Global Target Tracking Algorithm Based on Improved Twin Neural Network

The target tracking algorithm based on twin neural network is a classical deep learning target tracking method, and has attracted much attention due to its superior and stable tracking performance. GlobalTrack is a tracker for global instance search based on twin neural network architecture. Tracking globally can greatly improve the success rate of tracking. However, when the tracking target has motion blur, fast motion and other challenges, the tracker will lose track. To solve the above problems, a global target tracking algorithm based on improved twin neural network is proposed. Swin Transformer with better feature extraction ability, is introduced as the backbone network of feature extraction to improve the ability of target feature extraction. A regional proposal network that guides anchor points is introduced to improve the quality of generation of candidate frames for models in motion blurred scenes. A regional convolution neural network is used, According to the ROI characteristics of the target template, the prediction of its label and border is refined, and then the final prediction is obtained by classification and regression. Through a large number of experiments on OTB2015 and VOT2018 datasets, the tracking effect of this method is good, and it has strong robustness and adaptability in a variety of complex situations such as motion blur, fast motion, and meets the real-time requirements of the tracking algorithm.