Research On Robust Object Tracking Algorithm Based On Kernelized Correlation Filters

Visual object tracking technology is one of the research hotspots in the field of computer vision,which has achieved rapid development in recent years.Today,many excellent algorithms have emerged continuously,which are widely used in many fields such as video surveillance,human-computer interaction,and military.However,the actual tracking scene is often affected by various challenging factors such as motion blur,object deformation,illumination variation,and object occlusion,which greatly affects the tracking robustness of existing algorithms.In this paper,Kernelized Correlation Filters(KCF)tracking algorithm is taken as the research object.Focusing on how to improve the tracking robustness of KCF algorithm,this paper starts with two aspects of model robustness and anti-occlusion performance,and proposes a robust object tracking algorithm based on KCF.The main work of this paper is summarized as follows:(1)Starting from enhancing the robustness of the algorithm model,a robust kernel correlation filter tracker is designed.This tracker comprehensively utilizes the advantages of the improved Histogram of Oriented Gradient(HOG)in solving the problem of illumination variation,the advantages of normalized color histogram in solving the problem of object deformation,and the advantages of normalized histogram of Local Binary Pattern(LBP)in solving the problem of motion blur.After determining the extraction regions and extraction methods for these three features,the weights of each feature in the final correlation response map are determined by calculating Enhanced Peak to Side-lobe Ratio(EPSR)of the correlation response maps corresponding to different features,so as to achieve adaptive fusion of multi-feature at the response maps level.At the same time,in order to further enhance the robustness of the model,the spatial weight is introduced into the filter training to constrain the model,which strengthens the object information and suppresses the background information.The comparative experiments show that,compared with the original KCF algorithm,this tracker is able to track objects more robustly.(2)Starting from improving the anti-occlusion performance of the algorithm,a novel joint judgment and processing mechanism of object state is designed,and a matching model adaptive update method is proposed.This mechanism includes two parts:the object state determination method based on the peak of the correlation response map and the occlusion judgment method based on background-aware.On the basis of the above tracker,according to the difference between the peak value of the response map and the local maximum,the former is used to determine whether the current tracking state is appropriate.According to the difference between the object model and the surrounding background model,the latter is used to calculate the background-aware occlusion judgment indexes.Then,the object state can be jointly judged by combining the two results.At the same time,different treatment measures,continuing tracking or starting corresponding re-detection algorithms,are taken for the three states of no occlusion and good tracking state,local occlusion,severe occlusion,or even total occlusion.In addition,in order to avoid model pollution to the greatest extent,an adaptive adjustment parameter of the learning rate is designed,and a model adaptive update method that matches the mechanism is obtained,which can improve the algorithm's anti-occlusion performance.The comparative experiments show that,compared with the original KCF algorithm,the proposed algorithm performs better in complex scenes such as object occlusion,motion blur,and object deformation,and the tracking robustness is significantly enhanced.