Research On Target Detection Algorithm In Intelligent Video Surveillance

Intelligent video surveillance is a hot research direction in the field of computer visionrecent. It focuses on how to describe, distinguish and understand the contents of thesurveillance video by using computer vision technology, image and video processingtechnology and artificial intelligence techniques, and further control the video surveillancesystem according to the distinguish results, so that making the video surveillance system has ahigher level intelligence. It can be applied into security, bank, transportation, militaryfacilities and other fields, having great theoretical value and broad application prospect.The main research contents of intelligent video surveillance technology include: movingtarget detection, tracking, classification and behavior analysis. Among them, the targetdetection results will directly affect the effect of the rest modules, and the difficulty of targetdetection lies in the complexity of background, incompleteness of detection algorithm, andshadow of moving target and so on.The paper first introduces the application of intelligent video surveillance system, thenreviews the principle of three classic moving target detection algorithms: frame differentialmethod, optical flow method, background subtraction and compares their pros and cons.Among them, the codebook method and its improvements in recent years receive a moredetailed introduction.Thinking of the traditional codebook algorithm can’t well adapt to the illuminationchange, the paper puts forward a new codebook algorithm which processes brightness andcolor saturation information separately in YUV color space, and combines with LBP texturematching. In addition, the new algorithm uses color invariance which is not sensitive to lightand similar color background to detect the main outline of moving targets. Finally, the detectresults will be combined together in order to meet the effective robustness of target detection.In order to avoid the influence of shaking background, LUC and shaking measurement areproposed in the algorithm. Historical LUC is used for calculating shake measurement whichrepresents shaking level of each pixel, and then cluster the shake measurements to furtherdivide moving target, statistical background with the swing leaves. Experimental results show that using the detection method proposed in this paper, better effect has been achieved, thatmeans less noise would be found and targets appearance are more complete during theforeground detection.