Enhancing Safety in Industry 5.0: Human-Computer Collaboration Benefits through a Dataset of Protective Equipment Detection

Enhancing Safety in Industry 5.0: Human-Computer Collaboration Benefits through a Dataset of Protective Equipment Detection Alexandros Petropoulos∗, Georgia Apostolou∗, Georgios Tsoumplekas†, George Tziolas‡ Nikolaos Ntampakis†, Ilias Siniosoglou†, Vasileios Argyriou§ Panagiotis Sarigiannidis¶, Ilias Gialampoukidis∗, Stefanos Vrochidis∗ ∗ITI, CERTH, Thessaloniki, Greece Emails: {alpetrop, gapostolou, heliasgj, stefanos}@iti.gr †METAMIND INNOVATIONS P.C., Kozani, Greece Emails: {gtsoumplekas, nntampakis, isiniosoglou}@metamind.gr ‡Sidroco Holdings Ltd, Nicosia, Cyprus Email: [email protected] §Kingston University London, London, United Kingdom Email: v[email protected] ¶University of Western Macedonia, Kozani, Greece Email: [email protected] Abstract—This paper presents the development and implementation of a comprehensive dataset designed to enhance workplace safety in industrial environments through advanced computer vision technologies. The dataset focuses on the detection of essential protective equipment, such as helmets and vests, worn by workers in various industrial settings. Utilizing this dataset, a YOLOv8-based computer vision model is trained to achieve 82.1% mAP accuracy (70.5% for helmets, 93.7% for vests) in real-time identification of whether workers are equipped with the appropriate safety gear, demonstrating high reliability for safety compliance monitoring. This initiative is part of the European Research Project TALON, which aims to demonstrate the potential of collaborative efforts between humans and machines in achieving higher safety standards, through its 4th pilot. TALON system will enable automated, flexible, adaptable, programmable, explainable and energy-efficient edge Artificial Intelligence (AI) networking by developing complementary technologies such as AI orchestrator, blockchain, edge networking and digital twins (DTs) in an integrated and innovative way. The dataset, along with the developed predictive model, offers a significant contribution to the field of safety, showcasing how technological advancements can be leveraged to safeguard human lives in the workplace. Index Terms—protective equipment, safety, computer vision, industry, human-robot collaboration, deep learning, CNN I. INTRODUCTION In industrial environments, ensuring the safety of workers is paramount. Protective equipment, such as helmets and vests, plays a critical role in mitigating the risks associated with workplace hazards. However, manual monitoring of compliance with safety regulations can be labor-intensive and prone to human error. To address this challenge, we propose the development of a dataset tailored for the detection of protective equipment using computer vision technologies. This paper introduces the dataset creation process, which encompasses the collection, annotation, and validation of images depicting workers in various industrial scenarios. The dataset is designed to train a computer vision model capable of accurately identifying whether workers are wearing the required protective gear. This initiative is undertaken as part of the European research project TALON [1], which aims to showcase the synergy between human workers and machine learning systems in enhancing workplace safety. The TALON project highlights the importance of collaborative efforts between humans and machines to achieve greater safety standards. The power of machine learning and computer vision is leveraged to create a safer work environment and promote a culture of safety compliance. The dataset and the predictive model developed through this project represent a significant advancement in the field of safety, demonstrating how technological innovation can be harnessed to protect human lives in the workplace. The main objectives of this work are: •Create a dataset and fill the gap that exists in already existing datasets regarding changing lighting conditions and occlusions. An additional feature that differentiates it from the other corresponding datasets is that the images are captured by a drone, providing more diversity. •Train a state-of-the-art deep learning model with the above dataset. •Present a scenario in which the dataset and the model will be used within the TALON project. II. RELATED WORK There are several datasets on protective equipment, most of which focus on the class ’helmet’. However, other datasets with more classes have been created recently. The SHD dataset [2] comprises 5000 images featuring three object classes—helmet, head, and person. However, a significant portion of these images is partially labeled. SHEL5K [3] dataset on the other hand, introduced an enhanced version of the SHD dataset, containing a total of 75,570 labels. The hardhat dataset [4] is a safety helmet dataset provided by Northeastern University, comprising 7,063 labeled images. It consists of three distinct classes with about 27,000 labeled objects: helmets, heads, and persons. However, in