Obstacle Recognition with Ultra-Wideband Based on Integrated Learning

• Main Authors: Yang, Anning, Zhou, Jinglong, Li, Wenfeng
• Format: Conference Proceeding
• Language: English
• Subjects: channel impulse response; Feature extraction; Line-of-sight propagation; obstacle recognition; Prediction algorithms; Reliability; Sensors; spatial perception; Time series analysis; Training; ultra-wideband; XGBoost
• ISSN: 2766-8665
• DOI: 10.1109/ICNSC58704.2023.10318845

The error suppression and compensation of conventional ultra-wideband (UWB) ranging are less effective in non-line-of-sight (NLOS) environments. The contribution of obstacles that cause NLOS range errors to environmental perception is neglected. In order to achieve a comprehensive perception and understanding of the environment under dynamic positioning scenarios, this work proposes a UWB-based obstacle recognition method. Firstly, the channel impulse response (CIR) curve and communication bottom layer characteristics of UWB in a communication process are analyzed. The key parameters of signal features are extracted and combined with the integrated learning XGBoost classifier. Further, an efficient NLOS obstacle recognition method based on a loss function probability matrix weighted prediction label is proposed. The predicted probabilities of the XGBoost loss function are used as the weights of the labels, and the UWB data of continuous time series is weighted average. This algorithm mitigates the effect of low probability data on the overall prediction results. It also helps to improve the obstacle recognition accuracy. Finally, experiments under dynamic and static conditions are carried out to verify the reliability of the proposed method. The experimental results show that the recognition accuracy of line-of-sight (LOS) and NLOS under static conditions reaches 96.00%. Under different sampling steps, the average recognition accuracy of the three common obstacles, including wood boards, iron plates, and human body, is more than 95.63%. Compared with the origin, the average recognition accuracy of the proposed algorithm in this paper is improved by 18.87%.