Research On Object Tracking Algorithm Based On Kernelized Correlation Filters In Complex Scenes

With the rapid improvement of processing power,the continuous development of parallel computing architecture and the increasing demand for data processing in the era of big data,object tracking technology has achieved considerable development.Relying on excellent real-time performance and good robustness,the correlation filter tracking algorithm has set off an upsurge in the field of object tracking,and optimization and improvement algorithms based on it continue to emerge.However,there are various interference factors under complex environment,which seriously affect the improvement of algorithm performance.To this end,this thesis focuses on these key issues to carry out in-depth research on the Kernelized Correlation Filters(KCF)tracking algorithm.The main research contents and results obtained are summarized as follows:(1)Design of multi feature fusion method.KCF algorithm only uses the improved histogram of oriented gradient(HOG)to build the apparent model of the target,which has limitations in the complex tracking environment.A multi feature fusion method is designed.This method combines the strong complementarity between FHOG features and color names(CN)features,and makes full use of the good invariance of FHOG feature to target distortion and illumination change,and the good performance of CN feature in motion blurred,low resolution scene.This method combines the two complementary features linearly,which can significantly enhance the discriminability of the apparent model.It can effectively deal with the illumination change and target distortion in tracking.(2)Design and implementation of a fast motion scene tracking algorithm based on key points matching.To solve the problem that KCF algorithm does not work well in fast moving scenes,this thesis designs a fast moving scene tracking algorithm based on key points matching.This algorithm uses key points matching and forward-backward optical flow tracking to obtain stable key points in current frame,and then uses the key points after hierarchical clustering to estimate the center of the target,so as to achieve coarse target positioning.Then uses kernel correlation filter algorithm based on multi feature fusion to achieve accurate target positioning and further improve the accuracy of the algorithm.Secondly,using the matching relationship between the key points of the current frame and the key points of the target template,and assigning weights according to the response value of the key points in the response map,effectively estimating the target scale and adaptively adjusting the size of the tracking frame.Then,a template update strategy combining key points matching rate and maximum output response is proposed to adaptively update the target template key point set.Finally,the performance of the algorithm is evaluated through comparative experiments and based on the results of qualitative and quantitative analysis.The experimental results show that the algorithm has good tracking effect in the scene of illumination change,fast motion and scale change.(3)Design and implementation of kernel correlation filter tracking algorithm in target occlusion scene.In view of the lack of occlusion judgment mechanism of KCF algorithm,it is easy to cause template pollution when the target is occluded,which leads to the problem of tracking drift and even tracking failure,and puts forward a long-term kernelized correlation filters tracking algorithm based on the occlusion judgment mechanism.This mechanism uses the relationship between the average peak correlation energy and the peak sidelobe ratio of the current frame and the average value of the historical frame to effectively evaluate the target tracking status,such as no occlusion or slight occlusion,severe occlusion or complete occlusion.The corresponding processing strategy is adopted for the target's state.When the target is unoccluded or slightly occluded,the learning rate is adaptively adjusted to avoid the introduction of occlusion information into the template.When the target is severely occluded,multi-peak detection is used to achieve target re-detection.Finally,through the comparative experiments,it is proved that the algorithm has good anti occlusion performance,and the tracking effect is better in the scene of target occlusion,fast motion and scale change.Compared with the comparative algorithm,the algorithm has better robustness.