Research On Online Learning Algorithm For Abnormal Crowd Behavior Detection

With the social and economy development, city population density increasing, the peak flow will often occur in public places. Some abnormal crowd events will easily take place at the same time, such as the group of knife-wielding assailants attack in Kunming Railway Station, the Hajj stampede event of Mekka and Shanghai Bund in New Year’s Day of 2015. It will cause the very gravity of consequences if immediate treatment absence. Crowd anomaly detection and instant warning in public places is particularly important and significant. The abnormal detection of crowd behavior is full of challenges because of the complex scene and the various behavior of crowd in public places. Nowadays, the quickly and accurately crowd anomaly detection, as well as instant warning in public places has become a hot topic in the world. Some available algorithms for abnormal crowd behavior detection are difficult to apply to real complex environment because they are mainly based on ideal scenario and without learning and update online. Research and exploration on algorithms for abnormal crowd behavior detection is carried out, especially on online learning:1:Research on the algorithm for abnormal crowd behavior detection based on the sparse representation, and introduce the selection algorithm of overcomplete dictionary. Remove the atoms of great correlation determined by the similarity calculation of dictionary atoms, to extremely simplify the scale of the overcomplete dictionary and reduce the computational complexity in the case of unchanged expression ability. Improve the regularization norm constraint of L2, L1 through the introduction of atomic weight to determine the weight of every atom by the frequency of its normal sample. Then the atoms themselves have optimization capacity, i.e. the atom used often has lower cost in penalized function, means it has higher confidence level when expresses the normal sample. Experiment results of UMN database and amateur videos show that, relative to fixed dictionary, the modified algorithm for abnormal crowd behavior detection based on dictionary learning has higher sparse coefficient, faster recognition speed, higher recognition rate and more obvious advantage.2:To improve the rate and performance in real-time of sparse representation, a new crowd anomaly detection algorithm by Gaussian Mixture Model with online learning is introduced. From the concept, abnormal behavior is incidental and abrupt while normal behavior is steady and constant, which is similar to the definition of model extracted by Gaussian Mixture foreground. In Gaussian Mixture Model, the recurrent scenes are background, and the incidental and abrupt scenes are foreground. A crowd anomaly detection model is equivalent to a Gaussian Mixture Model. Detect the quantity and distribution of foreground in certain conditions of sensitive movement through parameters value changed. Spatio-temporal modeling is made combined timeline to detect the changes of regional crowd movement.The algorithm for abnormal crowd behavior detection based on Gaussian Mixture Model is created to extract sensitive movement point in certain speed of the video frame with the adaptive threshold. Form a spatio-temporal mode of behavior by introducing the temporal information, to identify the group anomaly classification by the time-variation analysis of sensitive movement point. The adaptive judgment of threshold and dynamic determination of initial center in the algorithm make the parameters acquired by learning, to realize the online learning and updating. This method not only can identify the group anomaly and distinguish the classification (scattered or mass brawl etc.) at a rapid rate, but also can adaptively update internal threshold and maintain a high accuracy under the influence of noise, such as illumination etc.Compare and analyze the advantages and disadvantages of those two online learning algorithms for abnormal crowd behavior detection in the end. The method of Gaussian Mixture Model has high recognition rate in the whole situation, well classification discrimination, speed online updating and good adaption to the rapidly changing scenes, but is difficult to distinguish the local anomaly. The method of sparse representation has well anomaly discrimination both in local and global, high stability at online updating but low speed, difficult to adapt to the rapidly changing scenes.