汽车工程 ›› 2024, Vol. 46 ›› Issue (7): 1208-1218.doi: 10.19562/j.chinasae.qcgc.2024.07.008

• • 上一篇    

面向自动驾驶场景的多目标点云检测算法

陶乐1,王海1(),蔡英凤2,陈龙2   

  1. 1.江苏大学汽车与交通工程学院,镇江 212013
    2.江苏大学汽车工程研究院,镇江 212013
  • 收稿日期:2024-01-23 修回日期:2024-02-29 出版日期:2024-07-25 发布日期:2024-07-22
  • 通讯作者: 王海 E-mail:wanghai1019@163.com
  • 基金资助:
    国家重点研发计划项目(2023YFB2504401)

Multi-object Detection Algorithm Based on Point Cloud for Autonomous Driving Scenarios

Le Tao1,Hai Wang1(),Yingfeng Cai2,Long Chen2   

  1. 1.School of Automotive and Traffic Engineering,Jiangsu University,Zhenjiang  212013
    2.Automotive Engineering Research Institute,Jiangsu University,Zhenjiang  212013
  • Received:2024-01-23 Revised:2024-02-29 Online:2024-07-25 Published:2024-07-22
  • Contact: Hai Wang E-mail:wanghai1019@163.com

摘要:

基于点云的三维目标检测算法是自动驾驶系统中关键技术之一。目前基于体素的无锚框检测算法是学术界的研究热点,但是大多数研究都侧重于设计复杂的二阶段修正模块,在牺牲巨大的算法延迟的情况下带来有限的算法性能提升。而单阶段无锚框点云检测算法虽然具有更加精简的检测流程,但其检测性能难以满足自动驾驶场景的需求。对此,本文基于无锚框检测算法CenterPoint,提出了一种面向自动驾驶场景的单阶段无锚框点云目标检测算法。具体来说,本文通过引入编码解码稀疏模块,极大地促进了三维特征提取器对于空间非连通区域的信息交互,保证了三维特征提取器能够提取到满足各类目标检测的特征。此外,考虑到现有的二维特征融合主干与基于中心点的无锚框检测头的适配存在挑战性,本文通过引入自校正卷积和大核注意力模块,能够有效提取到目标区域的点云特征,并将目标区域的点云特征聚集到中心点区域,从而提升算法对于目标的召回率和检测精度。本文所提出的算法在大规模公开数据集nuScenes上进行模型训练和实验验证,与基准算法相比,mAP和NDS分别提升了5.97%和3.62%。同时,本文将所提出的算法在基于自主搭建的实车平台上进行实际道路实验,进一步证明了所提出算法的有效性。

关键词: 自动驾驶, 深度学习, 点云检测, 无锚框

Abstract:

The three-dimensional object detection algorithm based on point cloud is one of the key technologies in the autonomous driving system. Currently, the voxel-based anchor-free detection algorithm is a research hotspot in academia, but most researches focus on designing complex refinement stage, at the expense of huge algorithm latency, to bring limited performance improvement. Although the single-stage anchor-free point cloud detection algorithm has a more streamlined detection process, its detection performance cannot satisfy the needs of autonomous driving scenarios. In this regard, based on the anchor-free detection algorithm CenterPoint, a single-stage anchor-free point cloud object detection algorithm for autonomous driving scenarios is proposed in this paper. Specifically, the encoding and decoding sparse module is introduced in this paper, which greatly promotes the information interaction of the spatial non-connected areas of the three-dimensional feature extractor, ensuring that the three-dimensional feature extractor can extract features that satisfy various target detection. In addition, considering that it is challenging to adapt the existing two-dimensional feature fusion backbone to the center-based head, in this paper self-calibrated convolution and large kernel attention modules are introduced in to effectively extract point cloud features of the target area, which are then gathered into the center point area, thereby improving the algorithm's recall and accuracy of the target. The proposed algorithm in this article is trained and experimentally verified on the large-scale public dataset of nuScenes. Compared with the benchmark algorithm, mAP and NDS are increased by 5.97% and 3.62% respectively. At the same time, the actual road experiments with the proposed algorithm are conducted on a self-built vehicle platform, further proving the effectiveness of the proposed algorithm.

Key words: autonomous driving, deep learning, point cloud detection, anchor-free