Extracting Gait Figures in a Video based on Markerless Motion

This paper proposes a new method to extract gait figures in a 2D video without using any markers. Such scenario is more feasible in a real-world environment than a traditional 3D cooperative multicamera system with reflective markers which is costly and complicated. The proposed method is developed to extract following information from a 2D gait video based on markerless motion: 1) a gait period; 2) key positions of a human body (i.e. head, waist, left-knee, right-knee, left-ankle, and right-ankle) in each frame within a gait period. This is processed by using statistical techniques including linear regression, parabolic regression and polynomial interpolation. Such extracted gait information is useful for many gait-based applications such as human identification in a surveillance system, injury analysis in a sport science, and disease detection and gait rehabilitation in a clinical area. The widely adopted CASIA gait database B is used to verify the proposed method. The extracted key positions are validated by comparing with a ground-truth which is manually generated by human observers. The experimental results demonstrate that the proposed method can achieve very promising performance.