Simultaneously detect and segment pedestrian

We present a framework to simultaneously detect and segment pedestrian in images. Our work is based on part-based method. We first segment the image into superpixels, then assemble superpixels into body part candidates by comparing the assembled shape with pre-built template library. A “structure-based” shape matching algorithm is developed to measure the shape similarity. All the body part candidates are input into our modified AND/OR graph to generate the most reasonable combination. The graph describes the possible variation of body configuration and model the constrain relationship between body parts. We perform comparison experiments on the public database and the results show the effectiveness of our framework.