Anomaly Detection And Tracking Robust To Occlusion And Imbalanced In Video Surveillance

Anomaly detection is one of the fundemental problems in computer vision,which benefits from the development of Image understatnding,action recoginition,visual tracking and image annotation.We need algorithms to understand video streams and de-tect abnormal events,such as human stampedes and illegal activities.In addition,with the widely-spread use of surveillance cameras in public places and rapidly increasing volume of surveillance video sequences,it is rather urgent for computer vision tech-nology to aid even replace human in supervision and analysis.As a result,anomaly detection in public places has evolved as a key technology.Anomaly detection in the surveillance video refers to understanding normal actions in a certain camera view and detect abnormal events and objects deviating significantly.Since the abnormal activity could be temporally occluded by other objects in the surveil-lance video,the anomaly detection methods might not perform well.Moreover,the imbalance between foreground and background is the instinct of the surveillance video.Such imbalance refers to the fact that most video clips from surveillance camera have an almost static background which lacks information and covers the most area.How-ever,the foreground varies a lot but covers little proportion area.Most existing methods ignore the above issue and treat them equally.As a result,the existing anomaly detec-tion algorithms could not meet the requirement of real application.Focusing on these challenges,this dissertation firstly carries out a research on perceptual constraints based anomaly detection in order to leverage semantic information.Then this dissertation car-ries out a research on anomaly detection framework based on partial convolution and spatial-temporal attention mechanism.Finally,this dissertation carries out visual track-ing research on tackling miss detection in crowded and highly occlusion scenarios.The main contributions and innovations could be summarized in following parts:1.Propose a framework of anomaly detection based on perceptual constraint and residual blocks.To solve the traditional anomaly detection's problem caused by lacking movement and perceptual information,it represents the normal video sequences from frame residual prediction and perceptual loss.Proposed algorithm labels abnor-mal frames by the difference between frame prediction and ground truth.Unlike deep convolutional network,the feature of background and stand-still foreground is passed to deep layer with identity mapping path.The pixel movement feature is modelled by residual mapping path.In order to make frame is consist to ground truth both in the spatial-temporal domain and in the semantic domain,proposed method introduces per-ceptual loss,which provides sharp edge and complicate texture.Extensive experimen-tal results illustrate that comparing to existing methods,the proposed method performs better,and achieves improvement in terms of ROC area.2.Propose a framework of anomaly detection based on partial convolution and attention fusion.Proposed method adaptively estimates object movement by Flownet-like network and predicts current frame with image warping operations.In order to prevent optical flow warp based alignment methods sometimes creating dis-tortions and artifacts,irregular mask and partial convolution operation is provided to eliminate possible noisy pixels.Existing frame prediction methods are usually based on deep convolutional layers,lacking spatial information.When fusing different frames,the reference frame and each supporting frame are probably not fully aligned due to occasional failure of optic flow warp operation.The attention mechanism weights each frame features spatially and temporally,so that the non-occluded area get more atten-tion by the algorithm.Therefor,the predicted frame is reconstructed from much more robust feature maps.Experimental results on several public challenge dataset demon-strate the effectiveness of the proposed algorithm which performs better than state-of-art methods.3.Propose a deep scale feature based single object visual tracking algorithm.To simultaneously leverage deep invariant features and shallow sensitive features,this approach designs two branches on top of deep convolutional network pretrained in an-other domain,such as image classification.The deep branch handles object direction,since it is based on deep convolutional features more robust against to object occlusion,appearance variation and illumination change.The shallow branch handles object scale estimation,since it is based on shallow convolutional features and is masked out some irrelevant feature channels by their feature channel sensitive to current object.Exten-sive experimental results illustrate that comparing to existing methods,the proposed algorithm could handle various challenges,and perform well under several evaluation criteria.4.Propose a single object tracking algorithm based on deep discriminative map.To solve the"ill samples" problem caused by occlusion,it exploits the well-known pyramid structure of deep convolutional feature maps,and formalizes object contour estimation as a heatmap area problem.First,the target is fed into deep neural network,and coarse shape map is constructed from class activation map.When occlu-sion occurs,class activation map would break into pieces and model parameter online update should be stopped.Extensive experimental results illustrate that comparing to existing methods,the proposed algorithm could handle occlusion challenges,and could achieve improvement under several evaluation criteria.