| Target tracking occupies a very important position in computer-based image processing and video analysis.It is an important processing technology and plays an important role in many fields such as video surveillance,human-computer interaction,unmanned driving,and medical diagnosis.With the rapid development of intelligence,target tracking technology is also constantly advancing and developing,and many achievements have been put into real life.Since 2010,after the idea of correlation filtering was introduced to the target tracking field,the excellent performance of the tracker based on the related filtering idea in real-time performance has attracted the attention of domestic and foreign researchers.The classic kernelrelated filtering target tracking algorithm(KCF)shows high tracking performance in relatively simple scenarios,but the KCF algorithm still has some problems: First,the size of the target area remains the same during the KCF tracking process.It is easy to cause the target to drift or lose;secondly,when the target is occluded by other people or objects in the scene during the tracking process,KCF cannot make a judgment,and will continue to track according to the algorithm itself;finally,KCF cannot judge the occlusion,after occlusion Can’t find the target again.In order to improve the adaptability of KCF to the above problems,the main work of this paper is as follows:1.After reading the video image,KCF will preprocess the image,convert the image into a grayscale image and perform gamma correction,calculate the image gradient,and extract the HOG feature.On this basis,if it is a color image,in addition to the HOG feature,the CN feature of the image is extracted,and the two are integrated as a feature model for target detection and tracking.Make full use of the pixel information of the image to get more accurate position prediction.2.Based on the idea of KCF,after predicting the target position using the classifier idea,the target is divided into 4 sub-blocks with this position as the center,and the center position of the 4 sub-blocks is still predicted by the classifier idea,and the predicted position of the subblock is used The relative change estimates the scale change ratio of the current target,and then uses the obtained scale to reposition the target and model it for subsequent video detection and tracking.3.While estimating the scale,use the difference in similarity values between images to judge whether the target is occluded.When the target is occluded,the similarity value will be greatly increased;if there is no occlusion,the peak of the similar value is basically Around the center of the target.If it is determined that the target is occluded,then stop updating the tracking template,extract the ORB feature value from the template,extract the next frame of image for ORB feature points matching,achieving the goal again positioning.4.Select the OTB data set for simulation experiments.The results show that the algorithm in this paper has greatly improved in handling occlusion.Compared with the KCF algorithm,the accuracy and coverage have been improved by 7% and 13% respectively,and it can still meet real-time performance.Finally,we selected popular algorithms in multiple fields for comparison,and achieved good results in terms of accuracy and success rate.And the algorithm in this paper has good adaptability when dealing with the target size change and target occlusion. |
