although first results are promising, we further need to investigate advantages and shortcomings as well as their feasibility to be implemented on edge devices. Moreover, we plan to combine object detection with tracking algorithms, to derive semantically higher information from videos such as animal counting or behaviour analysis. Another research direction we plan to investigate is to automatically classify grassing regions which enable farmers to strategically allocate grazing areas for their livestock. Acknowledgements This work has been supported by the SPADE project, funded by the European Union’s Horizon Europe Research and Innovation programme within HORIZON-CL6-2021-GOVERNANCE-01 under Grant Agreement no. 101060778. AgEng 2024 Proceedings 617 References Nordic Thingy:91 Cellular IoT prototyping platform, [Online]. Available: https://www.nordicsemi.com/- /media/Software-and-other-downloads/Product-Briefs/Nordic-Thingy-91-PB.pdf. [Accessed June 2023]. OpenRemote IoT Platform, [Online]. Available: https://www.openremote.io/. Chen A., Jacob M., Shoshani G. and Charter M., Using Computer Vision, Image Analysis and UAVs for the Automatic Recognition and Counting of Common Cranes (Grus grus), Journal of Environmental Management, vol. 328, 2023. Chen C., Zheng Z., Xu T., Guo S., Feng S., Yao W.and Lan Y., YOLO-Based UAV Technology: A Review of the Research and Its Applications, Drones, vol. 7, no. 3, 2023. CVAT, Open Data Annotation Platform, cvat.ai, [Online]. Available: https://www.cvat.ai/. [Accessed 7 June 2023]. de Lima Weber F., de Moraes Weber V. A., de Moraes P. H., Matsubara E. T., Paiva D. M. B., de Nadai Bonin Gomes M., de Oliveira L. O. F., de Medeiros S. R. and Cagnin M. I., Counting Cattle in UAV Images Using Convolutional Neural Network, Remote Sensing Applications: Society and Environment, vol. 29, pp. 2352-9385, 2023. Du Y., Zhao Z., Song Y., Zhao Y., Su F., Gong T. and Meng H., StrongSORT: Make DeepSORT Great Again, IEEE Transactions on Multimedia, 2023. Duan K., Bai S., Xie L., Qi H., Huang Q. and Tian Q., CenterNet: Keypoint Triplets for Object Detection, in IEEE/CVF International Conference on Computer Vision (ICCV), 2019. Fan C. and Lu R., UAV image crop classification based on deep learning with spatial and spectral features, IOP Conference Series: Earth and Environmental Science, vol. 783, 2021. Girschick, R., Fast-RCNN, Proceedings of the International Conference on Computer Vision (ICCV), 2015. Girshick, R., Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. ISO/IEC JTC 1/SC 41, Internet of Things and Digital Twin, 2021. [Online]. Available: https://www.iec.ch/dyn/www/f?p=103:7:601537151799192::::FSP_ORG_ID,FSP_LANG_ID:20486,25. it3915masterpreparatoryproject, Roboflow, 2022. [Online]. Available: https://universe.roboflow.com/it3915masterpreparatoryproject/sheep-detection-fromabove/browse?queryText=&pageSize=50&startingIndex=0&browseQuery=true. [Accessed 7 June 2023]. Liu W., Anguelov D., Erhan D., Szegedy C., Reed S. E. and Fu C.-Y., SSD: Single Shot MultiBox Detector, CoRR, abs/1512.02325, 2015. Manessis G, Gelasakis AI and Bossis I., The Challenge of Introducing Point of Care Diagnostics in Farm Animal Health Management, Biomedical Journal of Scientific & Technical Research, p. 4, 2019. Nolan G., Sheepcounter Dataset, 2021. [Online]. Available: https://universe.roboflow.com/riisprivate/sheepcounter. [Accessed 7 June 2023]. Redmon J., You Only Look Once: Unified, Real-Time Object Detection, CoRR abs/1506.02640, no. http://arxiv.org/abs/1506.02640, 2015. Ren, S. He, K., Girshick R. B., and Sun J., Faster R-CNN: Towards Real-Time Oject Detection with Region Proposal Networks, CoRR, abs/1506.01497, 2015. Tan M., Pang R.and Le Q. v., EfficientDet: Scalable and Efficient Object Detection, in Computer Vision and Pattern Recognition (CVPR) , 2020. Tan M.and Le Q. V., EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks, in International Conference on Machine Learning (ICML), 2019. Wang Y., Ma L., Wang Q., Wang N. and Wang D., A Lightweight and High Accuracy Deep Learning Method for Frassland Grazing Livestock Detection Using UAV Imagery, Remote Sensing, vol. 15, no. 6, 2023. Wojke N., Bewley A. and Paulus D., Simple online and realtime tracking with a deep association metric, in IEEE International Conference on Image Processing (ICIP), 2017. Wu X., Li W., Hong D., Tao R. and Du Q., Deep Learning for Unmanned Aerial Vehicle-Based Object Detection and Tracking: A survey, IEEE Geoscience and Remote Sensing Magazine, vol. 10, no. 1, pp. 91124, 2022. AgEng 2024 Proceedings 618 Zhang Y., Sun P., Joang Y., Yu D., Weng F., Yuan Z., Luo W.and Wang X., ByteTrack: Multi-object Tracking by Associating Every Detection Box, in European Conference on Computer Vision (ECCV), 2022. Zhou X., Koltun V. and Krähenbühl P., Tracking Objects as Points, in European Conference on Computer Vision (ECCV), 2020. Zhu X., Lyu S., Wang X. and Zhao Q., TPH-YOLOv5: Improved YOLOv5 Based on Transformer Prediction Head for Object Detection on Drone-captured Scenarios, in IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), 2021. AgEng 2024 Proceedings 619