| The research on artificial intelligence technology has been conducted for more than 60 years,and its applications have spread to all corners of human life.Among them,computer vision,as an important scientific research field,has been used in more and more field projects in military reconnaissance,UAV following,intelligent surveillance,autonomous driving,etc.However,target tracking technology as an application in computer vision,its algorithm for realtime,accuracy and robustness problems of decision-making is not fully mature.With the development of technology and technological advances,more and more challenging application scenarios are being explored.So far,most tracking algorithms are still unable to solve well the problems that arise in the actual application scenarios that have been explored,such as: light transformation,plane rotation,morphological changes,low resolution,fast movement,blurred movement,target occlusion,etc.In recent years,single-target tracking algorithms based on deep learning have achieved good experimental results in terms of experiments.Compared with traditional tracking methods,deep learning-based target tracking systems can not only locate the specified target accurately,but also have a large improvement in real-time performance.Although most of the current deep tracking algorithms have achieved good tracking results,there is still much room for improvement.In this paper,we study single-target tracking algorithms based on deep learning,and optimize the existing mainstream tracking algorithm models from the perspective of multi-view features and target segmentation to improve the robustness of single-target trackers in complex scenarios.The research in this paper focuses on the following two aspects.(1)In order to improve the robustness of single-target tracking,a single-target tracking method that incorporates detection and prediction information and Multi-view Collaborative Matching(MCM)is proposed.The method introduces a target prediction mechanism and a multi-view feature system,and implements a single-target tracker by extracting multi-view features from the candidate targets obtained by detection and prediction and performing collaborative matching.Compared with most mainstream tracking algorithms based on deep learning,the multi-view features extracted by the proposed method MCM are robust to different complex scenes and have a greater advantage in terms of accuracy rate.(2)Although the above MCM method have better robustness,it do not well in real-time.In order to achieve single-target tracking with both robustness and real-time performance for complex scenes,a single-target tracking algorithm based on Segmentation and Domain Feature Matching(SDFM)is proposed.In this method,by improving UNet segmentation network,a twin segmentation network that can assign multiple labels is obtained.Then,this network is combined with SiamMask,and a segmentation correction module is proposed to solve the problem that SiamMask is not robust in target occlusion,target rotation and other scenes without affecting the real-time performance of the tracker.Compared with the original SiamMask method,SDFM not only does not unduly degrade the real-time performance of the tracker,but also improves the accuracy rate in the application scenarios such as target occlusion. |
