A Fast Long-term Object Tracking Algorithm Based On Deep Learning For Local-Global Combination

Object tracking,as a research branch of computer vision,has a wide range of application prospects in multiple fields.Object tracking can be divided into short-term object tracking and long-term object tracking,among which long-term object tracking is closer to practical applications compared to short-term tracking,while facing more complex tracking challenges.However,there is limited research on long-term tracking algorithms at present,and there is still room for optimization in terms of performance and real-time capability.This thesis focuses on the tracking challenges of local similar object background interference,inaccurate bounding box positioning caused by object deformation,and object disappearance and reappearance in long-term tracking scenarios,and the main work includes the following three aspects:(1)To address the complex challenges in long-term target tracking tasks,this thesis proposes a deep network-based fast long-term object tracking algorithm that combines local and global search strategies.The local search is used to track the target in a local area,verify potential object,and further refine the bounding box position.The global search is used to re-detect lost tracking object.Both local and global searches are trained offline and can be used directly during the tracking process.The proposed long-term tracking framework is concise and efficient.(2)In the local search,this thesis proposes a lightweight verifier that can resist interference from similar objects and introduces a boundary box refinement module to address the problem of imprecise location caused by object deformation or occlusion in long-term tracking scenarios.By combining the temporal context position information of the tracking target with the real bounding box of the template frame,the proposed method prevents tracker drift and achieves high-quality feature representation for tracking results on a per-pixel basis,thus improving tracking accuracy.Experiments show that the proposed efficient local search algorithm has good adaptability to challenges such as interference from background similar objects and target deformation.(3)In the global search,this thesis proposes a fast re-detection algorithm based on a lightweight attention to address the problem of object disappearance and reappearance in long-term tracking scenarios.This method uses a lightweight attention to learn appropriate target representation features,enabling it to quickly and effectively estimate the potential location of lost targets.The global re-detection algorithm aims to find the most likely potential location of the target,and then sends the result to the local search for accurate tracking and positioning.The introduction of a lightweight attention network reduces the number of parameters and computations,enabling the global re-detection algorithm to run quickly and effectively.The proposed local and global search algorithms are combined to form a local-global combination fast long-term object tracking algorithm,which is compared with popular long-term tracking algorithms on mainstream large-scale long-term target tracking datasets.The experiments show that the proposed long-term tracking algorithm has good adaptability and real-time performance in dealing with challenges such as repeated target disappearance and reappearance,interference from background similar objects,and target deformation.