Research On Crowd Abnormal Detection By Using Feature Fusion And Sparse Representation

With the continuous development of social economy,crowd intensive scenes continue to increase.Intelligent monitoring and alarm for crowd status has become an important demand in current society,which has great research value and economic value.The work of this paper focuses on the algorithm of abnormal behavior detection in surveillance video.The extraction of foreground area,extraction of fusion features,construction and update of sparse representation dictionary,and detection of abnormal behavior are mainly studied.The main research work in this paper is as follows:First,serveral foreground detection methods are studied and analyzed.ViBe algorithm is used to extract the foreground target area.The extraction of foreground region is beneficial to reduce the time of subsequent feature extraction.Then the HOG features and MHOF features of the foreground area are extracted.When the MHOF feature is extracted,the Harris corner is used as the feature point to represent the movement crowd,and the extracted corner points are intersected with the foreground points set to reduce the redundancy of feature points.The MHOF feature extraction time is reduced by using the filtered feature points to construct the optical flow field.Finally,the fusion feature is obtained by normalize the HOG features and MHOF features.In addition,while the MHOF feature is extracted,the velocity distribution of feature points is statistically used for frame state estimation.After extracting the features,the overcomplete dictionary is constructed.In this paper,a method of sparse representation is proposed by using normal samples and abnormal samples to construct normal abnormal double dictionaries.The state of the video frame is estimated by the cost of reconstruction and the percentage of overspeed particles.When the dictionary reconstruction cost is unable to estimate the state of the video frame,the feature of current video frame and the number of consecutive occurrences are preserved.Though the statistical feature point velocity distribution information,the percentage of overspeed particles in the video is calculated,and the state of the video frame is estimated by comparing with the threshold of overspeed particles.When the number of consecutive occurrences exceeds the threshold,the dictionary is updated online using the saved features.For anomaly location,an anomaly location method based on the foreground area and speed of the moving target is studied in this paper,and the area weight is introduced to reduce the effect of perspective effect on the target area.The target foreground area and speed discriptor can locate the target which has fast moving speed and large volume.The UMN and UCSD standard dataset is used for experimental analysis.The experimental results show that the proposed method has good performance in real-time and accuracy.