Research And Application Of Target Tracking Algorithm Under Occlusion

As a research hotspot in the field of computer vision,visual target tracking has been widely used in the field of intelligent surveillance and man-machine interaction military strike,and has a wide range of application value and development prospect.Due to the complexity of the tracking environment,factors such as occlusion,scale variation and fast movements will affect the performance of the algorithm.This article focuses on solving the problem of occlusion in target tracking.It focuses on the research and analysis of tracking algorithms based on kernel correlation filtering,and improves the existing problems.The main contributions of this paper are as follows:(1)Aiming at the problem of performance degradation caused by partial occlusion of the target in the tracking process,a multi-feature fusion scale adaptive block tracking algorithm is proposed based on KCF.First,divide the subblocks from the target center,use a local kernel correlation filter that fuses the gradient and color features to track each target sub-block separately,and combine the position constraint relationship between the target sub-block and the whole to get a rough estimate of the target center position.The global filter is used as an initial estimate to determine the precise position of the target center.Secondly,the scale of the target is calculated by the change of the position of the corresponding sub-block between two adjacent frames.Experimental results show that compared with the original kernel correlation filter algorithm KCF,the overall distance accuracy is improved by 7.4%,and the overlap accuracy is improved by 7.8%.(2)Aiming atyingwen the tracking problem when the target is completely blocked or moved out of the field of view during the tracking process,a long-term tracking algorithm based on model update and target re-detection is proposed.First,calculate the confidence of the tracking result of each frame.When the tracking failure is detected,the target re-detection module is started,and the candidate target set is obtained in the re-detection area through the normalized product correlation algorithm(NCC).Then use the global filter to calculate the candidate target response one by one and select the reliable target from it as the re-detection result.At the same time,the learning rate of the classifier is adjusted according to the tracking confidence.Experimental results show that compared with the original kernel correlation filter algorithm KCF,the final algorithm in this paper improves the overall distance accuracy by 11.7% and the overlap accuracy by 16.2%.(3)In order to verify the application value of the proposed algorithm,it is applied to the UAV target tracking process.First,the algorithm of this paper is simulated and tested on the UAV123 dataset.Then the image captured by the fourrotor UAV is transmitted to the Matlab platform for processing.Experimental results show that the algorithm in this paper can accurately track the moving targets on the ground,especially in the scene of occlusion and scale changes.The algorithm has certain application value.