主办：陕西省汽车工程学会
ISSN 1671-7988  CN 61-1394/TH
创刊：1976年
汽车实用技术

汽车实用技术 ›› 2024, Vol. 49 ›› Issue (16): 40-43.DOI: 10.16638/j.cnki.1671-7988.2024.016.007

• 智能网联汽车 • 上一篇    

基于改进 YOLOv4-tiny 的行人检测算法研究

王 京，高浩宁   

  1. 河北建筑工程学院 能源工程系
  • 发布日期:2024-08-23
  • 通讯作者: 王 京
  • 作者简介:王京（1997－），男，硕士研究生，研究方向为机器视觉，E-mail:wangjing66322@163.com。

Research on the Pedestrian Detection Algorithm for Improving YOLOv4-tiny

WANG Jing, GAO Haoning   

  1. Department of Energy Engineering, Hebei University of Architecture
  • Published:2024-08-23
  • Contact: WANG Jing

PDF

摘要： 在汽车智能化进程中，对于道路行人的检测研究是必不可少的，文章基于 YOLOv4- tiny 提出一种改进的行人检测算法，应用于车载小型摄像头。将空间金字塔池化结构（SPP）引入 网络结构，通过 SPP 模块实现局部特征和全局特征的融合，丰富最终特征图的表达能力；在 特征层和上采样引入了坐标注意力（CA）机制，从通道和空间两方面对图像特征进行有效关 注；实验采用 PASCALVOC-2007 数据集进行训练和验证。实验结果表明，改进后的算法在 VOC 数据集中，平均精度提高了 3.84%，F1 值为 0.80，查准率提高了 0.77%，查全率为 73.95%， 平均准确率均值（mAP）提高了 8.79%，在保证算法速率的同时提高了检测精度。该研究为 汽车智能化行驶过程中的行人检测提供了建议。

关键词: 深度学习；注意力机制；智能驾驶；行人检测；YOLOv4-tiny

Abstract: In the process of automobile intelligence, the research of road pedestrian detection is essential. This paper proposes an improved pedestrian detection algorithm based on YOLOv4-tiny, which is applied to the small vehicle camera. The spatial pyramid pooling structure (SPP) is introduced into the network structure, and the integration of local and global features is realized through SPP module to enrich the expression ability of the final feature map; the coordinate attention (CA) mechanism is introduced into the feature layer and upper sampling to pay attention to the image features from the channel and space; the experiment uses PASCALVOC-2007 data set for training and verification. The experimental results show that the average accuracy of the improved algorithm is improved by 3.84%, the F1 value is 0.80, the accuracy by 0.77%, recall by 73.95%, and mean average precision (mAP) by 8.79%, which improves the detection accuracy while ensuring the algorithm rate. This study provides suggestions for pedestrian detection in the process of automobile intelligence driving.

Key words: Deep learning; Attention mechanism; Intelligent driving; Pedestrian detection; YOLOv4- tiny