Research On Trajectory Extraction And Anomaly Behavior Detection Of Pedestrian Based On Video

With the growing of surveillance video data, manual labor-intensive way ofanalyzing video not only wastes labor but also increases the cost of security, and it isunable to meet the actual needs increasingly. In this case, the intelligent method ofvideo analysis is particularly important, as a research direction, anomaly detection hasattracted widespread attention. Anomaly detection is a way of analyzing and detectingof target which has abnormal behavior or abnormal events through video. Thetrajectory of moving objects which has high research value contain a wealth ofspatio-temporal information in video, it can be used to infer the behavior of the targetbehavior characteristics. However, the most commonly used analysis method whichbased on the trajectory is so complex that it just only stays in the theoretical analysisphase.In this paper, we present and implement a system which can detect abnormalbehavior of pedestrian in real-time, which contain reliable trajectory extracted oftarget and the modeling of abnormal behavior and detection of anomaly, based on thestudy of the conventional step of abnormal behavior detection based on the trajectory.The system includes detection of moving targets in real-time, stable tracking andextracting trajectory, extraction and learning of trajectory feature and abnormalbehavior detection in four stages.Firstly, we introduce a Gaussian distribution with confidence and variablelearning rate which can accelerate the learning speed of the scene and the buildingprocess of background to improve Gaussian mixture model, then we use it to detectforeground. Since the shadow which detected as foreground will make irregular shapeand expand the target range, affects the tracking process, therefore, we use the methodbased on the maximum chroma difference in color space to eliminate shadow andobtain a more precise target shape, reduces the probability of occlusion among theforegrounds.Then, we set the objective correlation method, which combine the future positionmatching and mean-shift algorithm, with studying the stability tracking of target. Wedivide the target into several kinds of state according to the different situations thatmay occur in the process of tracking, and manage the state by using the finite-statemachine, conditions prompt state transition occurs, the whole process will be recorded when the target come into view until it leave, thus it can be extracted complete andreliable information. Meanwhile, we set the timer for the stationary target, to detect itwhether stay too long.Finally, we get the fixed length vector by using the information such as velocity,acceleration, abnormal trajectory growth, and build the self-organizing networktraining these eigenvectors, to get the general characteristics and topology structure ofthe normal track points in scene area, and the track points will be detected if abnormalwith regarding to the eigenvectors value between the normal track point and thecandidate track point. The trajectory will be set as abnormal if it contains multipleabnormal track points.The proposed method can test abnormal track points rapidly, which make itpossible for warning of suspected pedestrian path in real-time. Through a video takenfrom the outdoor parking area, we test and the experiments outcome shows that themethod for abnormal trajectory has good inspection rate.