Person Detection by Low-rank Sparse Aggregate Channel Features

Human detection in the video has several applications in security and surveillance. Human detection using video is desired to be robust against illumination, occlusions, scale, translation and view angle variations. In this paper, we develop an approach which can improve the performance of the aggregate channel feature for a high view angle. The foreground is estimated using a frame differences approach to identify the location of moving objects in the static camera scene. The sparse basis is included in the aggregate channel feature vector to describe the foreground region of each frame of the video. This approach provides better miss rate versus false positive per image as compared to the existing aggregate channel feature and histogram of the oriented gradient.