Research On Visual Object Tracking Based On Bilinear Fusion

With the development of technologies such as the Internet of Things,big data and artificial intelligence,visual object tracking is still a research hotspot in the field of computer vision.It has a wide range of application values,such as driverless,intelligent monitoring,medical diagnosis and behavior analysis.The visual object tracking task is to predict the position and size of the target in the subsequent video sequence accurately and reliably given the target position information of the initial frame.There are often many challenges in the visual object tracking process,such as occlusion between objects,blurred appearance of the object,rotation and illumination changes.These complex factors put forward a greater challenge to the discriminative ability of the visual object tracking algorithm.In response to this scientific issue,the research content of this paper is carried out in the following two aspects:(1)In order to solve the problem of tracking drift easily based on single response map,this paper selects the regression model based on convolutional neural network as the basic framework of our algorithm,and adds an innovative method of response map fusion,called bilinear fusion.The convolutional regression model is an improvement of the traditional correlation filter model,which can easily realize end-to-end network training.Therefore,this paper chooses to improve the algorithm under this framework.Different from the simple element-by-element fusion method,the bilinear fusion method can realize the position association and information interaction between different response maps,so as to predict the position of the target more accurately according to the response map,and improve the robustness of the algorithm effectively.(2)Based on the proposed visual object tracking algorithm,this paper further designs a user interaction prototype system for visual object tracking.The system includes a parameter adjustment module,an object tracking module,a statistical result display module and an arbitrary frame coordinate display module.This paper designs a user interaction interface based on MATLAB platform,integrates each module and forms a complete prototype system to meet the needs of user interaction.It can realize parameter adjustment of object tracking algorithm,observe target tracking status in real time,obtain the position and size of any frame of target and obtain the statistics of the video sequence accuracy and speed.