Research On Video Abnormal Event Detection Based On Online Adaptive Dictionary Learning

In recent years,abnormal event detection based on sparse representation has gradually become a hot research topic. Using this method usually needs to train a dictionary using the normal samples to deal with the problem of sparse representation.On the one hand,the scene and the events continue to change in the real application environment. The model that was trained offline can not adapt to the change,which will result in the decrease of the detection accuracy. On the other hand,the size of the dictionary is often relatively large in order to make the dictionary have good representation,which will result in the decrease of detection speed. Therefore,this paper studied a kind of abnormal event detection algorithm based on online adaptive dictionary learning. The main research contents included:An online adaptive dictionary learning algorithm is proposed. Online performance is reflected the dictionary learning speed is faster because of the use of a single step iterative method to update the dictionary. Adaptive performance is reflected in the process of using the algorithm to update the dictionary. In the process of updating the dictionary,the active degree of the current signal data is taken into account. The dictionary can has better representation for current signal data through this way. In the process of the dictionary learning, the updating strategy of the weight matrix and the online updating strategy of the dictionary are proposed. The updating strategy of the weight matrix can update the active degree of the basis in the dictionary to the signal data. In the process of updating the dictionary, the vector base in the dictionary is always consistent with the change of signal data using the activity of the vector basis.Therefore, the dictionary learning algorithm has a better adaptability and the dictionary keeps a good representation, while the size of the dictionary is reduced.The online adaptive dictionary learning algorithm is applied to the abnormal event detection,and an abnormal event detection algorithm based on online adaptive dictionary learning is proposed. Because the size of the dictionary is small,the dictionary is updated faster and the algorithm has a faster detection rate. In the process of abnormal event detection,the algorithm updates the dictionary continually. This method is more flexible than the models that are trained offline and maintain a better detection accuracy.In this paper,we use the public video dataset: UMN and Subway to test the performance of detection and the performance of other abnormal behavior detection algorithm is compared with our method. Experiments show that the method could gain higher detection accuracy and rate in this paper.