Research On Abnormal Event Detectionbased On Sparse Combination Learning

With the developing of cities and the increasing of population, video surveillance has been used widely. However, traditional video surveillance are detected by people primarily. This method can make people tired easily, and it may cause false alarm or miss some alarm. Finally, it brings huge challenge to social safety. Therefore, an abnormal detection system which has the properties of accuracy and real-time processing is more and more popular. At the same time, abnormal detection has been a hot research in the field of computer vision.Recently, the methods of abnormal detection has developed very well. Among them, the method which is based sparse representation theory has shown a better performance in the aspect of accuracy compared to other methods. However, by doing some research, we can find this method has some shortcoming. If we choose to use sparse representation to represent a signal, always need to find the suit bases combination from the redundancy space, and it can consume a lot of time which will lead to a slow detecting speed. So, we study a method which is called sparse combination to replace the original dictionary. By restricting the number of combination’s bases, and letting each combination represent the training as more as possible. At last, we will obtain many combinations, but its number is much less than the number of searching combination from dictionary. And we just use the combinations to represent testing data rather than using a dictionary. By this way, the abnormal detection system has a better detecting speed.Surrounding the sparse combination learning, we propose two strategy. The first one is updating the combinations online. Considering this situation, with the time spreading, the sparse combination of prior learning may not represent the current events very well. And according to the theory that the events in video are correlated with sequential N frames in space and time, and it has the property of consistency. We choose to save the events which has been proved normal, then use these events to learn some new combinations, so that the system can improve the accuracy of detection of abnormal events. The second strategy is modeling the background to remove redundancy. Since the surveillance video itself has a lot of redundancy of abnormal events, in the process of learning sparse combinations, the system has to handle a large number of recurring static background “events”, and this is very time-consuming. Considering this situation, we study the common background modeling method and extract the current background from the video. Then we use the background to learn sparse combinations, so that we can remove a lot of redundancy and reduce the time-consuming of learning sparse combinations. At the same time, it can help the system improve efficiency.Finally, we use some public video library in the experiment, such as AVENUE,UCSD and UMN database. At the same time, by comparing the detecting accuracy and speed with other algorithms, we achieve a precision which is closed to the optimal one, but the speed of detection is faster than them. Our method can achieve the requirement of actual.