the dataset, the ”person” class is not labeled accurately and there is an unequal distribution of images across the classes. Pictor-v3 dataset [5] focuses on the classes worker, hat, and vest. Images are collected through crowd-sourcing from construction projects and web-mining using search engines. Human annotators used LabelBox [6] and LabelMe [7] tools for annotation to ensure accuracy. The available crowd-sourced subset contains 775 images containing workers. Site Object Detection Dataset (SODA) [8], introduced a large-scale image dataset specifically designed for object detection in construction sites, containing 15 different object classes. The dataset was developed to address the need for large-scale, annotated images in construction sites, which are crucial for training and evaluating object detection algorithms SHWD [9] offers a comprehensive dataset specifically for safety helmet usage and human head detection. It comprises 7,581 images featuring 9,044 instances where safety helmets are worn (positive) and 111,514 instances of normal heads (negative). The positive examples were sourced from Google and Baidu, and were manually annotated using LabelImg [10]. Finally, a significant contribution to the field was made by [11] with their SH17 dataset, which focused on human safety and personal protective equipment detection. The dataset contains over 17,000 annotated images capturing various scenes in manufacturing environments, with bounding box annotations for different types of PPE, such as hard hats, safety vests, and safety glasses. However, as noted in their paper, challenges remained in terms of detecting PPE in varying lighting conditions and occlusions. Despite these developments, there are still several challenges in the field of personal protective equipment detection datasets, such as: the different lighting conditions, the existence of occlusions and the limited availability of drone imagery for use in corresponding scenarios. Our work aims to address these limitations by introducing a comprehensive dataset and exploiting it in real-world conditions within the TALON project. III. DATASET CREATION AND CHARACTERISTICS A. Introduction to the Dataset This dataset was created as part of the TALON project to support the 4th pilot. This pilot aims to expand the level of automation in the increasingly complex manufacturing landscape through human-robot collaboration. For the needs of the pilot, a warehouse was used where, depending on the scenario, it is necessary to detect in different areas of the warehouse (either indoor or outdoor) whether workers are wearing personal protective equipment such as helmets and vests. Before the use of TALON, the monitoring of the premises was carried out by a person on patrol. However, now the patrols are done by drones and the TALON system runs deep learning algorithms where it detects if workers are wearing the required protective equipment, while raising an alert on the platform’s dashboard. The goal is to create a diverse, well-labeled dataset that represents real-world scenarios, providing the foundation for training computer vision models that can be deployed for safety monitoring in industrial and construction settings. B. Data Collection Process The creation of the dataset started with the planning of the shoots that needed to be done to offer as much diversity as possible. Hence, based on the available space where the workers move, various scenarios were designed according to which the drones would fly. Of the 5 available spaces, 3 were indoors, 1 was outdoor and the last was indoors, but with direct access outside for trucks to enter/exit. The filming was done during the warehouse opening hours with the participation of the staff and the team conducting the measurements. Staff followed their daily working routine with no interruptions, to get a realistic result. Therefore to achieve diversity : 1) multiple different drone routes were made in all areas 2) in each run, parameters such as the height of the drone were changed, from 3 - 7 meters indoors and up to 20 meters outdoors 3) the angle of the camera of the drone was changing from 45 to 90 degrees 4) to have different lighting conditions, the shots were taken at different times of the day, both in the morning and in the afternoon, when the sun was less. In addition, where available, the artificial light flickered. 5) there were different numbers of people in each scene and each time some wore protective gear and some did not. Workers are captured in standing, sitting, walking, and operating machinery poses, simulating real-life scenarios where different body orientations may affect the visibility of PPE. 6) the dataset also contains images with varying background clutter and complex scenes (e.g., workers in groups, occlusions from machinery or other workers) to simulate real-world challenges for PPE detection systems. 