Research On Violence Detection Based On Gaussian Mixture Model Combined With K-Means Clustering

Violent incidents will bring social fears and turbulences.With the progress of the technology,surveillant videos are gradually used in public places.Using intelligent algorithm to detect whether the surveillan t videos have the nature of violence,and further more sending policemen to stop violence are effective ways to maintain social stability.Modern surveillant recognitions are mostly used in the recognition of basic actions such as running and jumping.It is difficult to identify violent behaviors because of their complexity and diversity.Therefore,aiming at maintaining social public order and supervising the viol ent behavior of criminals,and aiming at foreground detection,feature extraction,model training and behavior recognition of the videos,the article proposes a violence detection system based on gaussian mixture model combined with K-means clustering.The specific contents include:(1)Combining the traditional K-means clustering algorithm with gaussian mixture model,an improved gaussian mixture model was proposed.After the RGB three channel means of all pixels contained in the first 40 frames of the video were calculated,the K-means clustering algorithm was used to find out the clustering value of all pixels’ means.If the pixel mean of the created gaussian model was less than the cluster value,the cluster value would be updated and a new gaussian model was added.Then the combined vector s of pixel values in image frame’s R channel,G channel and B channel which were included in 40 frames of video,would be multiplied by its transpose vectors.The multiplied result value was compared with the cluster value,then cluster value would be updated,and the mean value,root of difference and other parameters of gaussian model would be updated too.Finally,the pixel value of the image frames in the original videos was compared with the generated threshold in the gaussian model,and the motion foreground was extracted.Then the extracted foreground results were compared with visual background extraction,codebook and gaussian mixture model,it was founded that the foreground results extracted by the improved gaussian mixture model were more complete,which solved the problems of incomplete foreground extraction in traditional gaussian mixture model.(2)The foreground contours of human body using by gaussian mixture model based on K-means clustering algorithm were calculated for feature extraction.The foreground features of the obtained binary contour picture s were extracted by Fourier description operator and Hu moment,and other features of the original video frames were extracted by optical flow method and Log Gabor filter.Then the mixed feature would be obtained from the foreground features and other features.And BP neural network was used for model training and behavior classification of the foreground features,the other features and the mixed features respectively.Finally,it was found that the classification result of mixed features was best in the three features,whose classification results were more than 80% in Weizmann and KTH databases,and the classification results were more than 70% in Hockey Fight databases.(3)A Long Short Term Memory network based on VGG16 network was proposed in the paper.The original image frames of the video s in the Hockey Fight databases were inputted into the VGG16 network for preprocessing to obtain the preprocessed data,and then the preprocessed data were inputted into the Long Short Term Memory network for classification.The classification results were more than70%,which is similar to the classifi cation results of BP neural network based on mixed features in(2).