Vision-based 3D Human Motion Analysis in a Hierarchical Way

Vision-based 3D Human Motion Analysis in a Hierarchical Way

Huo, F.

Reinders, M.J.T. (promotor)
Hendriks, E.A. (promotor)

Electrical Engineering, Mathematics and Computer Science

Intelligent Systems

2013-06-27

In the last decade, computer vision has drawn more and more attention because of its potential applications in our daily lives, such as health care, education, safety, and training. However the high complexity of many of the vision-based approaches hinders their practical applications, especially in applications where immediate feedback is required. This thesis focuses on developing fast and robust algorithms to track multiple people in a scene, which can be used in applications like interactive education, crisis analysis, serious games, or virtual reality. We proposed a complete system for people tracking and interaction recognition. This system consists of three modules: detection and tracking persons from a single camera view, position estimation of body parts of each person from multiple camera views, and human pose and interaction recognition. In the first module, we detect and track people in a single camera view. A simple and efficient approach is proposed to track a single person’s motion in real time. Based on this approach, we designed a pose-driven spatial game, which demonstrated a practical vision-based application. Further, we extended this single person tracking approach into multiple people tracking, by using color information to distinguish persons. Due the simplicity of the features and the simplified model, the tracker is able to track two persons simultaneously in close real-time. In the second module, we investigate multiple view approaches for multiple people tracking and pose estimation. The advantages of multiple view approaches over single view approaches are more camera views and better depth estimation. Multiple view approaches can lead to more accurate tracking compared with single view approaches, especially for multiple persons tracking. We proposed a novel approach that combines multiple views in such a way that we rely more on the features derived from clear views and less on those from occluded views. In this way, the proposed approach leads to accurate tracking results at much lower computational costs. In the third module, we use the tracking results (positions of body parts) to represent and recognize human interactions. We investigate what are the most informative features to distinguish the interactions while keeping good recognition performance. The necessity of informative feature lies in practical applications. By using the informative features, even a linear classifier is shown to be sufficient for interaction recognition. This is quite attractive from a computational point of view. Our proposed approaches do not only contribute to academic research in vision-based human motion analysis, but also provide practical solutions to innovative human computer interactions.

surveillance
computer vision
multiple people tracking
pose estimation

https://doi.org/10.4233/uuid:a14f2944-d922-4127-8ad9-81966190a227

9789491536090

Institutional Repository

doctoral thesis

(c) 2013 Huo, F.