Research Of Crowd Status Analysis And Abnormal Event Detection In Video Surveillance

With the worldwide population growth and rapid urbanization, massive crowdshave become increasingly frequent and how to effectively manage and control thehuman activities became the outstanding issues of public safety. Smart crowdsurveillance technology has become a hot research topic in intelligent monitoring,using the technigues like computer vision, image processing and pattern recognitionto achieve the crowd scene state analysis and abnormal event detection has importantpractical significance in crowd management,which is necessiry in safety managementof public places. This paper focuses on the research of crowd density estimation andmotion analysis algorithm in video surveillance and uses the population density levelsand the states of motion to realize the abnormal event detection and recognition.In crowd density estimation, the paper proposes a texture analysis method basedon local binary patterns combined with GLCM (LBP-GLCM) to extract theforeground density characteristics, and then apply the support vector machine (SVM)machine learning method for the classification of population density. Populationdensity state is divided into free, ristricted, dense and jammed four categories, usingthe foreground template texture analysis method effectively improve the accuracy ofdensity estimation for avoiding the interference of background. In order to achievelocal density analysis and positioning, this paper also proposes a method based on thesliding window technique based on image block processing. The proposed densityestimation algorithm is suitable in local and overall situation and has a higherclassification accuracy for low and high desity.In crowd motion analysis, this paper obtained the motion vector field using opticalflow technology based on sparse feature points, using adaptive Gaussian mixturemodel to extract foreground as mask, which received only target feature points motionvectors. In order to more accurately define the crowd exception, we usemulti-statistical characterization descriptors which calculate the kinetic energy of thecrowd, the direction of motion direction histogram entropy and mutual information ofthree adjacent frames, characterizating the degree of confusion in the crowd andsports mutations. To make the system capable of identifying the crowd gathered anddeffusion and other specific events, the paper also defines the key parameters of the regional distribution of motion for specific events identifying as an important role toimprove the recognition accuracy.Experimental results show that the proposed method of density analysis andmotion analysis model is simple, lower computational complexity, higher accuracy,robustness and has strong generalization ability. It can achieve a variety of automaticidentification of unusual circumstances, which with important research value.