Objects Recognition And Unusual Events Modelling&analysing In Intelligent Video Surveillance

Intelligent Video Surveillance is an application field of computer vision that attracted much attention in recent years. It can process, analyze and understand video signal and control the whole system using computer vision and image processing techniques. For example, a surveillance system can recognize different objects automatically, or detect unusual events in the surveillance scene. Then it provides the helpful information or alarms to surveillants. This can help the surveillants solving urgent problems quickly and effectively. Additionally, intelligent surveillance system can filtrate plenty of redundant information, only preserve key information. Intelligent video surveillance techniques can solve the problems such as huge processing data problem, long response times and human inherent weaknesses, which often occur in traditional video surveillance systems that make the work tedious and low effective.Intelligent video surveillance is an integrated application system that covers many research fields, including image processing, computer vision, pattern recognition, artificial intelligence, communication technology and computer networks. In this paper, we focus on one application field of intelligent video surveillance which is unusual event detection and recognition, and its relevant specific techniques. The main research problems and contribution are:1,Researched and implemented some basic but important techniques in intelligent video surveillance system. Adaptive Gaussian mixture model was used to extract moving objects. It had the advantages that restrained noises and promoted extracting effect. Mean-shift tracking algorithm was used to track objects, avoided the objects conglutination problem in moving detection. Multi-sub-block gray correlation match algorithm was used to solve the lost or confused tracking problem. The well done of these works promoted the credibility of the following works.2,Built a training database that contains more than seventeen thousands samples. It had three types of moving objects'profile image that were pedestrian, vehicle and the others. This database could be used for feature extraction of moving objects and classification researches.3,Researched the object detection and recognition problems thoroughly. They were divided into two research aspects which were static objects recognition and moving objects recognition. The research history were summarized and compared. The application of dimensionality reduction algorithms in pattern recognition is researched. We found that graph embedding can reflect the intrinsic feature of dimensionality reduction algorithm better. Based on the database we built and marginal fisher analysis, we proposed a general moving objects recognition framework, avoided the rigorous feature extraction process and the limitation of object types in traditional methods. The system has advantages of generality and effectiveness.4,Discussed and researched a new researched field in intelligent video surveillance called unusual event detection in videos. We analyzed and summarized the intrinsic feature of unusual events. Through the definition of behavior features and uniform modeling methods, we proposed an unusual event detection framework which can be used both in supervised mode and unsupervised mode.5,Trajectories of moving objects was one of the most important features in behavior features. In this paper, we proposed a new moving object trajectory similarity metric called Sectional Contextual Edit Distance (SCED). It added the difference information of moving objects'position, direction and speed into the penalty function to reflect the moving feature of different moving phases. SCED algorithm had the advantages that can provide more reasonable similarity metric of moving trajectories and had more effective computation ability.6,Based on the principle of Dynamic Bayesian Network, we proposed a new model called Sequential Multi-Layer HMM to model the sequential causal relationships of complex events and recognized unusual events. Events were expressed with the least state relationships. The expression and inference of model parameters were optimized; the complexity of computation was reduced and the system compatibility to noise was promoted based on this model.