Interest Lower Body Point's Detection For Markerless Gait Analysis
Résumé
Tracking articulated interest points for human motion analysis is a challenging task in computer vision. Most of existing works use artificial markers placed on the body to facilitate their location. Alternative markerless strategies present limitations caused by the use of some predictive models which are time consuming and sensitive to speed and moving way variations. To overpass these problems, we propose some new rules to automatically detect the interest lower body points from a frame-by-frame silhouette analysis. To evaluate the performances of our method, we perform a comparison with a ground truth manually extracted from a challenging database integrating large speed and moving way variations. This comparison is based not only on calculating the spatial deviations of the detected points compared to the ground truth ones, but also on measuring the impact of the detection errors on the gait characterization process. Obtained results clearly show the reliability of the proposed automatic detection