Abnormal Behaviors Understanding Of Moving Targets In Tansportation Surveillance System

In recent years, Intelligent Transportation Systems(ITS) with high efficiency and security has received extensive attention by international scholars. As an important part of ITS, abnormal behaviors recognition and understanding has great significance to forecast and deal with traffic accidents in time. However, most of exsit algorithms only apply to simple behaviors recognition and have low recognition rate in complicate scenes which contain pedestrian and vehicles. Based on the above background, this thesis studies the detection and tracking of motion targets in traffic scene, as well as abnormal behavior understanding.In targets detection aspect, a moving objects detection algorithm based on improved watershed method was proposed. First an adaptive background subtraction process was applied, and the initial denoised threshold images which contained moving objects regions were obtained by Otsu algorithm and morphologic methods. Then foreground and background were labeled respectively. Based on the above results, a labeled watershed transform was applied. Experimental results show that the algorithm can overcome the defect of excessive segmentations and extract the contours of moving objects accurately.In targets tracking aspect, the article proposed a template matching algorithm based on improved Hausdorff distance. First convert contours extracted by the watershed algorithm that aforementioned into multi-valued images and considered it as the template. Then achieved object tracking by matching the real-time image and template accrrding to improved weight Hausdorff distance. Experimental results show that proposed algorithm has the advantange in instantaneity and accurary over traditional algorithms.In terms of behaviors recognition and understanding, An abnormal behavior recognition algorithm based on Hidden Markov Model(HMM) in transportation scene was proposed. First movement characteristics were extracted as basic semantic units by objects detection and tracking. Then these semantic units were combined into simple behavior observation sequences. Then the abnormal behaviors models were built according to the Baum-Welch algorithm. Finally, abnormal transportation behaviors were recognized by the forward algorithm and thresholding. Experimental results show that the algorithm which has a high instantaneity can effectively recognize abnormal behaviors such as pedestrian crossing the fence, vehicle retrograding, parking and rolling the yellow solid lines.