Research On Algorithm Of Human Anomaly Behaviors Detection Based On Surveillance Videos

The algorithm of human anomaly behaviors detection based on surveillance videos is one of the hot research topics in the field of computer vision.The traditional artificial video surveillance system,which is limited by the physiological characteristics of human beings,has the disadvantages of poor real-time capability,low efficiency,high labor costs,prone to misjudge or leave out and so on.With the algorithm of anomaly behaviors detection being the core algorithm,the intelligent video surveillance system can automatically detect anomaly behaviors in videos and give an alarm without manual intervention.Therefore,it has great value of scientific application and application prospect.After the research status and development trends of current anomaly behaviors detection methods being deeply studied and fully described,corresponding improved algorithm is proposed in this thesis according to the existing algorithms.Complete experiments and data analysis has been given to verify the advantages on accuracy and robustness of the improved algorithm in this thesis.Firstly,a new spatio-temporal descriptor for characterizing the human behavior in the videos is proposed in this thesis.The descriptor is obtained using the neighborhood dense sampling method of each pixel in spatio-temporal volumes.To measure the temporal and spatial features of video,the original HOG feature is expanded from two dimensional pictures to three dimensional spatio-temporal volumes.Then it is extrated in the spatio-temporal volumes.Gaussian weights are used to measure the different contributions of the different gradient histograms of spatio-temporal volumes according to its position in spatio-temporal ensembles.Weights are calculated by Gaussian function according to the distance from the current spatio-temporal volume to the center spatio-temporal volumes of the spatio-temporal ensemble.The final experimental results indicate that the proposed descriptor can effectively express human behaviors and improve the robustness of the algorithm.Secondly,a weighted fuzzy clustering method based on symmetrical KL divergence is proposed for the distribution of various behaviors in the video.The entire cluster process is split to multiple stepwise clusters.The data participated in the cluster will be compressed into several weighted cluster centers according to the cluster results after each cluster.The requirement of the memory of the system will be reduced in this method.In order to measure the similarity and dissimilarity between feature and cluster centers better,symmetrical KL divergence method is used to calculate the similarity,making the fuzzy clustering method fitting the anomaly detection field better.Thirdly,the adaptive threshold anomaly detection method based on visual dictionary is proposed in this thesis.The weighted fuzzy c-means clustering method based on symmetrical KL divergence is adopted to obtain the visual dictionary and the membership matrix.The representative pixels are found according to the visual dictionary and the membership matrix.Finally,threshold of each visual word is calculated using all representative pixels.The thresholds are used to determine whether each pixel belongs to the anomaly behaviors.The algorithm is verified on the UCSD anomaly detection database by experiments.The result shows that the algorithm of this paper is of high accuracy.To sum up,three aspects are studied and then improved algorithms are proposed in this thesis:human behavior feature descriptors,feature learning method,the method of abnormal behavior detection.Experiments show that the entire algorithm can detect abnormal behavior from surveillance video accurately.