Shape based detection and retrieval

Human movement analysis is a long-studied, but still important and challenging research area in visual surveillance. It involves many fundamental problems in computer vision such as human detection, segmentation and tracking, and higher level problems such as human gesture, action and event recognition. Shape is the most dominant cue for detecting humans due to large appearance variability. In this paper, we present a shape matching approach for human detection and tracking. We search for approximate human shapes in still images by using shape templates. The templates are obtained from the silhouettes of the set of training images. The system is trained using codebook entries which consist of shape context and feature parameters of the training images. The features are extracted from the input images of different scales and are matched with the templates. The best matches are selected by computing the match score. We can also customize our training system by providing our own training set of images.