An Efficient Point Cloud Correlation Enhancement RCNN for 3D Object Detection

Authors

  • Jialong Du Guangxi University
  • Hanzhang Huang
  • Qingji Tan
  • Yong Li
  • Lu Ding
  • Feng Shuang

DOI:

https://doi.org/10.5755/j01.itc.54.1.35616

Keywords:

3-D Object Detection, Lightweight Proposal, Self-Attention, Point Cloud, Autonomous Driving

Abstract

To meet the requirement of 3D object detection task , an efficient point clouds correlation enhancement RCNN(EPCE-RCNN) is proposed. The proposed method reduces the computational complexity and time consumption of the network through a lightweight proposal generation module, and accelerates the generation of the 3D proposal box. Meanwhile, during region of interest feature coding, the relevance among different grid points is enhanced through an efficient self-attention pooling module, so that the limitation that the pooling operation is influenced by the radius of a neighborhood query sphere is addressed. In addition, the combination of an attention mechanism and a feedforward network ensures the nonlinearity of the model, so that the model can perform feature expression better. Thus, the synchronous improvement of the network detection efficiency and the detection precision is realized. On the KITTI dataset, the detection accuracy of three difficulty levels reaches 89.99%, 81.69% and 77.17% respectively. Compared with the baseline Voxel-RCNN, the detection efficiency of EPCE-RCNN is improved by 12%. To verify the generalization and application value of the proposed method, a power equipment dataset with 3D label information is constructed, the 3D label frame information of the YCB dataset is also supplemented. Experiments are carried out on these datasets. In the experimental results of the verification set, the mAP of a mug, gelatin box, single clip, wedge clip and C clip can reach 37.67%, 40.06%, 35.63%, 30.01% and 37.31% respectively. Compared with the baseline, the proposed algorithm has a significant improvement and its generalization has been fully verified.

Downloads

Published

2025-04-01

Issue

Section

Articles