Research On Target Tracking Algorithm Combining Background Subtraction And Correlation Filtering

Target tracking is an important research direction of computer vision,which is widely used in intelligent transportation,automatic driving and other fields.The target tracking algorithm based on correlation filtering not only considers the robustness but also has a very high running speed.Therefore,it has attracted the attention of researchers and made remarkable progress in recent years.However,the performance of correlation filtering tracking algorithms in complex environments such as target scale changes and similar background still needs to be improved.Based on the in-depth study of the basic framework of the correlation filtering target tracking algorithm,this thesis has completed the following tasks in response to the challenges of occlusion,scale change and similar background in the tracking process:(1)Aiming at the problem of inaccurate of filter-template learning in the scale change of Spatio-Temporal context target tracking algorithm,the Spatio-Temporal context target tracking algorithm based on the scale filter is studied.First,33 target rectangles with different scales are established through the scale coefficient matrix,and 31-dimensional FHOG feature vectors are extracted respectively,and then the maximum response position is obtained by convolution with the filter-template to determine the position and scale information of the tracking target,and finally The experimental verification improved method solves the problem that the filter template cannot effectively learn the image of the local context area when the target changes in scale.(2)To improve the problem of incomplete edge extraction in the Gaussian mixture model algorithm,the improved method of combining edge detection and Gaussian mixture model is studied.Firstly,the adaptive Gaussian background modeling and edge detection operations of the adaptive learning rate are performed on the input image respectively,and then the morphological operation is performed on the combined result of the two to extract the complete foreground information.Finally,the algorithm is verified by experiments to effectively extract the tracking The targetis used for the next tracking task.(3)In order to improve the tracking drift problem caused by complex backgrounds such as similar background and scale changes in the kernel-correlated filtering target tracking algorithm,an improved method combining background subtraction and correlation filtering was studied.First,perform background subtraction operations on the video sequence,extract the tracking target,and establish kernel correlation filters based on gray and color features in the first frame,and simultaneously establish an 11-bit scale pool,and then in the subsequent frame image set,if the image If the number of frames is a multiple of 5,the core target filter based on color features is used to obtain the maximum response to obtain the center target position and scale information of the current frame image,otherwise the core frame filter based on gray features is used to obtain the maximum response to obtain the current frame image The central target position and scale information of the;lastly use the current frame image to update the kernel correlation filter based on grayscale and color features.Experiments verify that in complex situations such as occlusion,scale change,and similar background interference,the improved algorithm improves the accuracy of tracking.