Research On Object Tracking Techniques Based On Moving Platforms

To achieve moderately accurate and efficient tracking,traditional algorithms typically assume that object's displacement and visual characteristics change uniformly across adjacent frames.These trackers,however,usually fail with high-speed vision platforms like UAV(Unmanned Autonomous Vehicle)and USV(Unmanned Surface Vehicle)where object's displacement and visual characteristics change irregularly or even fiercely.To address the displacement change,we propose a cascade tracking algorithm with a coarse-to-fine strategy for larger search area and a keypoint-matching-based tracking algorithm with adaptive full image searching.To address the visual characteristics change,we fuse long-and short-term information within deep models and design an improved adaptable tracking algorithm.The main contents of this paper are as follows:We propose a cascade coarse-to-fine tracking algorithm to address the background clutter which degrades traditional algorithms when searching large area.Our tracker first adopts a correlation filter to enlarge the search region(6 times the size of the target)and searches roughly for potential regions containing targets.A structured support vector machine(SVM)based tracker is introduced to center on the maximum response position within these regions and search within the resultant finer local area.Both two trackers are online learned.Consequently,our cascade tracker enlarges the search area from 2.5 times(typically adopted in traditional algorithms)to 6 times the size of the target to show improved adaptability under moving platforms.To cope with substantial movement of vision platforms,we propose a keypoint-matching-based tracking algorithm which adaptively searches over full image.Our tracker combines the advantages of full image searching of keypoint matching and local tracking of correlation filter.Besides,it innovatively judges abrupt motion based on tracking response map.This judge mechanism greatly enlarges search area but saves much computation.We also propose a confidence discrimination mechanism for online updating to avoid invalid update faced with occlusion and obtain a more robust tracker.Mainstream deep learning methods may easily fail to adapt to target dynamics without taking in any real-time(short-term)information.We accordingly propose to fuse long-and short-term information within the same deep model.This fusion considers both originally reliable and spatially,temporally changing information of the target,which can adapt to target's fast visual characteristics change across frames.