Detection of moving people with mobile cameras by fast motion segmentation

Detecting humans from video sequences is a key and difficult problem in computer vision. The problem becomes even more challenging when the camera is mobile, and thus, the background is not static. In this case, the traditional approach of background subtraction cannot be employed. Methods based on Histogram of Oriented Gradients (HOG) have been introduced and widely used for human detection due to reliable performance. Yet, these sliding window-based algorithms are computationally expensive to run on embedded platforms. In order to address this problem, we present a significantly faster moving human detection method that is based on fast motion segmentation, and uses HOG as the feature descriptor. First, edges are detected on two consecutive frames, and the difference between the edge images is computed. Then, a new edge-based frame alignment method is used to find the global minimum for motion compensation, and segment the motion region of interest (ROI). Instead of searching for the human(s) in the whole frame, the HOG features in sliding windows are calculated only in the ROI. Finally, Support Vector Machine (SVM) is used to classify human and non-human regions. Compared to the traditional method of using HOG and searching the whole frame, the proposed method significantly reduces the search region by motion segmentation, and speeds up the detection process. Experiments have been performed on three different scenarios, and the results show that the reduction in execution time of a frame can reach as high as 95.83%.