Research On Target Detection And Tracking Algorithms Based On Vision

Video surveillance technology have been developed rapidly in just two decades, it changedfrom the first generation of analog monitor in the early year to the second generation of digitalmonitor in recent year, and to the networking video surveillance now. Recently, researchercombined computer vision technology into the video monitoring field, and formed intelligent videosurveillance technology, thus fulfilling the video monitoring unmanned and automated. Intelligentvideo surveillance technology can detect, localized and track the target on the scene in real time,and through the analysis and understanding of the target’s behavior to achieve prediction. As acutting-edge research field of computer vision, intelligent video surveillance has become a focus inacademic and engineering, and has been successfully used in security, commerce, transportation,military, aerospace and other fields. Moving target detection and object target tracking is thefoundation the follow-up of high-level video image processing, understanding and application, andthe key technology of intelligent video surveillance.Based on the in-depth analysis of various moving target detection algorithm, a kind ofbackground subtraction method was proposed, which base on hierarchical codebook model. In theway of model construction, the hierarchical codebook model was based on Gaussian pyramid image.Results of the pyramid top-level image only with the outline of an object, which can be moreeffectively for denoise, while results of the pyramid bottom-level image contains more details,which can extract more precise contour of foreground. In the way of model description, the originalcode word was a cylinder in RGB color space. This paper presented a simplified model, which is abox in YUV color space.Based on the analysis of various objects tracking algorithm, Firstly, a kind of particle filteralgorithm which based on CenSurE feature point was proposed. The similarity between targetmodel and candidate model was measured by the combination of feature point’s number and theirrelative coordinate. The feature points were selected in the smallest rectangle area which onlycontaining the particles set, and using KNN algorithm to accelerate the matching process. In theprocess of tracking, dynamically adjust feature points of the target model to adapt to the changes oftarget appearance. And then, a kind of particle filter algorithm with adaptive combination of colorand CenSurE feature point information was proposed, which is to meet the problem low accuracyand might be failed in complex scene of single information based tracking algorithm. Finally, inorder to prevent degradation of multi-feature tracking algorithm into a single-feature trackingalgorithm, the credibility of each feature was analysed in detail, and add the results of the foreground subtraction as feedback to the tracking result.The experimental results and algorithm performance analysis show that the backgroundsubtraction method which based on hierarchical codebook model can deal with complexbackground situations, and the accuracy is better than Mixture Gaussian Model and the otherCodebook Model, and it can meet real-time requirements. The CenSurE feature points basedparticle filter tracking algorithm can achieve stable tracking in case of failure of obtain target’s colorinformation. The adaptive fusion of multi-feature tracking algorithm with feedback of backgroundsubtraction can meet high accuracy requirement in complicated scene.