Research Of Abnormal Behavior Detection Algorithm In Video Surveillance

Abnormal behavior detection is a key technology in the field of intelligent video surveillance.It can automatically detect the abnormal behavior occurred in the scene and alarm in time,which can effectively reduce the personal injury and property damage of the masses in public places.The abnormal behavior detection consists of global and local detection.Global anomaly usually refers to the crowd anomaly,which only need frame-level detection.It suffers from the background and illumination change etc.Meanwhile,local anomaly detection need pixel-level detection and behavior always shows the diversity.In this paper,we propose different detection framework respectively.The main contributions of this paper include the following aspects:1.In this paper,we propose a novel detection framework based on feature fusion for global anomaly detection.To learn more discriminative and robust feature representations,we exploit the complementary information of both appearance(saliency information)and motion(multiscale histogram of optical flow)patterns.Specifically,sparse autoencoders are proposed to learn the joint representation of appearance and motion features,which also has the effect of dimension reduction to reduce redundancy at the same time.Finally,in order to learn more effective global representation,the fisher vector is introduced to the framework instead of the bag of words.Based on the global feature,a linear support vector machine is adopted to detect anomaly,which achieves the auc of 99.65% in UMN dataset and 87.6% in Web dataset.2.Consider the problem of local abnormal behavior detection,this paper propose a novel detection framework based on convolutional neural network.As we know,most existing works are proposed based on hand-crafted features.However,it still remains challenging to decide which kind of feature is suitable for a specific situation.In addition,it is hard and timeconsuming to design an effective descriptor.In this paper,we propose a 7-layers deep neural network,which combined 2D convolution and 3D convolution.It can learn more discriminative and high level feature based on spatio-temporal information.In order to deal with the problem of gradient diffusion and speed up the convergence in train phase,this paper introduces the unsupervised slow feature analysis to pre-train the parameter of the 2D convolution layer.Finally,abnormal candidate area location is proposed to achieve lower conputational cost.In this paper,we experimentally evaluate our model on benchmark dataset USCD and achieve the detection rate of 82.1%.