Research On Context-aware Correlation Filter Object Tracking Algorithm

Object tracking is an important research direction in computer vision.It has a wide range of applications,such as video surveillance,human-computer interaction,and driverless driving.In the past two or three decades,the visual object tracking technology has made great progress.In particular,the object tracking method using deep learning in the past two years has achieved satisfactory results,and the object tracking technology has made breakthrough progress.object(single object)tracking is the prediction of the size and position of the object in subsequent frames given the object size and position of the initial frame of a video sequence.Moving object tracking is a very challenging task,because for moving objects,the motion scene is very complex and often changes,or the object itself is constantly changing,how to identify and track the changing scene in complex scenes.Goals are a challenging task.The object tracking method can be classified into a Generative Method and a Discriminative Method according to whether the observation model is a generative model or a discriminant model.The generation tracking method has been popular in previous years,and the recent discriminant tracking method has gradually occupied the mainstream position.The discriminant method represented by Correlation Filter and Deep Learning has also achieved satisfactory results.This paper improves the context-aware Correlation Filter(CACF)object tracking algorithm proposed by Mueller from two perspectives.The main research work is as follows:(1)When constructing the filter model for the original text algorithm,the context area around the target is treated equally.The generated filter model is not robust enough for fast motion and scale change.For this reason,the improved algorithm uses sparse optical flow.The method creates a context weight matrix,reconstructs the filter model with the weighted context,and finally verifies the effectiveness of the improved algorithm by comparing with other algorithms.(2)For the tracking instability phenomenon of the original algorithm model target in the background chaos,similar object interference andtarget deformation and scale change,the background suppression model and the color probability model are used to determine the target estimation position respectively.Finally,the two types are determined.The target position determined by the model is linearly weighted according to the respective response scores to determine the final position.The feasibility of the proposed algorithm is proved by comparative experiments.