Vision-Based Pedestrian Abnormal Crossing Behavior Recognition and Analysis
Hongyu Hu,Zhaowei Qu,Zhihui Li,Sheng Jiang,Jinhui Hu
2012-01-01
Abstract:At present, study of pedestrian behaviors usually depends on data collection and analysis, theoretical modeling as well as simulation. By these means, traffic safety diagnosis and evaluation, signal timing of mixed traffic, road design and pedestrian-vehicle collision could be well guided. In this paper, the authors developed a novel method to automatically detect abnormal pedestrian crossing behavior based on video processing. Based on object trajectories data extracted by video tracking, motion patterns could be learned automatically, which is an effective approach for understanding traffic scene as well as modeling and analyzing traffic behavior. Firstly, crossing pedestrian trajectories are obtained based on motion detection and tracking. Secondly, for estimating the crosswalk zone, the authors use origin points and destination points of pedestrian trajectories to construct the origin point set and the destination point set. These two separate point sets are fitted by 2-dimension Gaussian mixed model (2-D GMM) to extract the distribution of entry and exit zones. And then K-means algorithm is applied to acquire the initial parameters of GMM, and expectation-maximization algorithm is used to optimize the parameters. Based on these steps, the crosswalk region could be estimated automatically combined with hypothesis predefined. Moreover, pattern matching algorithm based on Bayesian classifier is presented for abnormal pedestrian crossing detection and analysis. At last, the entire approach proposed is implemented and tested at crosswalks in real-world. And the pedestrian crossing behaviors are analyzed. Two reasons leading to abnormal crossing are reduced and some helpful advices are given.