Moving Object Oriented And Patch Classifier Based Anomaly Detection In Crowd Scenes

With the progress in technology and the popularity of"smart city" as well as the development of "city brain",city surveillance is becoming more automatic and intel-ligent.And this requires computer vision techniques to solve problems in traditional surveillance videos,such as monitoring by human beings,storage redundancy,hard to retrieval,can't analyze videos automatically and so on.Anomaly detection is one of the sectors in computer vision.Since it can monitor videos and detect anomalies auto-matically to protect human from dangers,it is of great importance to the development of city security and captures much attention from both academic and industry.In crowd scenes,monitoring images captured by surveillance cameras can be very complicated,which is mainly reflected in:1)the monitoring images can be very dim because of weather,timing,which may make it difficult to distinguish objects and back-ground.2)Anomalies are highly-scene-related and vary a lot when scenario changes.Besides,the cameras are designed with wide-angle,which may cause depth of field problems.These problems bring a lot of challenges to anomaly detection:1)Ordinary object detection may miss some targets that are similar to background,which has an impact on object state analysis in later stage.2)Scenarios captured by surveillance cameras with wide-angles vary a lot.Anomaly varies in different positions,so single feature or classifier may lead to inaccuracy in anomaly detection.Aimed at those chal-lenges,this thesis solves the problem of how to extract processing units with semantic understanding of scenes to make the best of spatial and temporal information,and how to detect anomalies precisely while anomaly varies when position changes in one frame captured by cameras with wide angles.The main work and innovations are as follows:1.A moving object oriented anomaly detection algorithm in dim crowd scenes is proposed.Since processing units utilized in most anomaly detection methods lack semantic understanding of scenes,which may split targets into pieces and limit the per-formance of anomaly detection.To handle this problem,this paper proposes to extract moving objects with deep object detection algorithm as processing units.To improve performance for object detection in dim scenes,we fuse motion and appearance infor-mation to get dynamic image,which can make objects more distinguishable.Based on detected objects,we develop an effective and scale-insensitive feature,named his-togram variance of optical flow angle to find abnormal motion pattern.Meanwhile,this paper models the background in scenarios to get non-active area to find location anoma-lies.The experiment result shows that the proposed algorithm can detect and localize anomalies precisely when compared to state-of-the-art methods.2.A multi-feature and patch classifier based anomaly detection algorithm in crowd scenes is proposed to deal with the difficulty that anomaly varies when location changes in one frame captured by cameras with wide angles.To handle this problem,patch classifier is proposed to improve the performance of anomaly detection.Specifically,frames are divides into patches and each patch will learn its own normal pattern.Short-term as well as long-term motion features are utilized for detecting motion anomalies.Lastly,a new object re-targeting method is proposed to find the abnormal objects missed by occlusions in detection.The results show that the proposed method can achieve comparable accuracy in anomaly detection with state-of-the-arts methods and at the same time,localize anomalies precisely.