Yang Dongsheng, Zhang Dongmin, Yuan Yi, Lei Zhaoyu, Ding Binlei, Bo Lei
New Technology Research Institute, BYD Auto Industry Co., Ltd., Shenzhen, 518118, China.
Sci Rep. 2024 Dec 28;14(1):30913. doi: 10.1038/s41598-024-81666-7.
Effective road terrain recognition is crucial for enhancing the driving safety, passability, and comfort of autonomous vehicles. This study addresses the challenges of accurately identifying diverse road surfaces using deep learning in complex environments. We introduce a novel end-to-end Tire Noise Recognition Residual Network (TNResNet) integrated with a time-frequency attention module, designed to capture and leverage time-frequency information from tire noise signals for road terrain classification. Our method was evaluated on five distinct road types: asphalt, cement, grass, mud, and sand. The performance of TNResNet was rigorously compared against traditional machine learning techniques, including Decision Trees, K-Nearest Neighbors, and Support Vector Machines, as well as advanced deep learning models like Long Short-Term Memory and Convolutional Neural Networks. Experimental results demonstrate that TNResNet achieves superior classification accuracy of 99.48%, outperforming all comparative methods. This work not only establishes a robust framework for road terrain identification but also showcases the significant practical implications of TNResNet in the realm of autonomous vehicle navigation.
有效的道路地形识别对于提高自动驾驶车辆的行驶安全性、通过性和舒适性至关重要。本研究旨在应对在复杂环境中使用深度学习准确识别各种路面的挑战。我们引入了一种新颖的端到端轮胎噪声识别残差网络（TNResNet），该网络集成了时频注意力模块，旨在从轮胎噪声信号中捕获和利用时频信息进行道路地形分类。我们的方法在五种不同的道路类型上进行了评估：沥青路、水泥路、草地、泥地和沙地。TNResNet的性能与传统机器学习技术（包括决策树、K近邻和支持向量机）以及先进的深度学习模型（如长短期记忆网络和卷积神经网络）进行了严格比较。实验结果表明，TNResNet实现了99.48%的卓越分类准确率，优于所有比较方法。这项工作不仅建立了一个强大的道路地形识别框架，还展示了TNResNet在自动驾驶车辆导航领域的重大实际意义。
