北京大学学报（自然科学版） ›› 2025, Vol. 61 ›› Issue (4): 746-754.DOI: 10.13209/j.0479-8023.2025.055

上一篇    下一篇

基于单发多框机制的路端目标检测算法研究

许卓栋, 兰逸舟, 尚可, 万泽宇, 张飞舟†   

  1. 北京大学地球与空间科学学院, 北京 100871
  • 收稿日期:2024-05-13 修回日期:2024-12-22 出版日期:2025-07-20 发布日期:2025-07-20
  • 通讯作者: 张飞舟, E-mail: zhangfz(at)pku.edu.cn

Roadside Object Detection Algorithm Based on Single Shot Multibox Mechanism

XU Zhuodong, LAN Yizhou, SHANG Ke, WAN Zeyu, ZHANG Feizhou†   

  1. School of Earth and Space Sciences, Peking University, Beijing 100871
  • Received:2024-05-13 Revised:2024-12-22 Online:2025-07-20 Published:2025-07-20
  • Contact: ZHANG Feizhou, E-mail: zhangfz(at)pku.edu.cn

PDF

摘要：

针对路端图像背景高结构相似性及低特征重要性的特点, 基于单发多框检测器(SSD)的运行机制, 通过引入数据预处理和轻量级卷积注意力模块, 并调整注意力模块在算法中的位置, 建立最优化的路端检测算法。在车路协同公开数据集DAIR-V2X上进行日间和夜间场景以及不同交通流量下的路端目标检测, 结果表明, 在提取图像特征的第3级池化层后添加注意力模块, 优化算法的精度指标mAP@0.5:0.95可以获得1.67%的提升, 且仅损失2 FPS的检测速度, 能够满足路端目标检测任务的精度与速度需求, 有效地增强车端的目标识别能力。

关键词: 自动驾驶汽车, 目标检测, 注意力模块, 车路协同技术

Abstract:

To address the high structural similarity and low feature importance of the roadside image background, the object detection algorithm is improved based on the mechanism of Single Shot Multibox Detector (SSD). By introducing data preprocessing module and lightweight convolutional attention module, and adjusting the position of the attention module in the algorithm, an optimized roadside detection algorithm is established. The roadside object detection task is performed for daytime, nighttime scenarios and different levels of traffic flow, and the experimental results on the vehicle-road collaborative public dataset DAIR-V2X show that adding the attention module after the third pooling layer where the optimized algorithm extracts image features results in a 1.67% improvement in accuracy metrics mAP@0.5:0.95 with only a 2 FPS loss in detection speed, which can meet the speed and accuracy requirements of roadside object detection tasks and enhance the perception capability of the vehicle side.

Key words: self-driving cars, object detection, attention module, vehicle-road collaborative technique

版权所有 © 《北京大学学报（自然科学版）》编辑部
技术支持：北京玛格泰克科技发展有限公司