7) for the needs of the scenario helmets and safety vests were used as this was the equipment worn by the workers in the warehouse A sample of the images taken from the data set under different conditions is shown in Figure 1. C. Dataset Characteristics 1) Image Size and Resolution: All images in the dataset are in full HD (1920x1080) resolution. A better resolution than Fig. 1. Sample images from the dataset usual was chosen, because the drone was at some distance and information on both personal protective equipment and ambient lighting conditions needed to be better captured. In addition, starting with high-resolution images provides the flexibility to experiment with different resolutions during preprocessing without losing the original quality. 2) Labeling and Annotations: To annotate the dataset, the drone videos were converted into frames, keeping the most important ones. The open source-tool cvat [12] was used, with 2 people doing the annotation and a third person doing the final evaluation at the end. The defined classes are helmet and vest, as these are the most common means of protection in the warehouse. Next, special attention was given to labeling occluded or partially visible PPE, as these are common in realworld scenarios. Finally, the dataset was exported to YOLOv1 format, ensuring compatibility with widely used computer vision frameworks. 3) Class Imbalance: The graph in Figure 2, illustrates the distribution of classes within our dataset. As evident, there is a notable imbalance, with a certain class being significantly underrepresented compared to another. Given that protective gear might not always be visible or might be worn in different configurations (e.g., helmets alone, helmets and vests), we observed that the ’helmet’ class has about 1750 snapshots, while the ’vest’ class has only 500. 4) Privacy and Ethical Considerations: While the dataset is designed for industrial safety, privacy and ethical concerns regarding the inclusion of worker images were carefully considered. All workers appearing in the dataset have given explicit consent for their images to be used for research and training purposes. Additionally, any identifying information such as faces or logos on clothing were either obscured or excluded to protect privacy. 5) Dataset Splitting and Evaluation: To evaluate the performance of deep learning models trained on this dataset, it was divided into training, validation, and test sets. The split was as follows: Training Set: 75%, Validation Set: 15% and Test Set: 10%. Additionally, it was ensured that the splits were balanced in terms of the distribution of workers wearing full PPE versus those without, as well as the various environmental conditions. Fig. 2. Class imbalance diagram. 6) Challenges and Limitations: While the dataset has been designed to be diverse, there are still several challenges and limitations: •Partial Occlusion: In some scenarios, PPE might be partially covered by other objects or workers, making it difficult for models to correctly detect the equipment. •Image Quality: Variability in shooting conditions, such as different lighting conditions, artifacts, noise and reflections, can significantly affect the consistency, clarity and overall quality of images. Despite these challenges, this dataset serves as a comprehensive resource for training models that can identify PPE and ensure worker safety across a variety of environments and scenarios. IV. PPE DETECTION WITH YOLOV8 Following the introduction of the proposed dataset, the next step is to evaluate how effectively it can be used to train an object detector aimed at PPE detection in industrial settings. Automating PPE detection could enhance industrial workers’ safety by allowing for broader coverage of industrial workspaces through drone imagery. At the same time, it could reduce the possibility of errors due to human factors by using precise object detection models. In this section, we analyze how the proposed dataset can be utilized to train such an effective object detection model for PPE detection and assess its effectiveness through a series of evaluations. In the following subsections, first, a description of YOLOv8n is provided, a lightweight model that enables realtime object detection, which was utilized as the examined object detector in this study. Next, the experimental setting in which the model was trained, as well as the metrics used for its performance evaluation are outlined. Finally, the results obtained are presented, demonstrating the dataset’s suitability for training an effective PPE detection model that can be easily deployed in real-world applications. A. Model Description Over the years, various object detection models have been developed focusing on both general object detection [13] as well as application-specific cases, such as PPE detection [14]. Generally, the most common object detection models can be categorized as one-stage or two-stage detectors. One-stage detectors utilize anchors [15] or an initial set of potential object centers [16], [17] to make predictions. On the other hand, two-stage detectors generate proposals that are subsequently refined to obtain the final predictions [13]. Finally, in