汽车工程 ›› 2024, Vol. 46 ›› Issue (11): 2017-2027.doi: 10.19562/j.chinasae.qcgc.2024.11.008

• • 上一篇    下一篇

地下矿道无人车可行驶区域检测算法

陈志军1,2,王朝伟1,2,吴超仲1,3,4(),钱闯1,吴怀主5,申广俊5   

  1. 1.武汉理工大学智能交通系统研究中心,武汉 430063
    2.武汉理工大学计算机与人工智能学院,武汉 430070
    3.交通信息与安全教育部工程研究中心,武汉 430063
    4.湖北文理学院,襄阳 441053
    5.东风汽车有限公司东风商用车技术中心,武汉 430056
  • 收稿日期:2024-06-07 修回日期:2024-07-09 出版日期:2024-11-25 发布日期:2024-11-22
  • 通讯作者: 吴超仲 E-mail:wucz@whut.edu.cn
  • 基金资助:
    国家自然科学基金(52332010);湖北省重点研发计划项目(2022BAA078);武汉市科技计划项目(2023010402040022)

Algorithm for Detecting Free Space in Underground Mine Tunnels for Autonomous Vehicles

Zhijun Chen1,2,Chaowei Wang1,2,Chaozhong Wu1,3,4(),Chuang Qian1,Huaizhu Wu5,Guangjun Shen5   

  1. 1.Intelligent Transportation Systems Research Center,Wuhan University of Technology,Wuhan 430063
    2.School of Computer Science and Artificial Intelligence,Wuhan University of Technology,Wuhan 430070
    3.Engineering Research Center of Transportation Information and Safety,Ministry of Education,Wuhan 430063
    4.Hubei University of Arts and Science,Xiangyang 441053
    5.Dongfeng Commercial Vehicle Technical Center,Dongfeng Commercial Vehicle Co. ,Ltd. ,Wuhan 430056
  • Received:2024-06-07 Revised:2024-07-09 Online:2024-11-25 Published:2024-11-22
  • Contact: Chaozhong Wu E-mail:wucz@whut.edu.cn

摘要:

地下矿道可行驶区域检测是地下矿山自动驾驶系统的关键感知技术,然而地下矿道光照强度低、工况复杂的特点给该任务带来极大挑战。鉴于此,本文提出一种地下矿道可行驶区域检测算法。首先,为解决地下矿道细节退化导致图像特征难以提取的问题,提出一种双分支特征提取骨干网络;然后,针对地下矿道可行驶区域检测不完整问题,提出一种自适应多尺度空间空洞池化金字塔特征增强模块;最后,为解决地下矿道边界提取不准确的问题,设计一种双分支通道注意力机制融合模块。在自制矿道可行驶区域数据集上进行实验,相较于Deeplabv3+、UNet、DDRNet-23、PIDNet,本文算法取得最佳效果,在MIoU分数上分别提升2.07、2.39、1.87、1.92个百分点,在mAcc分数上分别提升1.78、2.45、1.84、1.86。本文算法已成功应用于地下无人驾驶矿车,验证了其在真实矿道场景下的有效性。

关键词: 自动驾驶, 井工无人矿车, 可行驶区域检测, 语义分割

Abstract:

The detection of free space in underground mine tunnels is the key sensing technology for underground mining autonomous driving systems. However, the characteristics of low illumination and complex working environment inside the tunnels bring great challenges to this task. In view of this, in this paper an algorithm for detecting free space in underground mine tunnels is proposed. Firstly, a dual-branch feature extraction backbone network is proposed to solve the problem of difficulty in extracting image features caused by the degradation of tunnel details. Secondly, for the problem of incomplete detection of drivable areas in underground mining tunnels, an adaptive multi-scale atrous spatial pyramid pooling feature enhancement module is proposed. Finally, a dual-branch channel attention mechanism fusion module is developed to solve the problem of inaccurate boundary extraction in the underground mine tunnels. The experiments are conducted on a self-made dataset specifically designed for underground mine tunnels. The results show that the proposed algorithm surpasses other existing methods such as Deeplabv3+, UNet, DDRNet-23, and PIDNet, with an increase of 2.07, 2.39, 1.87, and 1.92 percentage points in terms of MIoU scores, and 1.78, 2.45, 1.84, and 1.86 in terms of mAcc scores, respectively. The effectiveness of the proposed algorithm has been validated through its successful application in real mine tunnel scenarios, particularly for underground mining autonomous driving vehicles.

Key words: autonomous driving, unmanned mine truck, free space detection, semantic segmentation