Visual Tracking Based On Convolutional Neural Network Feature Sharing And Object Detection

Visual tracking is to track the target in the video with the corresponding position and size of target according to the initial detection result.Traditional tracking algorithms face the challenges due to occlusion and scale changes.When the target suffers from large occlusion or violent appearance's changes,the tracking result may not be robust.Since the object detection could provide useful information for target re-identification,it can help visual tracking to handle the occlusion and scale change problems.Therefore,this thesis introduces the target detection in the visual tracking.Deep Learning can automatically extract more powerful features that perform target detection and tracking accurately with high stability.So our method adopts the convolutional neural network method for visual tracking.In this thesis,we propose a visual tracking algorithm with feature sharing and object detection.By introducing the detection network,we can effectively solve the problem of scale change and occlusion in tracking.In addition,we propose to design a feature sharing network where the low-level features and high-level features in the network are used for both detection and tracking.The detected result are used to provide the candidates for tracking.As shown in our experimental results,the proposed method achieves more robust tracking results when compared to the existing tracking methods.