Research On Abnormal Event Detection In Crowds From A Moving Vechile

Intelligent video surveillance system plays an increasingly significant role in social life.With the cost of the monitoring equipment like cameras have fallen and the continuous developments of computer vision and machine learning,intelligent visual surveillance system has influenced people's lives.For example,many communities,schools,squares,bus stations and other public places have installed a video surveillance system with which the video analysis algorithm can analysis the video content,recognize the action and detect the abnormal behavior(e.g.,falling,fighting,robbing and other behavior)to supplement or even replace the traditional human security system.In the community outdoor environment,many cameras have been fixed in main roads,entrances and other public places to monitor abnormal event for better security.However,in some areas such as shade or turning corner,coverage monitoring can not be achieved due to the limited monitoring distance and angle constraints of fixed camera.When some abnormal behaviors(e.g.climbing,falling.)occur in these areas,it would cause serious property damage and personal injury for uneffective monitoring and belated alarm.In addition,the occurrence of abnormal behavior is a continuous and moving process,how to trace the center of the event is another challenge.In order to monitor the blind spot under the fixed camera and tracke the incident centroid,a rapid detection of abnormal behavior system based on mobile robot is proposed to detect abnormal event.The system can be embedded in a patrol robot which will be used in the community environment to shoot the real-time scene information by vehicle-mounted camera,then to analysis of the monitoring area through intelligent visual detection algorithm.If abnormal event happens,alarm will be triggered to notice the guard men.We have carried out our work around the detection of abnormal behavior from three main aspects:(1)Real-time running detection system based on a mobile vechile.Usually in the abnormal behavior,pedestrians do not constitute the focus of abnormal events,but the person of fast running.Running events is often accompanied in most abnormal behavior.Therefore,the detection of running targets is to prevent and predict the occurrence of abnormal events.Real-time and high detection accuracy are mainly considerated in the system.Firstly,a mean filter is used for image passivation and fast optical flow is calculated for movement information,and then closing operation in morphology are adopted to filter the background information.Finally,the contours of candidates are extracted as an input into the convolutional neural network(CNN)to recognize the running targets.Due to the influence of many factors,such as illumination,background and non-rigid body,it is a challenge to detect the moving targets in the outdoor environment with high detection accuracy.A two-stream CNN combining human body representation and motion information are used as a classifier to recognize running target.In order to train the network model,30 thousand running pictures as positive sample are manually collected.Through the outdoor test,it proves that the system achieves a real-time detection speed at 20 fps and 85.6% of the running detection accuracy.(2)Fast action localization based on spatio temporal path search.In video analysis,action recognition is a common and basic task,the purpose of which is to classify the video;action detection is an advanced task for it not only should identify the action type,but also localize the actor(s)spatially and temporally,i.e.,action localization.Fast and effective extraction of complete and coherent action is an important manifestation of intelligent video analysis.We intend to adopt a method similar to traction by detection.Firstly,a fast and effective body contour detector,Faster-RCNN is adopted,then each candidate is scored by a scoring function combined with the temporal continuity of the prior probability search and detection based on posterior probability confidence.Finally,the problem of action extraction is formulated as a max-path finding problem,and the dynamic programming algorithm is used to find the maximum value path.Experiments on UCF-Sports show that the method is effective for action localization.(3)Video sequences and behavior analysis based on LSTM.In general,crowd abnormal events are classified into two categories: individual abnormal events and group abnormal events.Individual incident refers that the goal of individual behavior is different from most of the behaviour of other individuals.While group abnormal events mean that the individuals or objects in the monitoring area is quite different from the usual group behavior.Long term memory network(LSTM)is a kind of special time recurrent neural network.For its gated network structure to select and forget the timing information,it is suitable for processing and predicting of time-series behavior or events.Abnormal behavior is a continuous process,we use the improved LSTM network based on mobile robots to detect some abnormal behavior(such as falling,climbing or fighting).