Human Behavior Understanding via Top-View Vision

Aiming to target occlusion problem in complex scenes, human action recognition via a top-view vision is proposed. However, in this view, the human behavior rotated will be mistakenly identified as another behavior. To address this situation, taking into account the rotation invariance of the moments, human static posture are represented by Hu moments, and using the SVM as trainer and classifier. According to the change of coordinate of the binary image centroid, semantic web of dynamic behavior is established. The experimental results show that this method can accurately identify the human dynamic information and has a high recognition rate.