Robot Visual Tracking Model Based on Improved GOTURN-LD Algorithm

Abstract

With the aging of the population and the shortage of labor, the demand for service robots is increasing. As the key performance of visual tracking, it still has the problems of low tracking accuracy and poor real-time performance. Therefore, this paper studies the use of the You Only Look Once series algorithm and the use of the regression network general target tracking algorithm to improve the detection and tracker part of the track-learning-detection algorithm. At the same time, three strategies for the attention mechanism, multi-level feature fusion, and regional overlap loss function are introduced to improve the visual tracking model. Therefore, a new robot visual tracking model is constructed. The experimental results showed that the average score of the research model in the success rate index was 0.68, which was significantly higher than other models. In the precision graph index score, the average score of the research model was 0.79 and the curves were all in the outermost circle. In practical application, the Expected Average Overlap rate index value of the research model was 0.4027 and the average mean Average Precision value was 91.07%. Compared with the other four models, its comprehensive performance was significantly better. The research model can maintain high robustness under different challenges and the visual tracking accuracy and real-time performance are better. It can provide more effective technical support for the service field of robots and promote the development of the robot field.

Keywords

• GOTURN algorithm
• TLD algorithm
• robots
• visual tracking
• real-time performance
• improvement strategy