Abstract :
[en] Human motion analysis using 3D skeleton representations has been a very active research area in the computer vision community. The popularity of this high-level representation mainly results from the large variety of possible real-world applications such as video surveillance, video conferencing, human-computer interaction, virtual reality, healthcare, and sports. Despite the effectiveness of recent 3D skeleton-based approaches, their suitability to real-world scenarios still needs to be assessed. Using these approaches in a real-world scenario can give new insights on how to improve them for reaching real-world standards. In this thesis, we propose new solutions to mitigate existing constraints for the deployment of 3D skeleton-based approaches in various real-world scenarios. For that purpose, we investigate two human motion analysis applications that are based on 3D skeletons, namely, home-based rehabilitation of functional activities and human motion analysis in the wild.
In the first part of this thesis, we propose a low-cost solution designed for supporting home-based rehabilitation of stroke survivors under the remote supervision of a therapist. To that end, we introduce the concept of color-based feedback proposals for guiding the patients in real-time while exercising. More specifically, color-based codes are visualized for informing the patient on the accuracy of the movement and on the adequacy of the posture. Feedback proposals are tailored to each patient's body anthropometry. An initial clinical validation shows an improvement of the posture and of the quality of motion when using the proposed feedback proposals.
In the second part of this thesis, we focus on human motion analysis in the wild in the context of cross-view action recognition. We propose and investigate different 3D human pose estimation techniques from a single RGB camera in order to take advantage of 3D skeleton-based approaches. Indeed, given their 3D nature, 3D skeletons can overcome more easily the challenge of viewpoint variability in contrast to 2D-based approaches. To show the relevance of 3D pose estimation techniques in the context of human motion analysis, two different pipelines are proposed. The first pipeline makes use of a per-frame pose estimation approach. Per-frame pose estimation shows temporal inconsistency and small fluctuations in the skeleton joint locations over time. Considering this, the second framework is then based on a sequence-to-sequence pose estimation, providing, therefore, temporally consistent skeleton sequences that are more robust to sensing in the wild. These two pipelines show an improvement in recognition accuracy as compared to state-of-the-art approaches on two different well-known datasets. However, despite their relevance, 3D human pose estimation methods present some limitations. For example, their accuracy drops significantly in the presence of unseen environments or situations, eg, challenging camera locations, and outdoor conditions. For that reason, we introduce 3DBodyTex.Pose dataset, an original dataset to address the challenges of camera locations and outdoor scenarios in the context of 3D human pose estimation. Moreover, 3DBodyTex.Pose offers to the research community new possibilities for the generalization of 3D human pose estimation from monocular in-the-wild images from arbitrary camera viewpoints.
