Research Of Dynamic Tracking Algorithm Based On TLD

Computer vision is a science to study how to make machine learn to “look”,the camera and computer is like our eyes to identify,track and measure the target object.While object tracking is a significant research direction for the computer vision,numerous issues remain to be addressed due to appearance change caused by pose illumination,occlusion,and motion.TLD(Tracking-Learning-Detection)is a long-term tracking of unknown objects in a video stream.This algorithm decomposes the long-term tracking task into tracking,learning and detection.The tracker follows the object from frame to frame.The detector localizes all appearance that have been observed so far and corrects the tracker if necessary.The learning estimates detector's errors and updates it to avoid these errors in the future.This algorithm is a semi-supervised learning,three different models become a intact tracking system.Therefore,this tracking system is more stable,efficient and robust.However,TLD trains only the detector and the tracker stay fixed.As a result the tracker makes always the same errors.Meanwhile,TLD is a real-time long-term tracking system based on the TLD framework and the P-N learning.Looking for an efficient data processing mechanism is a main problem for TLD algorithm.Based on these problems,we proposed two solutions as follows.(1)We proposed an algorithm to improve the efficience of data processing in a real-time long-term tracking system.In this paper,we apply the sparse representation to the TLD algorithm,the samples of foreground targets and the background would be compressed by using a very sparse measurement that can extract the features by a non-adaptive random projections efficiently,which can still preserve the structure of the image feature space of objects.The extracted features online update the classifier by a cascaded classifier.Building a real-time long-term tracking system based on the sparse representation could improve the efficiency of tracking algorithm,thereby solving the problem of efficiency decline in TLD with the time going.We also apply the mathematical morphology in the pre-processing stage for denoising and extracting intact object contour.In our algorithm,the sparse representation combines with the tracking,learning and detection,which can not only guarantee the ability of estimating errors,but only improve the efficience of data processing.(2)We proposed a novel training method to train the tracker based on TLD.For the problem of online training,we proposed an algorithm that combines naive bayesclassifier with the optical flow to train tracker.we sample some positive samples near the current target location and negative samples far away from the object center to update the classifier from the current frame.Using optical flow tracks the target from the next frame and draw some samples around the current target location and determine the one with the maximal classification score.