Research On Visual Object Real-time Tracking Algorithm Based On Correlation Filter

Visual object tracking technology has always been one of the research hotspots in the field of computer vision.Correlation filter based object tracking algorithms have attracted great attention in the field of visual object tracking due to its outstanding performance,real-time and ease of implementation.Nevertheless,it still cannot achieve good tracking effect in some specific scenarios.This thesis makes an in-depth study on the principle of object tracking based on correlation filter and summarizes its main problems.One is that it has poor tracking performance when the object is undergoing rapid deformation or rotation.Another is that there are boundary effect and model degradation in its process.In view of the above two problems,this thesis makes corresponding improvements to correlation filter based object tracking framework on the premise of real-time.The main work of this thesis is as follows:Firstly,an object tracking algorithm based on correlation filter and color statistical feature is proposed.Correlation filter framework restricts that the features used to represent the information of object must be template-like features,including hand-crafted template features(HOG,CN)and CNN features.Hand-crafted template-like features are sensitive to deformation and rotation.CNN features are robust to deformation and rotation because they contain rich semantic information,but the extraction process is time-consuming and their resolution is much smaller than that of the input image.Therefore,correlation filter based object tracking algorithm using CNN features has poor positioning accuracy and cannot meet the requirements of real-time tracking.Color statistical features describe the proportion of different colors in the whole image without paying attention to the spatial position of each color,which makes them sensitive to object deformation and rotation,and they can be extracted quickly.Based on these findings,this thesis uses color statistical features to make up for the defect that the object tracking algorithm based on correlation filter cannot cope with the rapid deformation and rotation of the object.The algorithm first builds a Bayesian classification tracking model based on color statistical features with color statistical features and Bayesian classifier,and then uses it to cooperates with correlation filter based tracking model to complete the object tracking task.In the stage of model fusion,in order to give full play to the complementarity of the two models,an adaptive response graph fusion method is proposed to fuse the output tracking response graphs of the two models in the stages of position estimation and scale estimation respectively,so as to obtain more accurate and robust tracking results.Finally,the effectiveness of the proposed algorithm is verified by the international visual object tracking algorithm test platform(OTB).The experimental results show that the proposed algorithm has better performance as the object is deforming or rotating rapidly compared with the mainstream object tracking algorithm based on correlation filter,while maintaining a high tracking speed(62.8fps).Secondly,an object tracking algorithm based on correlation filter and spatial-temporal regularization is proposed.Correlation filter based object tracking algorithm uses cyclic shift operation for dense sampling,which indirectly improves the discriminative ability of the tracking model and also introduces the problem of boundary effect.In addition,it adopts the method of updating the model every frame,which easily leads to model degradation.The BACF algorithm uses a predefined binary mask matrix to perform a dot product with the cyclically shifted training samples,so that the model only focuses on the center area of the training samples,thereby solving the boundary effect problem by expanding the training sample area and suppressing the background information,while preserving the real-time nature of correlation filter based object tracking algorithm.Based on the BACF algorithm,this thesis adds temporal consistency constraints to the model to limit the mutation of the model between adjacent frames,so as to solve the problem of boundary effect and model degradation simultaneously.At the same time,a method for evaluating the quality of tracking results is proposed.The model is selectively updated based on the tracking quality evaluation results of the current frame to avoid accumulation of tracking errors,thereby further suppressing model degradation during the detection phase and indirectly enhancing the tracking speed.In the model solving stage,ADMM(Alternating Direction Method of Multipliers)is used to solve the model iteratively,and the solving process is transformed to the frequency domain to accelerate the calculation by FFT.Finally,the proposed algorithm is evaluated on the international visual object tracking algorithm performance test platform(OTB、VOT)and compared with other mainstream object tracking algorithms based on correlation filter.The experimental results show that the proposed algorithm can solve the problem of boundary effect and model degradation in correlation filter based object tracking algorithm on the premise of ensuring real-time tracking(32.6fps).It is superior to the Baseline algorithm BACF in tracking accuracy,robustness and speed.