Target Tracking Algorithm for Mobile Robots Based on Siamese Neural Networks

Abstract

Target tracking faces various challenges in natural scenes, like changes in lighting, scale changes, occlusion, etc. In response to these issues, research is based on the use of Siamese neural networks for target tracking. And this paper uses channel attention mechanisms and high-quality bounding box models to improve the target tracking algorithm. Firstly, by combining attention mechanism with twin neural networks, useful channel information is enhanced and useless channel information is suppressed. Then, based on this, the classification branch and quality prediction branch are jointly represented, and the bounding box is represented through a general distribution form to further optimize the algorithm. The algorithm optimized based on attention mechanism has a success rate of 71.2% and an accuracy of 91.3%, which has better performance compared to other comparative algorithms. The algorithm optimized based on high-quality bounding box models has a success rate of 73.6% and an accuracy of 92.5%, which further improves its performance compared to attention mechanism optimization. The improvement strategy proposed in the study has improved the robustness and accuracy of target tracking, and the improved algorithm has better robustness and performance advantages in complex tracking scenarios. Meanwhile, the practical applicability of this algorithm has been effectively verified through the simulation experiments of robot control systems.