Journal of Computer Applications ›› 0, Vol. ›› Issue (): 0-0.DOI: 10.11772/j.issn.1001-9081.2019122227

    Next Articles

Development of 3D-object Detection for On-road obstacle based on Deep Learning Method

  

  • Received:2020-01-05 Revised:2020-02-25 Online:2023-07-27 Published:2023-07-27
  • Contact: Wei-Hong ZHENG

基于深度学习的道路障碍物检测方法研究

彭育辉,郑玮鸿,张剑锋   

  1. 福州大学
  • 通讯作者: 郑玮鸿
  • 基金资助:
    福建省科技厅产学合作重大项目

Abstract: Abstract: With rapid development of autonomous driving technology in automobile industry, Lidar based approaches for on-road obstacle detection play a more critical role than ever. Based on various point cloud data of on-road vehicles acquired by mobile Lidar, this paper proposes an on-road obstacle detection algorithm using deep learning. Firstly, the clustering algorithm is applied to filter and eliminate the outliers from the raw data, improving the roughness of point clouds. Then, an end-to-end trainable deep network is proposed to construct the object detection network, by optimizing RPN (region proposal network) and adding a max pooling layer. Experiments are conducted to show that the proposed approach significantly improves the accuracy of 3D-objects detection. For easy, moderate and hard detection tasks on orientation in the KITTI benchmark, the average precisions are improved by11.30%, 6.02% and 3.89%, respectively, compared with the VoxelNet.

Key words: autonomous vehicles, deep learning, mobile Lidar, 3D-object detection, point cloud

摘要: 摘 要: 随着汽车无人驾驶技术研究和应用受到业内的普遍关注,基于激光雷达的道路障碍物检测逐渐成为感知系统的重要研究内容。基于车载激光雷达获取的外部车辆三维点云数据,提出一种基于深度学习的路障碍物检测方法。首先,采用聚类算法对原始点云进行离群点过滤、剔除处理,提高点云的光整度;其次,提出一种端到端的深度神经网络,利用最大池化对RPN架构进行优化,构建改进的目标检测层;实验测试结果表明,提出的算法在三维目标的检测精度方面得到显著提高,尤其在KITTI数据集的简单、中等和困难任务的车辆定位中,平均精度比Voxelnet(Unofficial)提高分别提高了11.30%、6.02%和3.89%。

关键词: 无人驾驶, 深度学习, 激光雷达, 目标检测, 三维点云

CLC Number: