Research And Application Of Single Target Tracking Algorithm Based On Deep Learning In Complex Scene

Vision plays an important role in daily production life,so that machines like humans have vision to help people deal with some heavy,dangerous tasks have great application prospects,but also naturally promote the emergence and development of the field of computer vision,in which single object tracking has gradually become a research hotspot and key issues in the field of computer vision.The technology has been widely used in both military and civilian fields,and the penetration rate is also increasing,which largely enhances our national defense force and improves people’s quality of life.At present,it is more popular to use deep learning to track single targets in complex scenes in real time.Although the method has been significantly improved in terms of speed and accuracy compared with traditional tracking algorithms,there are still many problems,such as difficulties in tracking targets with small pixel areas,the large amount of calculation results in low real-time performance of the algorithm or even can not achieve real-time performance,correlation calculation the problems such as the serious loss of semantic information and the inability to capture global information during feature extraction.Through an in-depth analysis of the above problems,the following innovative points and research works are proposed and conducted in this thesis.1.A small object tracking algorithm based on bidirectional feature pyramid fusion framework in complex scenes is proposedTo address the tracking challenges such as weak feature expression ability of small objects and susceptibility to interference by similar objects in complex backgrounds,this thesis designs a bidirectional feature pyramid fusion framework on the backbone network,which is mainly composed of two parts: the forward feature pyramid and the reverse feature pyramid.The reverse feature pyramid is mainly for reinforcing semantic and contextual information,while the forward feature pyramid is for reinforcing location and spatial structure information for repairing the internal structure details of small targets and enhancing their localization ability.The differences in feature types between the two can both complement each other and meet the differential information requirements for classification and regression,and enhance the features of small targets.2.Proposed object tracking algorithm combining Transformer and Saliency encoderSiamese network-based tracking algorithms are highly prone to lose semantic information as well as detailed features of objects when applying correlation operations,and they lack global modeling capabilities to track objects in multiple complex scenarios.To address the above problems,this thesis proposes a feature repair strategy combining Transformer and saliency encoder to repair the feature loss from correlation operation.The strategy designs saliency encoder network to repair the severe loss of detailed features and semantic information brought about by correlation operations and reduce the interference of invalid targets.The encoder part of Transformer is also used to capture the contextual information with global association ability to strengthen the classification and localization ability,and to optimize the generation of Bounding Box in combination with DIo U loss function.3.Designed and implemented a single object tracking system based on Siamese network algorithmBased on the above algorithm,this thesis designs and implements a web-based single object tracking system with four functional modules: login/registration,online tracking,offline tracking and tracking record.Users can choose two tracking modes,online and offline,after logging into the system through the web terminal,and the system will automatically record the tracking information of these two modes.