基于自注意力机制的区域提案优化3D目标检测网络

打开文本图片集

中图分类号：TP399 文献标志码：A

3D Object Detection Network Based on Self-attention Mechanism for Regional Proposal Optimization

ZHANG Xin1 ，BI Bo-xue²，ZAN Guo-kuan³，ZHAO Jun-li1，WAN Zhi-bo1 （1. College of Computer Science 8. Technology，Qingdao University，Qingdao 266071，China; 2. Jimo District People's Hospital，Qingdao 266299，China; 3. Network Information Department， Chengyang District People's Hospital， Qingdao 266109，China）

Abstract： In the proposal refinement stage of the object detection algorithms there were two key problems including insufficient capture of contextual information and inadequate modeling of feature correlations. The improvement of detection accuracy was restricted by these problems. To address these challengs，a 3D object detection network for region proposal optimization， namely Proposal Refinement Optimization-RCNN （PRO-RCNN）， based on the self-attention mechanism was proposed. On the basis of the PV-RCNN is model，the Transformer model was utilized to dynamically learn feature weights，capture rich contextual information，and model the correlations between objects，thereby optimizing the proposal generation results. The experimental results on the KITTI dataset show that the accuracy of PRO-RCNN is improved in all test diffculty categories， the average precision of pedestrian category detection is increased by 2% to 3% ：

Keywords： object detection; deep learning; point cloud; self-attention; neural network

激光雷达点云在计算机视觉中至关重要，广泛应用于自动驾驶、增强现实和机器人等领域[1-]。（剩余9707字）