recent years, various novel approaches based on the transformer architecture [18] have been proposed that eliminate the need for initial guesses in the form of anchors or proposals, such as DETR [19] and DINO [20]. An important consideration for the examined use case is that the selected model should be able to produce real-time predictions, enabling its application in real-world environments that involve a continuous stream of input, such as video. At the same time, the model should also be efficient in terms of its computational requirements during training and inference, allowing its deployment in potentially resource-constrained devices. While two-stage detectors, such as Faster-RCNN [13], have demonstrated strong performance on various benchmarks and applications, the additional overhead introduced by their Region Proposal Network renders them less effective for realtime object detection. On the other hand, modern transformerbased object detection models alleviate the need for proposal generation. However, they typically leverage larger backbone architectures [19] that require significant computational power, even during inference, making them less appropriate for deployment in environments with restricted computational resources. Since the goal is to develop an object detection model suitable for real-time object detection in real-world industrial applications, we opt for YOLOv8n [21], a well-established one-stage object detection model that combines computational efficiency with real-time object detection capabilities. YOLOv8’s architecture, renowned for its balanced accuracyspeed trade-off, has been validated in industrial real-time applications [22], making it ideal for our safety monitoring context. In particular, YOLOv8n is the smallest variant of the YOLOv8 model family, consisting of approximately 3.2 million parameters. Building upon its predecessors in the You-Only-Look-Once [15] (YOLO) model family, the model leverages the CSPDarknet53 [23] network as its backbone feature extractor, which follows the Feature Pyramid Network structure introduced in [24], that enables the extraction of feature maps at different scales. Consequently, this feature allows for multi-scale object detection, which is particularly useful in the context of the examined dataset, where objects such as vests may be larger compared to smaller objects like helmets, especially in images taken from a significant distance. The second major component of YOLOv8 is its detection head, which processes the multi-scale feature maps generated by the backbone feature extractor using a series of convolutional layers. The resulting representations are subsequently utilized by three detection modules, each operating at different scales, which are responsible for producing the final classification scores and bounding boxes for the detected objects. As a final note, it is worth noting that within the context of the examined use case in the TALON project, the developed model is deployed in the cloud, which typically offers greater flexibility in terms of the available computational resources. However, due to YOLOv8n’s lightweight architecture and its one-stage detection method based on anchor generation, it can also be effectively deployed on resource-constrained environments, such as edge devices, with limited resource availability. This is a crucial requirement in most industrial applications, where there is a need for both performance in terms of object classification and localization accuracy, as well as reduced computational and energy needs [25]. B. Experimental Setting The proposed dataset contains two classes, namely helmet and vest, both of which are standard PPEs typically encountered in industrial sites. However, a critical factor in such use cases is the ability to detect the absence of these objects to ensure compliance with regulations that mandate the use of such PPE and potentially help reduce injuries in the event of accidents. As a result, we aim to enhance the selected YOLOv8n object detection model with this capability following a two-stage finetuning approach. Specifically, starting from a YOLOv8n model pretrained on MS-COCO [26], the first stage involves finetuning the entire model using the Worker-Safety dataset [27]. In particular, the dataset consists of five classes: person, helmet, no-helmet, vest, and vest, allowing us to detect both the existence or absence of critical PPE in industrial settings. We utilized 80% of the available data as the training set for this finetuning stage, while the remaining 20% served as our test set. Additionally, we follow a stratified approach during train-test splitting to ensure that class proportions are maintained across the two sets. As for the selected hyperparameters, we opted for the default values provided by the Ultralytics [28] framework, given their effectiveness in training a model that can successfully generalize within the specific dataset and no further improvement was achieved by using different hyperparameter values in our preliminary