汽车工程 ›› 2023, Vol. 45 ›› Issue (10): 1833-1844.doi: 10.19562/j.chinasae.qcgc.2023.10.006

所属专题: 智能网联汽车技术专题-感知&HMI&测评2023年

• • 上一篇    下一篇

面向路侧视角目标检测的轻量级YOLOv7-R算法

张小俊1(),奚敬哲1,2,史延雷2,袁安录2   

  1. 1.河北工业大学机械工程学院,天津  300401
    2.中国汽车技术研究中心汽车工程研究院,天津  300300
  • 收稿日期:2023-01-29 修回日期:2023-03-21 出版日期:2023-10-25 发布日期:2023-10-23
  • 通讯作者: 张小俊 E-mail:xjzhang@hebut.edu.cn
  • 基金资助:
    天津市科委新一代人工智能科技重大专项基金(18ZXZNGX00230)

Lightweight YOLOv7-R Algorithm for Road-Side View Target Detection

Xiaojun Zhang1(),Jingzhe Xi1,2,Yanlei Shi2,Anlu Yuan2   

  1. 1.School of Mechanical Engineering,Hebei University of Technology,Tianjin  300401
    2.Automotive Engineering Research Institute,China Automotive Technology and Research Center,Tianjin  300300
  • Received:2023-01-29 Revised:2023-03-21 Online:2023-10-25 Published:2023-10-23
  • Contact: Xiaojun Zhang E-mail:xjzhang@hebut.edu.cn

摘要:

针对V2X中的路侧感知单元在检测过程中,模型部署困难的问题、被测目标所呈现的多尺度问题及目标之间遮挡问题,提出了一种基于YOLOv7算法的轻量级检测算法YOLOv7-R。首先使用改进的EfficientNetv2-s重新构建YOLOv7的主干网络,减小模型参数量,提高模型的推理速度。其次,采用CA坐标注意力机制,保留精确的位置信息,加强模型对多尺度目标的检测性能;同时采用Focal-EIoU损失函数,提升算法精度。最后,在预处理阶段使用GridMask数据增强,提升算法对被遮挡目标的学习能力。实验结果表明:相较于基线算法YOLOv7,该算法在DAIR-V2X-I数据集上的map@0.5和map@0.5:0.95分别提高了3%与4.8%,检测速率达到了96.3 f/s,从而在满足轻量化要求的同时得到更优的检测精度,有效地实现了路侧单元对交通参与者的检测任务。

关键词: 深度学习, 路侧感知, YOLOv7, 轻量化, 注意力机制, 车路协同

Abstract:

A lightweight detection algorithm YOLOV7-R based on the YOLOv7 algorithm is proposed to solve the problems of model deployment difficulty, multi-scale problem of the measured target and occlusion problem between targets in the detection process of the road side sensing unit in V2X. Firstly, the backbone of YOLOv7 is rebuilt using the improved EfficientNetv2-s to reduce the model parameters and improve the model detection speed. Secondly, CA coordinate attention mechanism is adopted to retain accurate location information to enhance the performance of the model for multi-scale targets. At the same time, Focal-EIoU loss function is utilized to enhance the accuracy of the algorithm. Finally, GridMask image enhancement is used in the pre-processing stage to improve the learning ability of the algorithm for the blocked target. The experimental results show that compared with the baseline algorithm YOLOv7, the map@0.5 and map@0.5:0.95 value of the proposed algorithm on the DAIR-V2X-I dataset is increased by 3% and 4.8%, respectively, with the detection rate reaching 96.3 f/s, which can meet the requirements of lightweight and obtain better detection accuracy, and effectively implement the detection task of the road side unit for traffic participants.

Key words: deep learning, roadside perception, YOLOv7, lightweight, attention mechanism, IVICS