Detection Of Abnormal Crowd Behavior Based On Crowd Distribution And Kinetic Energy

The particle entropy method to describe the information of the crowd distribution.Firstly the particles are evenly sprinkled on each frame of the video frames,with the flow of particles in a video frame to describe the movement of crowd in the video.The velocity of the moving particles are computed by the optical flow of the pixels around the particles.Those whose velocity are greater than the threshold are defined as the moving particles.Secondly,the moving particles are projected to the horizontal and vertical axes,to calculate the probability of particles distribution in horizontal and vertical direction.The entropy of the particles is obtained by the particles' probability distribution,which is used to describe the crowd distribution information.Finally,the particles kinetic energy is computed by the particles' velocity.Usually the amount of the normal crowd behavior is far greater than the number of the abnormal crowd behavior,and thus the abnormal crowd behavior detection is an imbalance problem.By the advantages of Gauss mixture model(GMM)in dealing with the problem of imbalance,GMM is applied to model the behavior of normal crowd.The training samples used in the modeling phase contains only normal crowd behavior.So,the particle entropy describing the crowd distribution and the crowd motion kinetic energy are used to respectively establish Gaussian mixture model of normal crowd behavior.In the crowd anomaly behavior detection stage,the first step is to extract crowd distribution feature values and the crowd motion kinetic energy feature values.The second step is to use the extracted feature values to calculate the probability on the Gaussian mixture model established in the modeling phase.If the two probability values calculated from the models are both bellowed the threshold,the feature corresponding to the video sequences contains abnormal crowd behavior.Experiments are carried out to detect abnormal crowd behavior on the public available data sets UMN datasets and PETS2009 datasets which contain aggregation and dispersion events.The experimental results show that the algorithm can detect the abnormal crowd behavior under the crowd scenes effectively and accurately.