Research On Target Tracking Based On Kernel Correlation Filter

Target tracking technology has a wide range of applications in the fields of intelligent transportation,safety monitoring,human-computer interaction and motion analysis.In recent years,the target tracking technology has made rapid development,emerged many excellent target tracking algorithm,solve many difficult target tracking problem.However,the target tracking technology still faces many challenges.Due to the complexity of the real environment,the current tracking algorithms can not meet the practical application in real-time and accuracy.In this paper,the KCF target tracking algorithm is optimized,KCF tracking algorithm based on the cycle matrix for intensive sampling dates of the image feature extracting,use the regularized least squares classifier for training to improve the speed running.The advantage of KCF tracking algorithm is that the target tracking speed is fast.The average speed of OTB50 in the Benmark video sequence is 172 fps,with an average accuracy of 73.2%.It is not enough to keep track of the target when the target is occluded and the tracking is lost.The performance of target tracking is reduced.In this paper,the advantages of both CSK and TLD algorithm are analyzed and compared.KCF algorithm in CSK algorithm on the basis of improvement,inherited the CSK quickly realize the advantage of target tracking,and TLD algorithm by tracking,testing and learning modules,lost in target tracking,retargeting by detecting module.For KCF's lack of reference to the idea of TLD,the detection module is added,and the HOG feature is replaced with the FHOG feature,which improves the speed of the tracker.Graph Cut segmentation algorithm will contain the target image subgraph segmented from the original video image,as the subgraph of template matching,template matching tests to target new position,to update the tracker.The algorithm and other algorithms are performed on the OTB video series,respectively,on the video sequence sets with different characteristics,the experimental results show that this algorithm in the target under the condition of shade,can continue to track target in real time.