experimentation. Since both helmet and vest objects, along with their corresponding labels, were included in the Worker-Safety dataset, it is possible to immediately use the developed model for inference in the proposed dataset. However, since the proposed dataset is focused on a more realistic setting where images are captured by a drone from a distance, potential issues could arise due to distributional or domain shifts between the Worker-Safety dataset’s training set and the proposed dataset’s evaluation set. To address any performance degradation that might result from this distributional shift, we introduce a second finetuning stage where the model is further finetuned using the training set of the proposed dataset. To avoid catastrophic forgetting [29] of classes for which labels are not available in the proposed dataset, specifically person, nohelmet, and no-vest, we only finetune the three detection modules while keeping the rest of the model frozen. Similar to the first finetuning stage, we utilize the default hyperparameter values provided by the Ultralytics framework. For both datasets, we evaluate model performance using Mean Average Precision (mAP), which is defined as follows: mAP (%) = 1 N N X i=1 APi(1) where APiis the Average Precision (%) obtained for the i-th class, given Nclasses using an Intersection over Union (IoU) threshold of 50%. Additionally, we report Precision and Recall values in the test sets of both datasets, as well as their corresponding Precision-Recall curves and normalized confusion matrices. C. Experimental Results The developed YOLOv8n model is evaluated on the test sets of both datasets used during the finetuning stages after completing both stages. The following subsections present and analyze the results obtained for both datasets. TABLE I MODEL PERFORMANCE IN THE WORKERSAFETY DATASET Class Precision(%) Recall(%) mAP(%) person 97.9 91.7 99.3 helmet 98.7 98.0 99.3 no-helmet 95.8 100.0 99.5 vest 97.8 94.5 97.9 no-vest 99.5 96.0 99.2 all 97.9 96.0 99.0 1) Worker-Safety Dataset: Table I presents the results for each class, as well as aggregated results for all classes on the Worker-Safety dataset. Notably, the model achieves an overall mAP of 99.0% with the Average Precision for each class exceeding 99%, except for the vest class, which has a slightly lower Average Precision of 97.9%, demonstrating the proposed model’s effectiveness in detecting critical PPE with minimal errors. Additionally, the model exhibits significantly high precision and recall values, indicating that it produces only a very small number of false positive and false negative predictions across all examined classes. This is further corroborated in Figure 3, which illustrates the Precision-Recall curve for this dataset, showing that the model maintains very high scores for both precision and recall without suffering from any precision-recall tradeoffs that are typically encountered in deep learning models. Figure 4 presents a normalized version of the confusion matrix obtained for the model’s predictions, demonstrating that the model achieves remarkably high detection performance for Fig. 3. Precision-Recall curve for the Worker-Safety dataset. Fig. 4. Normalized confusion matrix for the Worker-Safety dataset. all classes of interest. It is also worth noting that the model makes minimal errors of misclassifying an object as belonging to a different class, with most errors arising from the model failing to detect the object and instead considering it part of the image’s background. Overall, it is evident that while the model is evaluated on the test set of the dataset used in the first stage after the second finetuning stage is completed, it manages to maintain its high performance. This result indicates that the developed model does not suffer from catastrophic forgetting, which can be attributed to the fact that during the second finetuning stage, only the detection modules of the model were finetuned, leading to only minimal changes in the representations learned during the first stage. 2) Proposed Dataset: Table II presents the overall results and the results per class obtained in the proposed dataset using the model after the two finetuning stages. The overall mAP of 82.1% is lower compared to the 99.0% value in the Worker- TABLE II MODEL PERFORMANCE IN THE PROPOSED DATASET Class Precision(%) Recall(%) mAP(%) helmet 80.5 71.6 70.5 vest 96.5 88.0 93.7 all 88.5 79.8 82.1 Safety dataset, which could be attributed to the proposed dataset’s more challenging and realistic setting, where objects appear significantly smaller because the images are taken from a greater distance, making detection more difficult. The performance discrepancy in the two datasets is more evident for helmet detection, where a mAP of 70.7% is achieved in the proposed dataset. Fig. 5. Precision-Recall curve for the proposed dataset. Regarding performance discrepancies between helmet and vest detection, it is generally expected that performance for helmets is going to be lower compared to vests (93.7% in the proposed