Research On Target Tracking Algorithm Based On Kernel Correlation Filter

Target tracking is one of the hot topics in the field of computer vision research,and it is widely used in defense military and intelligent cities.The aim of the tracking is to accurately calculate the position of the target in the next frame according to the target specified in the first frame in the sequential video sequence.In the process of tracking,the target may appear the scale variation,shape variation,illumination variation or occlusion and so on.So the target tracking algorithm is to allow the target to be accurately tracked when these problems occur.This paper is a research improvement for the Kernel Correlation Filter(KCF)algorithm.Firstly,the improved algorithm was to fuse the color feathers and self-adaptive scale variation based on the KCF,and then further improved in three aspects.The first aspect joined the detection mechanism,and choosed whether to re-detection by judging the anomaly value of the detection result.The second aspect added an update strategy,choosed the high confidence update strategy during the model update stage of the algorithm.And the third aspect was fusion detection mechanism and update strategy.Finally,experimental tests show that the accuracy and success rate of improved tracking algorithms are significantly better than KCF algorithm.The main work is as follows:1.Firstly,a self-adaptive scale feature fusion of anomaly re-detection tracking algorithm(SFAR)was proposed.Through multi-feature fusion and scale variation strategy to improve the multi-feature scale kernel correlation filter,the filter estimated the image obtained multi-scale images,and extracted multi-channel features for multi-scale images.Then the filter trained the multi-scale target model,and detected the optimal scale model.Finally,the outlier detection was introduced,to judge the peak value in response map of the optimal scale model.If there was an abnormal peak,then re-detection mechanism was taken.Through the test of OTB-50 dataset,the accuracy and success rate of SFAR algorithm was better than KCF algorithm.2.Secondly,a self-adaptive scale feature fusion and model update tracking algorithm(SFMU)was proposed,the SFMU algorithm detected the optimal scale model using the multi-feature scale kernel correlation filter.Using multi-peak detection judged the overall oscillation degree of the response map to determine the peak value.Based on the evaluation of the confidence degree of the tracking results,the algorithm stopped updating model timely in the case of low confidence of the tracking results such as occlusion and deformation.In the high confidence model update,the algorithm introduced the initial model to the alignment operation to suppress model drift.Through the test of OTB-50 dataset,the accuracy and success rate of SFMU algorithm was better than KCF algorithm.3.Lastly,the analysis of SFAR and SFMU algorithm shows that the SFAR algorithm is more stable under the condition of target deformation,occlusion and out of view,and the SFMU algorithm has better performance under background clutter,fast motion and blur motion.Therefore,the improved method of the SFAR and SFMU algorithm forms the anomaly re-detection and model update tracking algorithm(ARMU).Through the test of OTB-50 dataset,the accuracy and success rate of ARMU algorithm is significantly improved compared with KCF algorithm.