Research And Implementation Of Multi Feature Fusion Objecttracking Algorithm Based On Correlation Filtering

In recent years,the research of artificial intelligence has increased.As one of the basic research tasks in the field of artificial intelligence,visual object tracking is widely used in business scenarios such as,intelligent uav monitoring and national defense aviation safety monitoring.The national emergency department puts forward the demand for UAV tracking algorithm in the UAV rescue mission under the emergency disaster relief scenario in 2021.Visual object tracking still has the following research difficulties:on the one hand,due to object being blocked,complex video background and the interference of similar objects in the background,the object tracking algorithm based on deep learning can not learn the feature correlation between the non occluded part and the occluded part of the object,which affects the judging of the overall motion direction and size of the object;on the other hand,when the object is blocked and tracking is difficult,the weight of object tracking algorithm based on correlation filter focuses on learning the characteristics of occlusion.The correlation filter cannot be automatically adjusted and optimized,which affects the tracking effect of subsequent frames of video.To solve the above problems,this thesis learns the feature changes of the object that is not occluded,predicts the correlation between the unobstructed and occluded parts and judges the overall motion state of the object.This thesis studies the template updating method of adaptive correlation filter.When the correlation filter tracking is offset,the motion correlation information of the target between the front and back frames of the video is constructed,and the tracking model is adjusted and optimized.The main contributions of this thesis are as follows:(1)Multi feature fusion target tracking algorithmIn order to solve the problem of how to learn the object features better in the absence and interference of target features under the object being occluded and interference in the background,this thesis studies the object tracking algorithm based on multi feature fusion.Through the local feature change of the object to learn the correlation between the occluded part and the unoccluded part of the object,and using the strong expression ability of the deep learning feature layer,this thesis integrates the gradient feature with the deep feature and the shallow feature respectively,so that the feature layer has strong expression ability and the model has strong generalization.In order to make the fusion features express the target features layer by layer and have stronger expression ability,this thesis gives corresponding weights to each feature layer according to the clustering results of feature points in the fusion feature layer.The experiments show that the tracking accuracy and success rate are improved by0.22%and 0.37%respectively in the OTB100 dataset(object tracking benchmark)proposed by the University of California in 2015.(2)Template updating method of adaptive correlation filterIn order to solve the problem of how to adjust the correlation filter template when the tracking performance of correlation filter is poor,this thesis studies the updating method of adaptive correlation filter.Through constructing the motion correlation model between adjacent frames,the motion state of the object in the current frame is predicted and the correlation filter of the current frame is adjusted and updated by using the correlation of the motion characteristics of the object in the previous and subsequent frames.The experiments show that the tracking accuracy and success rate of this algorithm are improved by 0.56%and 0.63%respectively,which verify the effectiveness of the algorithm.(3)Feature selection object tracking algorithm based on inter frame correlationIn order to realize the object tracking of the initial frame target in the UAV tracking scene simply and efficiently,this thesis reseaches on the initial frame target by the feature selection of inter frame correlation and evalutes the above two tracking algorithms(1)and(2)by experiments.Due to the optical flow changing and the object characteristics between video initial frames,the initial frame position of the object is frame selected simply.In the self-made cloudy,sunny,urban and suburban UAV data sets,tracking accuracy and tracking success rate shall not be less than 78%and 69%.The experiments evalute that the(1)and(2)tracking algorithms proposed in this thesis have a good application prospect.