dataset) due to their smaller size, which makes them more challenging to detect. Figure 5 further illustrates this discrepancy by presenting the Precision-Recall curves for each class. For vest detection, the model maintains strong performance for both precision and recall, while for the more challenging helmet class there is a clear tradeoff between these two metrics, particularly when a recall is set above 0.8, resulting in a significant drop in precision. However, lowering the confidence detection threshold during inference could help strike a better balance between false positive and false negative errors for the helmet class. Figure 6 presents the normalized version of the confusion matrix for the model’s predictions on the proposed dataset. In accordance with our previous findings, the model’s performance is lower compared to the Worker-Safety dataset, which can be attributed to the more challenging setting introduced in the proposed dataset. Once again, this discrepancy is more evident in the detection of helmets due to their significantly Fig. 6. Normalized confusion matrix for the proposed dataset. smaller size. It is also worth noting that, unlike the WorkerSafety dataset, in the proposed dataset, a significant number of the misclassification errors occur during background and foreground classification where elements in the background are incorrectly identified as helmets. Overall, although the performance measures in this dataset are lower, the results obtained for both classes can be considered highly competitive, especially given the challenging evaluation setting, underscoring the suitability of the dataset for training object detection models that can be effectively deployed in real-world industrial applications. Finally, it is important to note that even though the model can predict all five classes introduced in the Worker-Safety dataset, we only report evaluation metrics for helmets and vests in the proposed dataset due to the absence of labels for the rest of the classes. V. SCENARIO This section introduces the TALON project, its objectives, and benefits. It explains how the data set and prediction model will be integrated and used in a real-world scenario, demonstrating the system’s benefits under real conditions. A. Introduction to TALON project and use case scenario TALON addresses Industry 5.0 challenges by combining edge-cloud AI for intelligent automation. Its 4th use case focuses on human-robot collaboration in manufacturing, deploying drones and computer vision to enhance safety in warehouses where workers must wear helmets/vests. Current safety checks rely on manual inspections, risking errors. TALON automates monitoring using drones and realtime deep learning to detect non-compliance, such as missing safety gear, and trigger alerts. This reduces response times, ensures faster hazard identification, validates safety protocols, and secures the workplace through efficient human-robot collaboration. B. Overview of the scenario pipeline Thus, this scenario demonstrates integrating deep learning and edge computing for real-time worker safety monitoring. As a result, drone devices are deployed in a warehouse setting to monitor workers’ compliance with protective equipment (PPE) standards. The data collected by these drones is into a YOLOv8 [22] model that detects whether workers are wearing the required protective gear, such as helmets and vests. The system achieves 30 FPS processing with sub-100ms latency during drone-based deployment, enabling immediate safety interventions. Additionally, privacy concerns are addressed through face anonymization, and the results are visualized in a dashboard user interface (UI) that provides actionable safety insights. Figure 7 shows a schematic diagram of the scenario. Fig. 7. Diagram for the UC4 scenario As shown in the diagram, the TALON components that participate in the scenario are divided into 2 main categories, those located at the edge and those located in the cloud. On the edge are the drones, while in the cloud are the main components that will be used, which are the prediction model, the anonymisation tool and the dashboard. All these systems communicate with each other and are organised in a unified way under the TALON umbrella. C. Step by step analysis This section presents the scenario step by step, along with pictures of the results after running it. Initially, the drone is piloted into the warehouse, equipped with high-definition cameras to capture live video footage of the workplace. These videos are subsequently fed into the YOLO detector, as detailed in the previous section. Figure 8 depicts the drone’s flight path within the warehouse, while Figure 9 presents a screenshot of a video taken in a corridor. Fig. 8. Drone paths Fig. 9. Drone stream Next, TALON’s AI orchestrator takes charge, intelligently allocating resources either in the cloud or at the edge to ensure all scenario modules run seamlessly. By optimizing the available resources at any given time, the system achieves optimal performance and scalability. Once