汽车工程 ›› 2023, Vol. 45 ›› Issue (7): 1222-1234.doi: 10.19562/j.chinasae.qcgc.2023.07.013

所属专题: 车身设计&轻量化&安全专题2023年

• 精选论文 • 上一篇    下一篇

基于图像识别与动力学融合的路面附着系数估计方法

张雷,关可人,丁晓林(),郭鹏宇,王震坡,孙逢春   

  1. 1.北京理工大学,北京电动车辆协同创新中心,北京  100081
    2.北京理工大学,电动车辆国家工程研究中心,北京  100081
  • 收稿日期:2022-12-08 修回日期:2023-01-03 出版日期:2023-07-25 发布日期:2023-07-25
  • 通讯作者: 丁晓林 E-mail:dingxiaolin@bit.edu.cn

Tire-Road Friction Estimation Method Based on Image Recognition and Dynamics Fusion

Lei Zhang,Keren Guan,Xiaolin Ding(),Pengyu Guo,Zhenpo Wang,Fengchun Sun   

  1. 1.Beijing Institute of Technology,Collaborative Innovation Center for Electric Vehicles in Beijing,Beijing  100081
    2.Beijing Institute of Technology,National Engineering Research Center for Electric Vehicles,Beijing  100081
  • Received:2022-12-08 Revised:2023-01-03 Online:2023-07-25 Published:2023-07-25
  • Contact: Xiaolin Ding E-mail:dingxiaolin@bit.edu.cn

摘要:

准确的路面附着系数估计是车辆主动安全控制的前提。首先,建立了单轮动力学模型,利用卡尔曼滤波实现了轮胎纵向力精准估计,并结合魔术轮胎模型建立了基于粒子滤波的路面附着系数估计器;其次,提出了基于图像识别的前向路面附着系数预测方法,通过DeeplabV3+、语义分割网络和MobilNetV2轻量化卷积神经网络实现路面分割和路面类型辨识,并利用查表获取前向路面附着系数。最后,建立了图像识别与动力学估计时空同步方法和融合规则,实现了两种估计方法的有效关联与可靠融合。CarSim-Simulink联合仿真表明,本文所提出的基于图像识别与动力学融合的方法可有效提高不同工况下的路面附着系数估计精度。

关键词: 路面附着系数, 粒子滤波, 图像识别, 语义分割, 卷积神经网络

Abstract:

Accurate estimation of tire-road friction is a prerequisite for vehicle active safety control. Firstly, a single-wheel dynamics model is established, and precise estimation of the longitudinal tire force is realized using the Kalman filter. Then a particle filter (PF)-based tire-road friction estimator is developed based on the Magic Formula tire model. Secondly, a forward road adhesion coefficient prediction method based on image recognition is proposed. The DeeplabV3+, semantic segmentation network and the MobilNetV2 lightweight convolutional neural network are used for road segmentation and classification, based on which the tire-road friction is obtained through table look-up. Finally, the spatiotemporal synchronization method and fusion mechanism of dynamics and image recognition are established to realize effective correlation and reliable fusion of the two estimation methods. The Carsim-Simulink co-simulation shows that the proposed estimation method based on image recognition and dynamics fusion can efficiently improve the tire-road friction estimation accuracy.

Key words: tire-road friction, particle filter, image recognition, semantic segmentation, convolutional neural network