Research On Kernelized Correlation Filter Tracking Algorithm In Complex Scene

In recent decades,target tracking technology has developed rapidly,and it has been widely used in both military and civilian applications,such as smart city monitoring,human-computer interaction and precise guidance.The kernelized correlation filtering algorithm has become more and more important in the target tracking algorithm due to its fast speed and high accuracy.However,because of the complex and diverse tracking environment,there are multiple factors exist,such as lighting changes,occlusions,background and scale changes,which affect the accuracy of the target tracking algorithm,and even lose the tracking target in severe cases.In view of the above problems,this paper improves the kernel correlation filtering algorithm,making it 28.26% higher in tracking accuracy than the original algorithm.The specific content is as follows:Firstly,for the changes in the external environment and the problem of fast moving targets,the LLE(Locally Linear Embedding)dimensionality reduction color feature model is added to the directional histogram feature model of the kernel-related filtering algorithm,and the filter response in the frequency domain is allocated according to the two models Respective weights.In addition,in order to prevent the overfitting phenomenon between the models,the penalty term is added to the weights.Through quantitative and qualitative experimental analysis,it is shown that the improved algorithm has a good effect in the processing of target jitter,illumination and rapid target translation.Then,for the problem of target scale change in the tracking process,it is proposed to transform the logarithmic polar coordinate information of the target coordinate information,so that the non-linear change of the target scale can be transformed into a linear translation of the logarithmic polar axis,and the introduction of subpixels to improve Image quality after log polar transformation.At the same time,the phase correlation filter is trained by the logarithmic polar coordinate information between adjacent frames,and the scale and rotation angle of the target are calculated according to the maximum response of the filter.For the occlusion problem,a judgment mechanism is introduced,which combines the peak oscillation and peak sidelobe ratio of the targetimage corresponding to each frame,and decides whether to update the target model based on this mechanism.Experiments show the effectiveness of the improved algorithm.Finally,in order to further improve the accuracy of the kernelized correlation filtering algorithm,the position and scale of the target tracking algorithm were improved based on deep learning.The two convolutional layer depth features in the trained VGG-Net-19 network model are extracted,the two features are weighted together,and the rich information of the first frame target is introduced to obtain the best filtered position response.At the same time,a scale pool is established by using scale factors to solve the scale problem of kernelized correlation filtering algorithms under deep learning.The experiments show that the improved algorithm has a good effect on improving the tracking accuracy.