Research On The Target Tracking Of Soccer Robot In Complex Background

With the development of science and technology,machine vision based target tracking system is becoming more and more important for high-quality life.When applying such techniques into applications,however,the existence of complicated and dynamic background brings in severe mutual restrictions between accuracy and realtime requirements.As a result,it is of great significance to figure out robust target tracking models to deal with the impact from uncontrolled background.The thesis focuses on generating solutions to such problem by designing goalkeeper robot system,and the main contributions could be summarized as follows:To design effective binocular vision system,detailed theory analysis is carried out followed by system construction.To solve the imaging problem with the system reflecting in image distortion,three main distortion occurrences are modelled and corrected by specifically designed methods.In the procedure of stereo matching,stateof-the-art algorithms are illustrated and compared,based on which the robust matching model based on LoG operator and structure tensor analysis is proposed.Regarding the design of robust target tracking algorithm dealing with the influence of complex and changing background,an improved optical flow model is proposed for accurate and real-time target tracking.To improve the speed of the model without affecting the accuracy greatly,the improved particle filter(PF)algorithm by incorporating Mean Shift for clustering is introduced.Apart from adopting Mean Shift for clustering,the Gaussian model optimized with maximum entropy and particle resampling strategy is realized to deal with the existence of invalid and degrading particle in traditional PF model,and the real-time requirement of target tracking task is solved simultaneously.To verify the superiority of the optimized algorithm for robust tracking and binocular vision system for high-quality imaging,the goalkeeper robot system is designed and tested under different conditions,and satisfactory results in terms of speed and success rate fully demonstrates the which fully demonstrates the effectiveness of the algorithm and imaging systems.Besides,we also argue that the proposed models could also be adopted to solve other machine vision problems.