Research And Application Of RGBT Target Tracking Algorithm Based On Correlation Filter

RGBT target tracking based on correlation filtering framework is a hot research topic in recent years,which combines visible(RGB)and thermal infrared(T)image information and has a broader application scenario and better tracking performance than unimodal target tracking.The classical correlation filtering framework uses constant weights for RGBT feature fusion and filter update,which does not take into account the unique advantages of visible and thermal infrared features in different situations,and is prone to filter degradation and thus drift in complex environments.To address these problems,this paper proposes an adaptive RGBT feature fusion and filter update based on the background-aware correlation filtering framework,and proposes the AFRTB algorithm.In order to apply the algorithm to the Electro-optical targeting system(EOTS)scenario,we propose a camera motion detection mechanism(MD)for the camera motion problem of the EOTS,which improves the feasibility of the AFRTB algorithm in the EOTS scenario.The main work of this paper contains the following two aspects.(1)The AFRTB algorithm is proposed to address the problem that the constant hyperparameters in the traditional correlation filtering framework affect the tracking performance.In this paper,we first introduce the background-aware framework BACF to weaken the boundary effect of the correlation filtering algorithm,and train independent one-dimensional scale filters to achieve fast multi-scale fitting.Then the response maps of visible and thermal infrared features are adaptively weighted fused according to the KL scatter to assign higher weights to the better-performing features in different scenes,while the fused response maps are used to make confidence judgments based on the improved APCE criterion to achieve a more robust filter update.The experimental results of the last two public RGBT target tracking datasets show that the tracking accuracy and success rate of AFRTB are higher than all the compared algorithms.(2)In the EOTS scenario,the camera motion detection(MD)mechanism is proposed to fuse with the AFRTB algorithm for the problem of tracking failure due to camera motion.The images taken when the camera moves will be blurred.In this paper,the image entropy is used to detect the blurring degree of the image,and when it is determined that there is image blurring,the adaptive number of corner point matching is used to calculate the camera motion vector and adjust the position of the filter search area according to this vector.Meanwhile,in order to simulate the tracking characteristics of the EOTS more accurately,this paper makes a home-made RGBT target tracking dataset EOTS-RT,and then compares the AFRTB+MD algorithm with other advanced algorithms in this dataset.The experimental results show that the camera motion detection mechanism significantly improves the tracking performance of the tracking algorithm in the case of camera motion and image blur,and the feasibility and robustness of the AFRTB+MD algorithm in the EOTS scenario have obvious advantages.