Research On Target Tracking Algorithm Based On Superpixel Segmentation

Object tracking is one of the key research in computer vision,which has showed great practical value in a wide range of fields such as autonomous driving and intelligent traffic surveillance.Currently,most tracking algorithms are based on discriminative correlation filter.However,these algorithms remain to be further examined in some tricky tracking conditions.Meanwhile,multi-target tracking requires the algorithm to track a good many of targets simultaneously,which cannot be achieved merely by a target detection algorithm or a single target tracking algorithm.Therefore,a multi-target tracking algorithm with high robustness is required.Considering the above-mentioned points,this thesis proposes a superpixel-segmentation-driven tracking algorithm based on correlation filter in order to solve the traditional correlation filter algorithms' poor tracking performance in rotation,scale variation and other situations.The proposed algorithm processes input video sequences with superpixel segmentation and establishes a superpixel-based appearance model which can effectively capture the appearance change and solve the problem of scale variation when target tracking.To handle the target rotation problem,we adopt a combination of PHOG and CN features with noise and rotation resistance.Appling the algorithm to OTB-2015 data set,the AUC index can achieve 65.3%,which is 21.3% higher than the AUC of traditional KCF algorithms and also exceed ECO-HC(64.7%).Comparing to other algorithms,the proposed algorithm also shows a better performance in tracking accuracy with an outcome of 0.805.According to the experimental results,the proposed algorithm can not only meet the real-time tracking requirements,but also improve the tracking accuracy and effectively solve the problems of rotation and scale change.Aiming at the disadvantages of the correlation filter tracking algorithm in which the samples are easily tainted and the fast moving targets cannot be handled,this thesis puts forward a multi-target tracking algorithm on the basis of Markov Decision Process.In this algorithm,each target is modeled as a Markov Decision Process and transitions between states are driven by maximizing the reward function.In addition,the algorithm adopts reinforcement learning drill in similarity function of data association,which has effectively solved target occlusion.Besides,to eliminate the target loss due to fast movement objects in tracking algorithms,superpixel is used to establish an appearance model which takes full account of the historical image information and improves the accuracy and reliability of the tracking algorithm.The experimental evaluation shows that the tracker has a good performance in the MOT15 data set.The tracker proposed in the thesis achieves 36.5 in MOTA index,which is evidently higher than other algorithms,while the ID switch index is only 308 times,which is lower than other compared algorithms and significantly reduces the target loss rate and identity exchange rate.