Research On UAV Long-term Tracking Algorithm Based On Re-detector And Bounding Box Refinement

With the rapid development of unmanned aerial vehicle technology,it has been widely used in many fields,among which the visual target tracking technology of unmanned aerial vehicle has become one of the research directions of great concern.However,in complex environments,unmanned aerial vehicle visual target tracking is affected by factors such as illumination changes,camera motion,partial occlusion,and scale changes,resulting in reduced tracking performance.Aiming at the above problems,this paper proposes a long-term tracking algorithm for aerial photography targets in UAV scenes.The main contributions are as follows:(1)Aiming at the problems of boundary effects and filter degradation in correlation filter tracking algorithms in unmanned aerial vehicle scenarios,a correlation filter tracking algorithm based on adaptive spatiotemporal regularization is proposed and used as a basic tracker.Firstly,an adaptive spatial regularization method is designed in the correlation filter to alleviate boundary effects.Secondly,a time regularization method based on high confidence samples and an optimization update strategy are designed to prevent filter degradation.Then,ADMM is used to optimize the solution of the objective function to ensure the operational efficiency of the algorithm.Finally,a fast scale filter is designed to estimate the target scale.Experimental results show that the proposed basic tracker effectively improves short-term tracking performance.(2)Aiming at the problem of tracking failure when the basic tracker faces long-term tracking scenarios such as occlusion and moving out of view,a long-term tracking framework for unmanned aerial vehicles based on local-global region re-detection is proposed.Combining with the basic tracker,a tracking state judgment module and a re-detector are designed.The framework performs target tracking through the basic tracker,and uses the tracking status judgment module to determine the tracking status of the basic tracker.In the event of a tracking failure,the re-detector is started,while the recovery tracking of the target is completed through the re-detector.The re-detector consists of two parts: local area redetection and global area re-detection,which can recover tracking of the target when it reappears after being occluded or moved out of view.Local area re-detection is completed by searching within the local area of the target.If the recovery target tracking fails,the recovery of the target tracking state is completed through global area re-detection.Experimental results show that the proposed algorithm can better cope with long-term tracking of complex scenes such as occlusion and moving out of view while maintaining real-time performance.(3)Based on the long-term tracking framework composed of the basic tracker,tracking state judgment module,and re-detector described above,a bounding box refinement module is designed,and a long-term tracking framework for unmanned aerial vehicles that combines re-detector and bounding box refinement is further proposed.The tracking bounding box obtained by the basic tracker or re-detector is refined through the bounding box refinement module,improving the quality of the predicted target bounding box.Experimental results show that the proposed algorithm can improve the success rate and accuracy of tracking,and has good long-term tracking performance while ensuring real-time performance.