Reaserch On Object Tracking Based On TLD Algorithm

Computer vision is composed of four sub-topics: object image genaration,object detection,object tracking and object recognition.Among them,the research progress of object tracking is lagging behind to other sub-topics,which needs to deal with the challenges of complex tracking scenes,such as appearance changes,illumination changes,scale changes,motion blur,occlusion and so on.Tracking-Learning-Detection is an online object tracking algorithm based on detection which has good robustness and reliability.Although TLD achieves long-term tracking of any unknown object,it still has some shortcomings and deficiencies.For the problem that TLD has bad performance in dealing with severe occlusion and median-flow tracker drift easily due to rough generation strategy of feature points,this paper proposes an enhanced TLD object tracking algorithm under linear motion.Firstly,when the object is seriously occluded,the enhanced algorithm can accurately detect the situation,then TLD provide the object's state observations for Kalman filter,Kalman filter recursively to output the object state;Secondly,Oriented FAST and Rotated BRIEF algorithm is used to generate the feature points with scale invariance and noise immunity to enhance the median flow tracker.The experiment results show that the enhanced algorithm decreases the origin TLD's center pixel error by about 50%.For the problem that TLD is poor in real-time as a result of time-consuming detection module,this paper proposes a fast TLD object tracking algorithm based on VisualBackground-Extractor.Firstly,the algorithm uses Vi Be to obtain the approximate region of all the moving objects in the current frame,and then the motion object corresponding to the obvious difference in the size of the object bounding box is removed,and the redundancy of the image patch is reduced by generating an efficient image patch for detection module;Secondly,in order to enable patch variance classifier to filter out the image patches effectively which do not contain the foreground object,the algorithm realizes the threshold of patch varience classifier form constant to dynamic by means of the online object model.Finally,the algorithm reduces the computational burden of the nearest neighbor classifier by improving the update strategy of object model.The experiment reults show that the algorithm increases the origin TLD's processing rate by about 30%.