Research On Robustness Of TLD Target Tracking Algorithm Based On Enhanced Target Learning Accuracy

There are always shortcomings and limitations in target tracking using any mechanism alone.Once the algorithm itself cannot handle well,the robustness of the tracking algorithm is difficult to be guaranteed.How to track the complex motion of the target stably and robustly for a long time has become a research difficulty.In this dissertation,TLD algorithm is used as a framework to improve the accuracy of target tracking results and learning samples.(1)Firstly,multi-tracking fusion is proposed to determine the learning target.On the spatial dimension of the same frame,the Median-flow method is combined with KCF algorithm.The overlapping similarity is used to judge the tracking effectiveness,and the bounding box is adjusted accordingly to ensure the accuracy of the learning samples.At the same time,the KCF model is used to calculate the confidence of image region and realize multi-objective judgment(2)Secondly,feature matching is proposed to calibrate bounding boxes,which optimizes the adaptability of TLD algorithm to target changes.On the time dimension of front and back frames,the size change ratio and rotation change angle of SIFT matching points in the target range of two frames are counted,and the average value is obtained by weighting the distance,the target bounding box is calibrated to ensure the accuracy of tracking results(3)Finally,combining a cascade matching filtering method based on three-channel similarity,this dissertation utilizes the three-channel information of color image ignored in the process of SIFT algorithm,weights the credibility of the original matching results,sets threshold,and realizes preliminary eligible screening,and then cascades with the RANSAC method to screen better results.This method is applied to improve the reliability of the bounding box calibration.Experiments show that all kinds of improvements proposed in this dissertation have achieved good results.By integrating these methods into TLD target tracking algorithm,the accuracy and the robustness of TLD target tracking can be effectively improved.