Research On Fast Multi-scale Target Tracking Algorithm

With the development of image processing and machine learning technology,target tracking technology is attracting more and more researchers' attention.In recent years,the target tracking algorithm also has obtained the remarkable result,but as the target tracking is applied in more fields,the environment is also facing complicated,such as target information variability,complex background and partial occlusions,and so on.Therefore,the target tracking algorithm is still a very challenging task at present.In this paper,the theory based on kernel correlation filtering is studied in deeply.Some effective improvement methods are proposed to solve the difficult problems in the existing target tracking algorithm.The main work is summarized as follows:1.A target detection and tracking method based on improved HOG-Color features is proposed.Aiming at the problem of poor real-time performance of the target tracking algorithm due to the large computational complexity of the fusion HOG-Color feature in the construction target appearance model,the method of feature selection is used to reduce the dimension of the feature dimension.First,the HOG features and color features are extracted from the target samples respectively.Secondly,the Bhattacharyya distance is used as the standard of feature selection,and the Bhattacharyya distance of each feature is calculated,and the suitable feature is selected as the HOG feature.Then the selected HOG features are fused with the color features.Finally,the filter is learned by kernel correlation filtering,and the image is detected by the filter,and the response output is obtained.Experiments show that under various complicated scenarios,this method can ensure more accurate tracking accuracy,and the tracking speed has obvious advantages.Therefore,this method can reduce the feature dimension fundamentally,and reduce the time of feature extraction,so that the efficiency of target detection and tracking can be further improved.2.A fast Multi-scale target tracking method based on correlation filtering is proposed.To solve the problem of inaccurate target estimation and slow tracking speed in the existing target tracking algorithm,the target scale is quickly and accurately estimated by using the method of quickly constructing the target-scale feature pyramid.The traditional feature of Pyramid is to build the image Pyramid,and then calculate the target image features layer by layer.This paper first extracts the features of the single image,then use the resampling function to construct target image features with different scales,and greatly reduces the time overhead of the computing feature.Experiments show that the proposed method can not only effectively cope with scale changes in complex environments,but also meet the tracking requirements in real-time scenes.