Research On Real-time Object Tracking Algorithm In Complex Environment

With the development of computer,image processing,artificial intelligence,pattern recognition and electronic technology,the application of object tracking technology is more and more.Object tracking has been widely used in the field of intelligent human-computer interaction,intelligent monitoring and unmanned driving.It is a difficult task to design a robust method for the object tracking system due to the influence of illumination change,object scale change,occlusion,fast motion and low resolution.Aiming at these difficulties,we studies two kinds of strong real-time object tracking algorithm: Real-Time Compressive Tracking(CT)and High-Speed Tracking with Kernelized Correlation Filters(KCF).In view of the shortcomings of these two algorithms,the corresponding improvement strategy is putted forward in this thesis.In this thesis,we first analyze the five components of object tracking algorithm:feature extractor,observation model,motion model,updating model and ensemble post-processor.We analyze the functions of each part,the realization principle and the commonly used model,and summarize the basic flow of the whole target tracking algorithm.Then we introduce the CT algorithm in detail.In view of the problem that the CT algorithm can not effectively deal with the scale change,this thesis proposes to adaptively change the scale size of the acquisition frame and the relative position of the rectangular filter.The CT algorithm can not effectively deal with the occlusion,so we used Bhattacharyya coefficient for online feature selection and build adaptive learning factors to update the classifier.In the OPE evaluation,the average accuracy of the SCT algorithm is 57.5%,which is 11.43% higher than that of the FCT algorithm.The average success rate is 42.6%,which is 11.52% higher than that of the FCT algorithm.In view of the problem that the KCF algorithm can not effectively deal with the scale change,we propose a new target appearance model Rs by using the original KCF algorithm to construct the scale pyramid and use Rs to find the scale of the current frame.KCF algorithm can not effectively deal with the occlusion,out-of-view and other issues,we use the random fern classifier to build re-detection module,and use the target appearance model Rs to accurately locate.In the OPE evaluation,the average accuracy of the LST algorithm is 85.1%,which is 13.77% higher than that of the KCF algorithm.The average success rate of the LST algorithm is 62.8%,which is22.42% higher than that of the KCF algorithm.Finally,this thesis summarizes the research content of the object tracking algorithm and the deficiency of the new algorithm,and looks forward to the future research direction.