The Study Of Vehicle Trajectory Recognition And Online Anomaly Detection With Incremental Modeling

Vehicle trajectory recognition and online anomaly detection are an important research direction of intelligent transportation systems. They play a supporting role to timely and effective rescue and post-processing of real-life accidents, and they can reduce traffic delays and secondary accidents cause, meantime they can provide important information for urban traffic monitoring and security management.Usually, the vehicle trajectory recognition and online anomaly detection are performed based on the trajectory modeling which are divided into two types, statistical model and motion model. Unsupervised learning is proposed in recent years, which is a new trajectory modeling approach. This method can be effective to trajectory recognition and online anomaly detection in the condition of the data excluding abnormal trajectory and lots of training data, but for the initial trajectory set containing the abnormal trajectory and a small amount of initial trajectory set, recognition rate and accuracy of the abnormality detection is relatively low.To solve these problems, incremental EM algorithm is applied to trajectory modeling based on unsupervised. An incremental trajectory modeling method based on batch-mode model initialize was proposed, and was applied to vehicle trajectory recognition and online anomaly detection. Firstly, the proposed method utilized improved Hausdorff distance to measure the similarity between trajectories and spectral clustering algorithm to cluster the initial trajectories, and in consequence we can get the distribution patterns of trajectory.Secondly, For each type of clustered sample in the initial trajectories, Hidden Markov Model for each behavior class is established using multiple observation sequence training methods,thus the initial trajectory models are obtained. Thirdly, for a new trajectory captured form the video image, we find the most likely normal trajectory class using maximum a posteriori estimate. Online anomaly detection was performed using intelligent threshold method. The trajectory class was recognized, and the Hidden Markov Model parameters were updated with incremental EM algorithm. Finally in order to increase the adaptability of this method, the model structure was updated by trimming mixture component.Outdoor real world shooting video have been tested by the method. Experimental results show that, comparing with the classic batch-mode algorithm, incremental trajectory modeling can get more accurate trajectory models and faster speed. At the same time, the proposed algorithm has higher recognition rate in the condition of the initial trajectory set containing the abnormal trajectory, and higher detection rate and lower false alarm rate in anomaly detection. The algorithm realizes online anomaly detection and it is insensitive to the initial trajectories.