Research On Target Tracking Algorithm Based On Kernel Correlation Filtering

Object tracking is a research hotspot in the field of computer vision,which has been widely applied in many fields including intelligent transportation,human interaction,intelligent monitoring,automatic driving,military and so on.After many years of research on object tracking,a large number of classical and excellent algorithms have been proposed,which made great progress in this field.However,on account of the interference due to deformation,scale variation and occlusion,the current tracking algorithms are hard to meet the requirements of accuracy,robustness and real-time in practical applications.Therefore,it is a challenging research topic to achieve robust and efficient tracking in the presence of interference.In recent years,correlation filtering has become a research hotspot due to its robust and efficient calculation in frequency domain.through the correlation operation between the learned filter model and the object block in the frequency domain,the object position can quickly detected by the correlation filter.Moreover,in order to adapt to the change of the object,the linear interpolation is used to update the filter model.However,there are still some problems in this algorithm,for instance the insufficient feature utilization,the detecting precision decreased while object scale changed and occlusion occured.Therefore,aming at the shortcoming of HOG feature,multiple features is fused in this thesis;and then for the purpose of reducing model drift,model updating rate adaptive adjustment module is introduced;after that,the change of the object scale is coped by using scale adaptive adjustment module.On the basis of this,TLD algorithm and Kalman filter are combined to carry out object detection,then the tracking state discrimination mechanism is added to detect the presence of interference.in this way,stable single object long-time tracking can be achieved,and the challenging difficult problems due to object occlusion,scale variation and deformation can be solved to a certain extent.The main contents and innovations of this dissertation are summarized as follows:(1)In the kernelized correlation filter(KCF)object tracking algorithm,only using HOG feature can result in insufficient feature expression.In addition,model drift is easily to be caused by the linear interpolation model updating strategy employed in this algorithm.for the purpose of solving these problems,an improved kernelized correlation filter tracking algorithm based on adaptive feature fusion and model updating is proposed in this thesis.Firstly,PCA is applied to reduce the dimension of HOG feature and CN feature,which improved the speed of the algorithm.Secondly,the response maps of the two dimensionality reduction features are calculated independently.After that,the peak value and the average peak-to-correlation energy(APCE)of two response maps are used to obtain their weights.Thus,the fused response map is obtained by using weights.Finally,the model updating rate is determined by the similarity of CN feature between two frames,so the accuracy of model updating is enhanced.Through the comprehensive application of feature adaptive fusion and update rate adaptive adjustment,the tracking accuracy and processing speed are increased.(2)An improved TLD object tracking algorithm based on KCF is proposed.Firstly,KCF tracker is applied to replace the optical flow tracker in TLD algorithm,and scale discrimination module is added.Secondly,for the sake of reducing a large number of meaningless windows in the detector,Kalman filter is used to estimate the object position,after that,the cascade classifier is employed to locate the object more accurately around the estimated position.Finally,tracking state judgment mechanism is adopted to determine whether to update the filter model by setting the threshold value of maximum output response,the threshold value of APCE and the threshold value of the random fern classifier.Besides,the improved cascade classifiers including variance classifier,random fern classifier and KCF classifier are employed.Based on the combination of tracker and detector,the tracking performance of the algorithm is improved under the conditions of occlusion,deformation and illumination variation.Qualitative and quantitative experiments on OTB-50 dataset are carried out,it is shown that the improved algorithm has better robustness and accuracy in the case of object deformation,scale variation and occlusion.