the resources have been allocated, the YOLO object detecion model as already mentioned, is loaded. This model processes the drone’s video input, identifying workers and checking if they are wearing helmets or vests, based on their specific work areas. The output is a video annotated with bounding boxes and text that highlight the objects of interest—the workers and their protective equipment. Figure 10 illustrates a screenshot of the module’s results. Fig. 10. Results of security compliance for PPE detector The output video from the detector is subsequently processed by an anonymization tool to ensure privacy and adhere to ethical standards. This tool blurs the faces of employees, safeguarding their identities. Figure 11 demonstrates the results after anonymization. Finally, the dashboard, where alerts are displayed, is the most crucial component of the system. During the scenario, the system continuously monitored the processed data to detect any instances of non-compliance with MAP regulations. If an employee was found not wearing the required PPE, an alert was generated and displayed on the dashboard, accompanied by the relevant image from the anonymization tool. Figure 12 shows the alert for a worker found without a helmet and vest. This enabled safety managers to take immediate corrective actions, thereby enhancing overall workplace safety. It is important to note that all the scenario steps, along with their results, are presented without blurring faces to Fig. 11. Resuls of Anonymization tool Fig. 12. Alerts dashboard demonstrate the function of each module individually. In the final system, only the final image with obscured faces accompanied by alerts would be seen by the operator on the dashboard. Additionally, all participants were informed about and provided consent for the processing of their data. VI. CONCLUSION AND FUTURE DIRECTIONS Firstly, we addressed the creation of a dataset designed to bridge the gap in existing data, particularly concerning lighting conditions, partial occlusions, and drone shots for various scenarios. The dataset was then utilized to train a detector, yielding highly competitive results despite the challenging conditions of similar scenes. Looking ahead, we could expand the dataset by introducing additional classes such as goggles and protective gloves, while also increasing the overall number of data. Subsequently, a use case scenario was presented within the context of the TALON project. This scenario demonstrated significant benefits in terms of security for warehouse workers. Based on the early results of this scenario and compared to the earlier approach, where the security officer manually patrolled the site and recorded worker status on paper, the new system has led to a remarkable improvement—inspection time has been reduced by 75%, while accuracy has increased by 70%. Moving forward, the scenario could be enhanced by incorporating additional elements like explainable AI, which would provide better justification and explanation of the results to the security officer. Finally, testing the scenario in different workplaces would present several challenges that need to be addressed. VII. MATCH AND CONTRIBUTION Our research aligns closely with the scope of ICE IEEE 2025 conference research objectives. By addressing the management of emerging technologies, we showcase the integration of drones, edge-cloud AI, and computer vision to automate PPE compliance in industrial environments. The development of a diverse, drone-acquired dataset and deployment of a real-time YOLOv8 model provides a practical framework for enhancing safety through human-machine collaboration. We analyze implementation challenges such as lighting variability, occlusion, and domain adaptation, offering realistic solutions and metrics for performance. Ultimately, our work focuses on value creation by reducing workplace hazards and improving safety protocol enforcement through intelligent automation. ACKNOWLEDGMENT This work was supported by the European Union’s Horizon Europe Research and Innovation programme TALON, under grant agreement No. 101070181. Acknowledgments to our colleagues at the VANOS warehouses [30] for their support in the measurement procedure and their collaboration. REFERENCES [1] “TALON — talon-project.eu,” https://talon-project.eu/, [Accessed 0801-2025]. [2] “Safety Helmet Detection — kaggle.com,” https://www.kaggle.com/ andrewmvd/hard-hat-detection, [Accessed 07-01-2025]. [3] M.-E. Otgonbold, M. Gochoo, F. Alnajjar, L. Ali, T.-H. Tan, J.-W. Hsieh, and P.-Y. Chen, “SHEL5K: An Extended Dataset and Benchmarking for Safety Helmet Detection,” Sensors, vol. 22, no. 6, p. 2315, Jan. 2022. [Online]. Available: https://www.mdpi.com/1424-8220/22/6/2315 [4] L. Xie, “Hardhat — dataverse.harvard.edu,” https://dataverse.harvard. edu/dataset.xhtml?persistentId=doi:10.7910/DVN/7CBGOS, [Accessed 07-01-2025]. [5] N. D. Nath, A. H. Behzadan, and S. G. Paal, “Deep learning for site safety: Real-time detection of personal protective equipment,” Automation in Construction, vol. 112, p. 103085, 2020. