Research On Object Tracking Based On Correlation Filters In Complex Background

The rapid development of computer vision technology is widely used in daily life.Object tracking,as an important branch of computer vision,has attracted the attention and research of a large number of scholars.Object tracking has shown strong advantages in many fields such as intelligent surveillance,autonomous driving,medical imaging,and military affairs,and related algorithms have been endlessly produced.In the early generation of generative model tracking algorithms and the recent development of discriminative model tracking algorithms,the object tracking technology continued to break through in performance.The latter(that is,discriminative model tracking algorithm)includes a correlation filter tracking algorithm that takes both speed and accuracy into account,and has attracted much attention.Although the correlation filter and other object tracking algorithms show good tracking results,the tracking results for complex background situations,that is,involving multiple scale changes,lighting changes,target deformations,background clutter,and fast motion,are still not ideal.Inspired by the discriminative scale space tracker(DSST)and spatial regular correlation filters(SRDCF),this paper starts from the perspective of multi-correlation filters detection fusion and multi-sample combined training,and seeks breakthroughs in the consideration of multiple scenarios.The main work includes:(1)Aiming at scenes such as fast motion and background clutter,this paper proposes a mask-based correlation filter tracking algorithm.First,for the boundary effect caused by the cyclic shift of the correlation filter,this paper processes the trained correlation filter according to the mask matrix extracted from the first frame of target information.Secondly,for the problem that the target drift is easy to occur in the background clutter scene,this paper uses the mask to process the samples,and then uses the mask-processed samples to train the correlation filters.Finally,the detection response map of the two correlation filters are linearly combined,and the maximum response position is used as the estimated target position.(2)In order to make the spatial regular correlation filter tracking algorithm suitable for the tracking environment of a variety of scenes,this paper proposes a correlation filter tracking algorithm that is extended to multi-scene boundary effect suppression.First,considering the problem of poor tracking effect of spatial regular correlation filter tracking algorithm in multiple scenes such as background clutter and occlusion,this paper proposes a tracking component adapted to multiple scenes,and integrates this component into the spatial regular correlation filter tracking algorithm to train correlation filters.This paper analyzes the two methods mentioned above on the OTB-2013,OTB-2015 datasets and several current tracking algorithms with better tracking performance,including accuracy,success rate,and average overlap rate.Experimental results show that the two target tracking methods proposed in this paper are feasible and have good accuracy for tracking scenes such as fast motion,background clutter,object deformation,and occlusion.