Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (17): 195-204.DOI: 10.3778/j.issn.1002-8331.2205-0023

• Graphics and Image Processing • Previous Articles     Next Articles

KPP3D:Key Point Painting for 3D Object Detection

WANG Mingming, CHEN Qingkui, FU Zhibing   

  1. 1.School of Management, University of Shanghai for Science and Technology, Shanghai 200093, China
    2.School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China
  • Online:2023-09-01 Published:2023-09-01

KPP3D:基于关键点信息融合的3D目标检测模型

汪明明,陈庆奎,付直兵   

  1. 1.上海理工大学 管理学院,上海 200093
    2.上海理工大学 光电信息与计算机工程学院,上海 200093

Abstract: In order to fuse image information to improve the accuracy of the 3D object detection model using lidar sensor and reduce runtime while improving detection accuracy, a 3D target detection model KPP3D based on key point information fusion is proposed. KPP3D uses a lightweight key point generation module to replace the non-derivable sampling process, so that the entire model can be trained end-to-end to generate task-related key points. Aiming at the problem of high sampling repetition rate caused by the introduction of this module, a sampling loss function based on the evolution of Bernoulli distribution entropy is proposed to reduce the repetition rate and control the attention distribution of key points. Experiments in the autonomous driving dataset KITTI show that the fusion of image and point cloud can improve the accuracy of 3D object detection, and the use of the KPP3D model with 512 key points can reduce the time consumption by 38%. In addition, experiments verify that the newly proposed loss function can make the key point non-repetitive rate reach more than 98%, thereby reducing the redundancy of key points.

Key words: sensor fusion, autonomous driving, 3D object detection, lidar, sampling

摘要: 为了融合图像信息以提高单激光雷达传感器模型的3D目标检测准确率,使融合感知模型在提升检测精度的同时减少运行时间,提出了基于关键点信息融合的3D目标检测模型KPP3D。KPP3D使用轻量级关键点生成模块取代不可求导的采样过程,使整个模型可以进行端到端的训练从而生成任务相关的关键点。针对引入该模块产生的高采样重复率问题,提出了一种基于伯努利分布熵演变来的采样损失函数降低重复率并聚焦关键点的注意力分布。在自动驾驶数据集KITTI的实验表明,融合图像信息到点云能提升3D目标检测准确率且使用512个关键点的KPP3D模型能减少38%时间消耗。此外,实验验证了新提出的损失函数能使关键点的不重复率达到98%以上,从而减少了关键点的冗余。

关键词: 传感器融合, 自动驾驶, 3D目标检测, 激光雷达, 采样