| At present,object tracking technology based on visual image has a broad application prospect in military guidance,visual navigation,security monitoring and medical diagnostics,and many other fields,which has played an important role on improving people’s lives,enhancing national defense strength,and promoting social progress.In this paper,with the subject of continuous online single-target tracking on video sequences,we studies and analyzes the framework of TLD(Tracking-LearningDetection)algorithm for object tracking.The traditional object tracking method can’t achieve online object tracking for a long time under complex background when objects undergo appearance variations or occlusions,including Recursive Tracking and Tracking-by-Detection.Therefore we research a novel online object tracking approach--TLD,which only need the initial state of object in the first frame.Object tracking method based on TLD mainly consists of four parts,namely tracking,detection,synthesizer and learning.Tracking estimates the current state of object according to the target trajectory for recursive tracking.Detection employs an exhaustive search in order to find the object by the sliding-window approach.The subwindows are tested through a cascade mechanism which contains variance filter,ensemble classifier and template matching four stages.Only if a subwindow is accepted by one stage,the next stage is evaluated.Synthesizer fuses the result of tracking and detection as the final output.With a semi-supervised training of learning update classifier and target template in detection,so that constantly adapt to the change of the camera scene,as well as re-initialize recursive tracking when necessary.Finally,we propose two measures,perception hash and Kalman filter,to improve the TLD performance with respect to appearance variations and occlusions on the premise of ensuring system real-time,and then enhance the stability and robustness of the algorithm during long time online object tracking. |
PDF Full Text Request