Research On Correlation Filter Tracker Based On Multi-peak Detection

Visual object tracking technology can obtain the motion parameters of the object in the subsequent frames according to the object information given in the initial frame.Visual object tracking technology has a wide application prospect in many fields,such as precision guidance,intelligent navigation and so on.However,there are various complex situations in the actual tracking process,so that the visual object tracking algorithm has to face more challenges.Aiming at the multi-peak problem of response map in the challenges of object occlusion,background clutter and fast motion,this thesis studies the correlation filter tracking algorithm based on multi-peak detection technology.The main tasks of the thesis are as follows:First,in the object occlusion and object deformation challenge scene,the maximum peak of correlation response map is not always the real position of the object.This thesis presents a tracking method based on the combination of peak change ratio and state discrimination.This method first discriminates the state of the current frame response map.When it is judged that the current response map is in the state of multi-peaks,the algorithm no longer blindly selects the highest peak in the response map as the response peak corresponding to the tracking object,but uses the peak selection module to select the response peak corresponding to the tracking object from multiple peaks,so as to reduce the impact of multiple peaks on the tracking algorithm.Comparative experiments based on public benchmark video sets show that this method effectively promotes the tracking performance of the tracking algorithm in complex scenes such as occlusion and deformation.Second,the algorithm aims at the problem of inaccurate object location in the background clutter challenge scene.Based on the previous improved method,a tracking method based on improved peak selection and condition update is proposed in this thesis.In this method,the time regular term and adaptive regular term penalty coefficient are introduced into the establishment of tracking template,and the peak selection module is improved combined with historical peak information.The update of the current template is judged.When the algorithm judges that the current response map meets the update conditions,the tracking template and adaptive penalty coefficient are updated.When the algorithm judges that the current response map does not meet the update conditions,it skips the frame and does not update the tracking template and adaptive penalty coefficient,so as to avoid the tracking template and penalty coefficient being polluted by error information.Comparative experiments based on the published benchmark video set show that this method further improves the tracking performance of the tracking method in complex scenes,especially in chaotic scenes.