Research On Object Tracking Based On Siamese Network

With the rapid development of the multimedia technologies,visual data including images and videos have rapidly increased,which contain much valuable information.Studying such information by computer vision technology can make life of people more convenient and safer.As a classic task in the field of computer vision,object tracking aims to infer the location in subsequent frames given its initial bounding box in a frame of a video.Object tracking has great practical value in autonomous vehicles,security system,etc.This paper deeply studied object tracking based on Siamese network in deep learning,and propose three improvements in terms of enhancing feature extraction of backbone network,improvements for tracking performance under limited training resource,and improving adaption for appearance changes of target.The main work and corresponding conclusions are summarized as follows:1.Propose Siamese network for tracking with multi-stage training.For the problem of feature extraction of backbone network,this paper adapts network structure and training methods to capture features which are more applicable to object tracking.When applying Siamese networks in object tracking,it is necessary to consider the marginal effect introduced by deepening network.Since object tracking has a high requirement on real-time speed,we need to balance the relationship between size and precision of networks.As a result,we explore optimum structure design by long-short period tracking module.Moreover,we propose a multi-stage training method,which uses single frame and multiple frames to train long-time tracking module and short-time tracking module respectively.By this way features captured from network are more applicable for tracking to further design suitable network structure for object tracking.2.Propose Siamese network for tracking with action-selection.For the problem of improving tracking performance under limited training resource,we propose to embed a stride-varied action-selection mechanism into Siamese networks,to obtain bounding boxes effectively and flexibly.Bounding box of object from the last frame choose a series of actions by actionselection mechanism to obtain the optimal position in current frame.Before conducting each move,a similarity metric is used to determine the value of stride.After performing action we need to filter action set to reduce computation burden,and compare similarity of sub-regions in candidate areas by region of interests pooling,to infer accurate location for target without amount of training data.Besides,this paper compares action-selection mechanism with Gaussian sampling,sliding window sampling,regression and other sampling methods in terms of characteristics and differences,and shows its superior in public databases.3.Propose Siamese network for object tracking with multi-granularity appearance representations.For the problem of adaption to object appearance changes,we propose an updatable object tracking algorithm,which combines a generative model and a discriminative model to capture semantic features and appearance features respectively.During tracking,we fixed robust semantic features and update appearance model.This paper apply a convolutional block attention module to enhance semantic features and obtain multigranularity appearance features by color histogram to describe object changes in detail,and update appearance template pool selectively by preserving recent appearance of targets.Furthermore,this paper proposes an adaptive model fuse method,which measure historic performance by Hedge algorithm and estimate current performance by drift movements in current frame,and then integrate two metrics by selective traverse method.The tracker proposed in this paper can perform well under occlusion,deformation and other conditions of appearance changes substantially.