Fast Target Tracking Algorithm In Complex Scenes Based On Spatial Information Minin

Video object tracking is one of the main research directions in the field of computer vision and has been widely concerned by researchers at home and abroad.The basic task of object tracking is to accurately predict the state of the object in subsequent video frames given the object information in the first frame.Over the past decades,the overall performance of object tracking algorithm has been excellent,but the problem is that the algorithm complexity increases,which seriously limits its practical application.At the same time,most of the tracking algorithms pay more attention to the target area information and ignore the important background information,and the performance of these algorithms is still not satisfactory for the complex scenes,such as similar background interference,fast movement and occlusion.Therefore,how to improve the tracking performance of target tracking algorithm in complex scenes and balance the contradiction between robustness and real-time performance of the algorithm is still an urgent problem to be solved.In this paper,the main research efforts are to achieve fast target tracking in complex scenes.For the above problems,the main research work is summarized as follows:(1)A high-speed spatial constraint(HSC)is proposed: spatially regularized discriminative correlation filters(SRDCF)introduce spatial regularized weights to alleviate the boundary effect caused by cyclic convolution and showed excellent performance in solving similar background interference and fast motion problems.However,the optimization cost of spatial regularization is high,which limits the real-time performance.To solve this problem,a highspeed spatial constraint discriminant correlation filter(HSCDCF)is proposed.By introducing a posterior space constraint to punish the coefficients of the correlation filter,and simplifies the optimization process,the boundary effect caused by cyclic convolution is effectively alleviated,and the tracking algorithm is more efficient.On OTB-2013 and OTB-2015 benchmarks,the performance of HSCDCF in accuracy and success rate was slightly reduced by 2.7% and 3.1%,respectively,compared with SRDCF.In terms of speed,HSCDCF achieved a real-time speed of 62.5FPS,10 times faster than SRDCF,and achieved a good balance between real-time performance and robustness.(2)A similar semantic Distractors Assisted object tracking algorithm(DATracker)is presented.Currently,most object tracking algorithms only use appearance models to locate object,however,these algorithms tend to lose the target in complex scenarios with similar semantic distractors and rapid appearance changes.It is found that some prominent objects in the scene can also serve as anchors for locating targets.Based on this finding,this paper proposes a target tracking algorithm assisted by similar semantic distractors.Firstly,target and similar semantic distractors are detected simultaneously,and the relative position relationship between target and distractor is encoded by Kalman motion model cluster to eliminate the influence of camera motion and convey these important spatio-temporal information in video sequence.In the detection stage,the relative position information of the predicted target is matched with all candidate targets,and the target with the highest matching degree is the target.The proposed model is simple,efficient and has strong applicability.On the La SOT benchmark dataset,the proposed algorithm achieves a 1.2% improvement in the success rate compared with the basic algorithm,which verifies its effectiveness.To sum up,by fully mining the spatial information in the scene,this paper effectively solves the complex scene problems such as similar background interference,fast motion and occlusion.At the same time,the algorithms proposed in this paper maintain the real-time speed and can well balance the real-time performance and robustness of the algorithm.