Research And Implementation Of Target-aware Tracking System Based On Siamese Networks

Target tracking is an important research topic in computer vision,where given the target state in the first frame,the tracker needs to predict the position and scale change of the target in subsequent frames.It is widely used in military and civilian applications,such as intelligent driving,human-computer interaction,intelligent video surveillance,medical diagnosis,and smart cities.However,visual tracking still faces serious challenges due to partial occlusion,scale variation,background clutter,motion blur and illumination changes.In recent years,Siamese trackers have been widely used in target tracking.Most Siamese trackers compute target templates and search images separately,and the tracker will fail to recognize the target when there are a lot of appearance changes in the target,and to solve this problem,we propose an effective Siamese target-aware algorithm.First,we adaptively encode the target information into the search image to improve the network discrimination,and then adaptively contribute valuable features by spatial and channel locations.Finally,to improve the target tracking frame accuracy,we propose dual response map fusion to obtain the final results.Experiments on challenging publicly available datasets such as GOT-10 k,UAV123,La SOT,and OTB100 show that the proposed algorithm outperforms many state-of-the-art trackers,running at about 45 fps,meeting real-time requirements.Based on the study of the above algorithms,a Siamese target-aware tracking system is implemented,and the implementation of the system interface is completed using Py QT5.According to the system feasibility analysis and system requirement analysis,the system implementation mainly consists of three modules,including user management module,image management module and target tracking module.The user logs into the system after completing registration,and uses the camera and video files to obtain video data as the input for target tracking.The system uses the tracking algorithm to complete the tracking by feature extraction and target matching of the target,and realizes the real-time tracking of the target in the video by predicting and updating the target position and motion status.Meanwhile,the system is tested for its functionality and stability,and the results show that the system achieves the expected goal.