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0926580519308325 [6] “Labelbox — The data factory for AI teams — labelbox.com,” https: //labelbox.com/, [Accessed 23-01-2025]. [7] “Label Me &x2013; Label Printing — labelme.gr,” https://www.labelme. gr/, [Accessed 23-01-2025]. [8] R. Duan, H. Deng, M. Tian, Y. Deng, and J. Lin, “SODA: Site Object Detection dAtaset for Deep Learning in Construction,” Feb. 2022, arXiv:2202.09554. [Online]. Available: http://arxiv.org/abs/2202.09554 [9] “GitHub - njvisionpower/Safety-Helmet-Wearing-Dataset: Safety helmet wearing detect dataset, with pretrained model — github.com,” https: //github.com/njvisionpower/Safety-Helmet-Wearing-Dataset, [Accessed 07-01-2025]. [10] “GitHub - HumanSignal/labelImg: LabelImg is now part of the Label Studio community. The popular image annotation tool created by Tzutalin is no longer actively being developed, but you can check out Label Studio, the open source data labeling tool for images, text, hypertext, audio, video and time-series data. — github.com,” https: //github.com/HumanSignal/labelImg, [Accessed 24-01-2025]. [11] H. M. Ahmad and A. Rahimi, “Sh17: A dataset for human safety and personal protective equipment detection in manufacturing industry,” 2024. [Online]. Available: https://arxiv.org/abs/2407.04590 [12] B. Sekachev, N. Manovich, M. Zhiltsov, A. Zhavoronkov, D. Kalinin, B. Hoff, TOsmanov, D. Kruchinin, A. Zankevich, DmitriySidnev, M. Markelov, Johannes222, M. Chenuet, a andre, telenachos, A. Melnikov, J. Kim, L. Ilouz, N. Glazov, Priya4607, R. Tehrani, S. Jeong, V. Skubriev, S. Yonekura, vugia truong, zliang7, lizhming, and T. Truong, “opencv/cvat: v1.1.0,” Aug. 2020. [Online]. Available: https://doi.org/10.5281/zenodo.4009388 [13] S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-time object detection with region proposal networks,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 6, pp. 1137– 1149, 2017. [14] V. Isailovic, A. Peulic, M. Djapan, M. Savkovic, and A. M. Vukicevic, “The compliance of head-mounted industrial ppe by using deep learning object detectors,” Scientific Reports, vol. 12, no. 1, p. 16347, 2022. [15] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 779–788. [16] Z. Tian, C. Shen, H. Chen, and T. He, “Fcos: A simple and strong anchor-free object detector,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 4, pp. 1922–1933, 2022. [17] X. Zhou, D. Wang, and P. Kr¨ ahenb¨ uhl, “Objects as points,” arXiv preprint arXiv:1904.07850, 2019. [18] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” 2023. [Online]. Available: https://arxiv.org/abs/1706.03762 [19] N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End-to-end object detection with transformers,” in Computer Vision – ECCV 2020, A. Vedaldi, H. Bischof, T. Brox, and J.-M. Frahm, Eds. Cham: Springer International Publishing, 2020, pp. 213–229. [20] H. Zhang, F. Li, S. Liu, L. Zhang, H. Su, J. Zhu, L. Ni, and H.-Y. Shum, “DINO: DETR with improved denoising anchor boxes for end-to-end object detection,” in The Eleventh International Conference on Learning Representations, 2023. [21] J. Terven, D.-M. C´ ordova-Esparza, and J.-A. Romero-Gonz´ alez, “A comprehensive review of yolo architectures in computer vision: From yolov1 to yolov8 and yolo-nas,” Machine Learning and Knowledge Extraction, vol. 5, no. 4, pp. 1680–1716, 2023. [22] R. Varghese and S. M., “Yolov8: A novel object detection algorithm with enhanced performance and robustness,” in 2024 International Conference on Advances in Data Engineering and Intelligent Computing Systems (ADICS), 2024, pp. 1–6. [23] A. Bochkovskiy, C.-Y. Wang, and H.-Y. M. Liao, “Yolov4: Optimal speed and accuracy of object detection,” arXiv preprint arXiv:2004.10934, 2020. [24] T.-Y. Lin, P. Doll´ ar, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature pyramid networks for object detection,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 936–944. [25] G. Tsoumplekas, V. Li, I. Siniosoglou, V. Argyriou, S. K. Goudos, I. D. Moscholios, P. Radoglou-Grammatikis, and P. Sarigiannidis, “Evaluating the energy efficiency of few-shot learning for object detection in industrial settings,” arXiv preprint arXiv:2403.06631, 2024. [26] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll´ ar, and C. L. Zitnick, “Microsoft coco: Common objects in context,” in Computer Vision – ECCV 2014, D. Fleet, T. Pajdla, B. Schiele, and T. Tuytelaars, Eds. Cham: Springer International Publishing, 2014, pp. 740–755. [27] computer vision, “Worker-safety dataset,” jul 2022, visited on 2025-0203. [Online]. Available: https://universe.roboflow.com/computer-vision/ worker-safety [28] G. Jocher, A. Chaurasia, and J. Qiu, “Ultralytics yolov8,” 2023. [Online]. Available: https://github.com/ultralytics/ultralytics [29] M. McCloskey and N. J. Cohen, “Catastrophic interference in connectionist networks: The sequential learning problem,” in Psychology of learning and motivation. Elsevier, 1989, vol. 24, pp. 109–165. [30] “VANOS S.A.” https://www.vanos.gr/, [Accessed 23-02-2025].