利用分層俯視角深度特徵應用於日常活動辨識

Abstract

In this thesis, two novel features for activity recognition from top-view depth image sequences are firstly proposed. Most of previous works are focusing mainly on dealing with the side-view depth image sequences, which unfortunately may encounter occlusion problems. Therefore, top-view camera setting is adopted in our thesis. Based on the notion of computed tomography, the top-view depth images are segmented to different layer along z-axis. Then, the representative body points which are found on each layered image will be a meaningful feature as the substitute of body parts for the activity postures. Besides, a discriminative shape descriptor is also proposed to describe the human shape for different activity postures. Based on the occupancy value of small region, the cylinders-sector occupancy grid with saturation function is proposed to capture special characteristic of top-view human shape. To make our proposed features invariant to orientation, the human orientation is also calculated by extracting the regions of head and shoulders, and then refines the above two features according to the orientation. Finally, dynamic time warping algorithm is applied to address the problem with different sequence lengths and the SVM classifier is trained to classify our activities. To verify our performance, 2 new top-view datasets are constructed. In our experiments, challenging cross-subject tests are conducted, and the effectiveness of our representative body points and layered sector-based shape descriptor are demonstrated. The result shows that the accuracy can achieve up to 96%, which is quite promising while being compared with those from the state-of-the-art methods in the literature.