Improved TLD Target Tracking Algorithm Based On LBP

Computer vision is the basis of machine cognition and the basic component of artificial intelligence technology.As an important research direction of computer vision technology,target tracking has a wide range of applications,including intelligent driving,intelligent transportation,intelligent security,medical imaging and so on.In the process of tracking,we need to deal with the influence of light intensity change,occlusion,scale change and other factors,so the performance requirements of the tracker are high.How to improve the performance of the tracker and improve the tracking algorithm is still of great significance at this stage.This paper aims to study the tracking learning detecting(TLD)algorithm,which combines four parts: target detection,target tracking,learning module and synthesis module.In view of the defects of TLD algorithm in some scenes,the detection and tracking modules are studied,and some improvement measures are put forward,which are verified by experiments and analysis.The main work of this paper is as follows:Firstly,the improved TLD algorithm based on LBP(local binary mode)is proposed for the influence of illumination change,target deformation and other factors.LBP algorithm has good performance in texture.In the second classifier of the detection module,instead of comparing pixel values to get feature descriptors,LBP algorithm uses LBP rotation invariant mode to calculate the image corresponding to LBP operator.The detection window is divided into several regions,the histogram of each region is calculated,and the feature descriptor is obtained by comparing the frequency of LBP value in the histogram.The improved classifier can filter out more target independent detection areas and reduce the task of the next classifier.At the same time,when the target has good texture attributes,the improved classifier has better classification effect.Secondly,in view of the large number of scanning windows in TLD algorithm,the target is easy to be occluded or swayed,we propose to use LSTM(short and long term memory network)in the algorithm to predict the area where the target is located.The image features and location information of the video sequence are used as the input of the neural network.When output,the network is trained according to whether it can match the learning classifier or not.The trained model is used to predict the target position information in the next moment.Aiming at the occlusion problem caused by the cross movement of the target,Markov predictor is introduced to predict the moving direction of the target.Compared the predicted rectangular region with the original TLD scanning window,the intersection part is marked as positive samples,and the rest areas are classified as negative samples.Combined with the analysis of learning module and synthesis module,the final target location information is obtained.The improved algorithm greatly reduces the detection range of the detector,reduces the number of tracking windows,and enhances the efficiency of the algorithm.Finally,some data sets are compared.The experimental results show that the improved tracking algorithm has better robustness,higher accuracy of correct frame and faster tracking speed than the original TLD.