Résumé
Musculoskeletal disorders (MSDs) represent a major occupational health issue, accounting for a significant proportion of work-related illnesses. They pose a considerable challenge in industrial environments due to their adverse effects on workers health and overall productivity. In the context of Industry 5.0, which emphasizes a human-centered approach, preventing these disorders requires practical and accessible ergonomic assessment tools. In this paper, we aim to investigate video-based estimation of load handling to advance automated ergonomic evaluation, rather than relying on inertial sensors. Specifically, two models are proposed: an image-level classification model based on MobileNetV2 and a video-level classification model based on R3D-18. Both models are trained and validated on ViLoad, a custom dataset collected under controlled conditions replicating typical industrial handling tasks. ViLoad is introduced as a new resource to encourage further research on vision-based load estimation. The obtained results are satisfactory and demonstrate the relevance of deep learning architectures for load estimation from visual information. Together with the creation of the ViLoad dataset, these results open new perspectives for future research and provide a valuable foundation for advancing video-based ergonomic assessment using deep learning.