Detection Of Crowd Abnormal Events Based On Block Vibe And Feature Fusion

The detection of abnormal events in the crowd is mainly to recognize abnormal events in surveillance environment automatically by computer and alarm timely.This thesis focuses on shadows and variety of abnormalities in crowd abnormal events detection,and proposes a crowd abnormal event detection method based on block ViBe and feature fusion.The main work and innovations include:In the stage of target detection,aiming at the shadow problems in the ViBe detection process,this thesis improves ViBe method by blocking.First,after object detection by the Vibe method,the current frame is partitioned,and furthermore divided into a foreground block map and a background block map.The HOG feature is calculated for the foreground block map and the background block map individually.The misjudged foreground blocks are reduced by HOG feature,and then the foreground and background maps are transformed from RGB to HSV color space.For foreground blocks,count the pixel V component values in the background block in the 3 ? 3 neighborhood of foreground block,and the background thresholds in the neighborhood are obtained by calculating the V component values of neighboring background blocks.Finally the shadow is removed by this threshold.In the feature extraction stage,this thesis extracts both optical flow and texture features and cascade them to construct event feature.The histogram of the optical flow direction is performed block by block,and then cascaded in horizontal and vertical directions to constitute the optical flow motion vector.For texture features,this thesis extracts Scale Invariant Local Ternary Pattern(SILTP)feature in each block in multiscale Gaussian pyramid images,and then the histogram vectors of all the blocks are connected in series to form a texture feature vector.Image feature vectors are inputted SVM classifier for classification and recognition.The method of this thesis was performed on the database of two abnormal events,UCSD and UMN.ROC curves and AUC values are used as evaluation indicators.Experimental results show that this method can effectively detect abnormal events in these database.