Research On Object Tracking With Fully Conventional Anchor-Free Siamese Network

Visual object tracking is one of the most important tasks in computer vision.It plays an important role in other researches such as assistant driving system,target behavior analysis,and intelligent video surveillance and control.The purpose of visual object tracking is to continuously and accurately estimate the position,shape change or occupied area of the target in the video sequence,and determine the motion information such as the speed,direction and trajectory of the target,so as to complete more difficult tasks.Caused by partial or complete occlusion,scale deformation,background clutter and many other factors,it is still a great challenge to build a general and robust visual tracking system.In this paper,we propose an efficient framework which is end-to-end trained offline,for real-time object tracking: fully conventional anchor-free Siamese network(FCAF).Specifically,the deep siamese Res Net is adopted,as the backbone network in the Siamese trackers is relatively shallow which leads to fewer features acquired by the tracker and lower accuracy,to provide richer feature representation for the tracker.Meanwhile,the introduction of multi-layer feature fusion effectively combines low-level detail information with high-level semantic information to solve the localization problem.In addition,we propose the anchor-free proposal network(AFPN)to replace the RPN network.The AFPN network consists of correlation section,using depth-wise cross correlation,and supervised section which has two branches,one for classification and the other for regression.In order to reduce the prediction of low-quality bounding boxes,the center-ness branch is added.We conduct extensive experiments on the OTB-2015 and VOT-2016 public datasets,demonstrating the effectiveness of these improvements and that our proposed tracker achieves state-of-the-art performance.