Research On Correlation Filter Tracking Algorithm Based On Multi-channel FHOG Feature

Visual target tracking is a very important branch in the field of machine vision,which has wide applications in intelligent manufacturing,security monitoring,unmanned reconnaissance,unmanned driving,human-computer interaction and intelligent diagnosis.After the video target initialization information is given,the visual target tracking aims at the location and size estimation of the subsequent video target.At present,many researchers have studied it and achieved good tracking accuracy in some scenes.However,in specific applications,video targets are often in very complex scenes,which may involve one or more interference factors such as occlusion,high-speed movement,target jitter,illumination change,target deformation and background pollution,etc.,making efficient,accurate and robust tracking still a great challenge.In recent years,Correlation Filter based Tracer(CFT)has achieved remarkable results.In this paper,an improved approach was proposed and verified by experiments aiming at the shortcoming of weak feature representation capability used by Kernelized Correlation Filter(KCF).The main work of this paper is as follows:(1)Analysis and theoretical derivation of the kernel correlation filter tracking algorithm.The overall flow of the algorithm is introduced,with emphasis on the principles of circulatory matrix,ridge regression and kernel space dichotomy in the one-dimensional case,and the processing of kernel correlation and multi-channel features is derived in the two-dimensional case.(2)MMG?KCF,an optimal weighted multi-feature fusion tracking algorithm based on information entropy,was proposed.First,the feature of the Histogram of Orientated Gradient(FHOG)of the R,G and B three-way was extracted,and then the feature was fused at the feature level by means of concatenated Mosaic.Considering the different functions of the three feature channels,they were assigned to the three feature channels in accordance with the weight ratio for fusion.In this paper,the image information entropy is used to find the optimal weight proportion,and the image information entropy corresponding to the weight proportion of ownership is calculated by constructing the weight proportion pool,in which the maximum information entropy corresponds to the optimal weight proportion.The optimal weight ratio is used to fuse the three-channel FHOG features of R,G and B to obtain the MMG?HOG features,and the grayscale FHOG features in the KCF algorithm are replaced to obtain the MMG?KCF algorithm.Experiments show that the MMG?KCF algorithm using MMG?HOG feature has certain improvement in accuracy and success rate.28 color videos were randomly selected from the standard data set OTB100 for testing.According to the experimental results,MMG?KCF algorithm improved 5.5% in accuracy and 6.2% in success rate compared with KCF algorithm.(3)MCC?KCF multi-feature fusion tracking algorithm based on Hue Saturation Intensity Color Space(HSI)and Lab(CIELab Color Space,Lab)was proposed.First,the RGB images were transformed into the HSI and Lab color space,and then the FHOG features of H,S,a and b channels and the mean FHOG features of HSI and Lab three channels were extracted,Gray1 and Gray2.Finally,the six-channel features were fused to obtain the MCC?HOG features,and the original features in the KCF algorithm were replaced to obtain the MCC?KCF algorithm.Experiments show that MCC?KCF algorithm has higher accuracy and higher success rate than KCF algorithm,with a large improvement.In the experiment part,this paper selects all color videos(75)of the standard data set OTB100 to pass the test at one time.From the perspective of quantitative analysis,MCC?KCF algorithm improves 6.8% in accuracy and 6.4% in success rate compared with KCF algorithm.From the qualitative analysis,MCC?KCF algorithm has certain advantages over KCF algorithm in the same video in terms of anti-occlusion,anti-jitter and anti-rotation.It can be seen that the MCC?KCF algorithm proposed in this paper has stronger target representation ability and better tracking effect compared with the KCF algorithm.