Research On Visual Object Tracking Algorithms Based On Correlation Filters

In computer vision,object tracking is one of the fundamental problems.Object tracking is widely used in human-computer vision,intelligent cities,military and so on.Excellent object tracking algorithms have emerged.However,with the increasing applications of object tracking technology,the research of object tracking is facing more and more challenges,such as occlusion,fast motion,background clutter and so on.The kernel filtering object tracking algorithm has been widely concerned with its high rate and high precision.It constructs a large number of candidate samples by cyclic shift,and introduces multi-channel features to achieve the robustness of the object and high performance.However,during the movement,when the object encounters difficult scenes,such as blur,occlusion,deformation and so on,there will occur object drift and weak adaptability to the appearance of the object.considering the shortcomings of the kernel correlation filter tracking algorithm in the object tracking process,this paper proposes an improved scheme and designs the experiment to verify performance.The main work of this paper is summarized as follows:(1)The execution flow of the generative and discriminative object tracking algorithm and the characteristics of the classical algorithms are analyzed.The basic principle of the object tracking algorithm based on correlation filtering is discussed.The core and execution flow of the algorithm are analyzed and elaborated.(2)In order to improve the performance of the tracking model,a multi-strategy correlation filter tracking algorithm is proposed.By combining the HOG,LBP and CN features of the object in the feature representation,the feature pool of the object is obtained,and a more robust object feature representation is formed.Then the corresponding position of the object is calculated by using the detection model.According to the robustness evaluation of the detection model,the best predicted position of the object is selected.In the phase of model update,the adaptive update strategy is selected by calculating the parameters that apply to the current frame.Finally,the scale filter is used to obtain the best scale of the target to complete the object tracking task.(3)Aiming at the problem that the traditional artificial features have low tracking accuracy in the scenes of rotation and fast motion,a correlation filter tracking algorithm combining deep feature and HOG feature is proposed.The deep learning is introduced into the correlation filtering framework.By combining the deep feature with the traditional HOG feature,the high-level,middle-level and low-level expressions of the object are formed,and then the response of the detection model to the high,medium and low features is linearly combined to obtain a more comprehensive object prediction.As a result,the maximum position in the integrated response diagram is taken as the object center position,and the model is updated with a high confidence update strategy to further improve the robustness of the algorithm.Finally,the scale filter is used to solve the optimal scale of the target.As a result,tracking accuracy has been improved.