文摘
This paper presents a real-time pedestrian counting approach in dense visual surveillance scenarios with a fixed camera. Firstly, we build a set of generic pedestrian models, which is irrelevant to application scenarios. Feature points are sampled from the edge images of pedestrian, which are further used to determine the potential positions of pedestrian by matching the edges with the head-shoulder template. The template is generated through utilizing the Parzen Windows method on a self-made sample dataset of head-shoulder edge images. Secondly, we refine the frame based counting results with MHT (Multiple Hypothesis Tracking) under a Bayesian framework. And two effective improvements are proposed to resolve the missing tracking problem caused in traditional MHT tracking method. With the data association in the tracking process, false positions in the matching process could be removed and missed ones could be added.