汽车工程 ›› 2025, Vol. 47 ›› Issue (6): 1198-1206.doi: 10.19562/j.chinasae.qcgc.2025.06.018

• • 上一篇    

基于极坐标的环视视觉稀疏化时序3D目标检测

魏超1,2,随淑鑫1(),李路兴1   

  1. 1.北京理工大学特种车辆研究所,北京 100081
    2.特种车辆设计制造集成技术全国重点实验室,北京 100081
  • 收稿日期:2024-10-16 修回日期:2024-12-20 出版日期:2025-06-25 发布日期:2025-06-20
  • 通讯作者: 随淑鑫 E-mail:13021028800@163.com
  • 基金资助:
    第二十七届中国科协年会学术论文。国家自然科学基金(52002026)

PolarSparse4D: Polar Parametrization for Vision-Based Surround-View Temporal Sparse 3D Object Detection

Chao Wei1,2,Shuxin Sui1(),Luxing Li1   

  1. 1.The Special Vehicle Research Institute,Beijing Institute of Technology,Beijing 100081
    2.National Key Laboratory of Special Vehicle Design and Manufacturing Integration Technology,Beijing 100081
  • Received:2024-10-16 Revised:2024-12-20 Online:2025-06-25 Published:2025-06-20
  • Contact: Shuxin Sui E-mail:13021028800@163.com

摘要:

在自动驾驶领域,针对基于环视视觉的3D目标检测方法准确性和实时性之间的矛盾,本文提出了一种极坐标参数化的基于稀疏查询的时序3D目标检测方法PolarSparse4D,该模型由图像编码器、3D锚框解码器以及辅助训练的质量检测分支组成。首先为避免参数归一化带来的检测距离限制,设计了3D锚框中心距离与方位角参数解耦的特征编码方式。其次,通过设计锚框空间信息交互自注意力模块以及锚框时序信息融合模块,高效高精度地完成了极坐标系下环视相机图像时空信息融合过程。最后,通过设计锚框参数质量检测分支,显著提高了检测精度和模型收敛速度。在nuScenes数据集上进行实验验证,本文模型的mAP和NDS均得到了极大的提升,分别为41.3%和52.5%,模型速度为19.2 FPS,证明了本方法在精度和速度方面的优越性和有效性。

关键词: 3D目标检测, 环视视觉, 极坐标参数化, 自动驾驶

Abstract:

To address the trade-off between accuracy and real-time performance in vision-based surround-view 3D object detection for autonomous vehicles, PolarSparse4D, a sparse query-based method using polar parametrization, is proposed. The model consists of an image encoder, a 3D anchor decoder and an auxiliary quality assessment branch for training. Firstly, to avoid the detection distance limitation caused by parameter normalization, a feature encoding method that decouples the center distance and azimuth angle of the 3D anchor boxes is designed. Secondly, by designing an anchor spatial information interaction self-attention module and a temporal information fusion module, the spatiotemporal information fusion process of anchors is completed efficiently and accurately. Finally, an anchor box parameter quality assessment branch is established to improve the detection accuracy and model convergence speed significantly. The experimental results on the nuScenes validation set show that the proposed model achieves 41.3% and 52.5% on mAP and NDS, respectively, with a speed of 19.2 FPS, demonstrating high accuracy and real-time capability.

Key words: 3D object detection, surround-view camera, polar parametrization, autonomous vehicle