Crowd Analysis On Video Surveillance Systems

Recent years,as the increasing of large-scale events and big crowd gatherings,the surveillance and management of crowds are becoming a serious problem for the society.Usually,the governors will set security cameras in the key areas to monitor the crowd status in real time.However,most of precent surveillance systems rely heavily on human labour to find out effective information from the surveillance videos,resulting in the process to be tedious and inefficient.Therefore,we hope that the relative knowledge of computer vision can be used to process the images and videos from surveillance systems and then make the process of crowd surveillance to be automatic and efficient.Based on the computer vision technique,the surveillance system is able to describe and estimate the current status of crowd,and furthermore warn the areas which have to be paid close attention to,or analyze the statistic characteristics of crowd over a long period of time.This automatic process is effective to assistant the governor to manage the crowd,divert the crowd,or guide the crowd to avoid possible danger,thus ensure the security of crowd.Although the development of computer vision technique makes the automation of crowd analysis to be possible,the existing crowd analysis algorithms have some disadvantages,which make them not practical enough.First of all,they are not applicable for complex scenes.The real scenes and real crowd behaviors are usually not regular,consisting of structured and unstructured scenes or low density to high density crowd.Secondly,the crowd surveillance application require real-time analysis,while many algorithms are too complex to calculate in real-time.Thirdly,the users need intuitive and semantic analysis results for the assistance of crowd management.So relative crowd analysis algorithms have to extract high-level semantic descriptions and practical informations rather than a simple depiction of crowd.The First part of this paper introduces the general framework for crowd analysis,and summarizes the characterictics of existing crowd analysis algorithms,as well as introducing some algorithms to describe low-level crowd gathering and crowd motion.The algorithms include pedestrian count algorithm and crowd density estimation algorithm.We discuss the characteristics of the algorithms,apply them to real surveillance videos,and then analyze experiment results and propose possible improvement of them.Secondly,based on the discussion of existing algorithms,we propose a spatial-temporal semantic segmentation algorithm and an abnormal detection algorithm based on the calculation of density-combined social force.Both of the two algorithms are based on the estimation of crowd density and description of crowd motion.The proposed semantic segmentation algorithm first generates distribution segmentation and motion group for crowd based on density consistency and motion coherence.Then the temporal results are combined together to get the final semantic segmentation results,while the semantic information of every segmented group is described.On the other hand,the proposed abnormal detection algorithm first derives a new calculation of social force which utilizes the density information.Then,the new calculation of social force is used to calculate the force of every sample point.The distribution of force in a frame is sent to SVM classifier to determine whether it is abnormal or not,while the crowd semantic segmentation results are utilized to detect the local areas of abnormal.Experiments show that these two proposed algorithms are both effective algorithms and applicable to crowd surveillance applications.Finally,based on the algorithms discussed above,we accomplished a crowd surveillance system on the real surveillance platform.The system consists of three crowd analysis functions,including crowd density estimation,crowd semantic segementation and crowd anormaly detection.The results of crowd analysis will be sent into three parts,including the user interface,the database and the socket to sent.By means of various engineering techniques such as cache and parallel caculation,the system is able to work stablely and efficiently on the real surveillance platform and provide assistance to the management of crowd.