《计算机应用》唯一官方网站 ›› 0, Vol. ›› Issue (): 240-245.DOI: 10.11772/j.issn.1001-9081.2024030272

• 多媒体计算与计算机仿真 • 上一篇    下一篇

基于改进YOLOv5s的车辆-人员检测方法

王凯1,2,3,4, 吴清潇1,2,3,4(), 关键1,2,3,4, 方英健1,2,3,4, 李思聪1,2,3,4   

  1. 1.中国科学院光电信息处理重点实验室,沈阳 110016
    2.中国科学院 沈阳自动化研究所,沈阳 110016
    3.中国科学院机器人与智能制造创新研究院,沈阳 110169
    4.中国科学院大学,北京 100049
  • 收稿日期:2024-03-15 修回日期:2024-04-14 接受日期:2024-04-16 发布日期:2025-01-24 出版日期:2024-12-31
  • 通讯作者: 吴清潇
  • 作者简介:王凯(2000—),男,河北唐山人,硕士研究生,主要研究方向:模式识别、智能系统、机器人视觉、深度学习、目标检测
    吴清潇(1978—),男,山东滨州人,研究员,博士,主要研究方向:机器人视觉
    关键(2000—),女,河南平顶山人,硕士研究生,主要研究方向:模式识别、智能系统、深度学习、目标检测
    方英健(1998—),男,河北承德人,硕士研究生,主要研究方向:模式识别、智能系统、深度学习、目标检测
    李思聪(1987—),男,山东蓬莱人,副研究员,硕士,主要研究方向:机器人视觉、模式识别、智能系统。
  • 基金资助:
    国家重点研发计划项目(2021YFC3002002)

Vehicle-passenger detection method based on enhanced YOLOv5s

Kai WANG1,2,3,4, Qingxiao WU1,2,3,4(), Jian GUAN1,2,3,4, Yingjian FANG1,2,3,4, Sicong LI1,2,3,4   

  1. 1.Key Laboratory of Opto-Electronic Information Processing,Chinese Academy of Sciences,Shenyang Liaoning 110016,China
    2.Shenyang Institute of Automation,Chinese Academy of Sciences,Shenyang Liaoning 110016,China
    3.Institutes for Robotics and Intelligent Manufacturing,Chinese Academy of Sciences,Shenyang Liaoning 110169,China
    4.University of Chinese Academy of Sciences,Beijing 100049,China
  • Received:2024-03-15 Revised:2024-04-14 Accepted:2024-04-16 Online:2025-01-24 Published:2024-12-31
  • Contact: Qingxiao WU

摘要:

在交通事故等紧急情形下,及时定位车内被困人员对救援工作至关重要。针对救援中车内人员检测的难题,提出一种基于YOLOv5s的检测模型,以实现对车与车内人员的两阶段检测。在第一阶段,检测出车辆位置,并通过动态缩放聚焦感兴趣区域(ROI);然后采用直方图均衡化增强过暗图像的对比度,并使用非局部均值滤波去除噪声,从而提升图像的质量;在第二阶段,准确判定车内是否有被困人员,并定位人员的确切位置。在BIT-Vehicle数据集和UA-DETRAC数据集上的实验结果表明,相较于Faster-RCNN、YOLOv7-tiny等模型,所提模型的精度、召回率等多个指标都表现最佳,展示出更强的鲁棒性和更高的准确率。此外,所提模型的实时性能可以满足智能救援场景的需求。

关键词: 图像处理, 图像增强, 目标检测, 深度学习, YOLOv5, 车内被困人员定位

Abstract:

In emergency situations such as traffic accidents, the timely location of trapped passengers inside vehicle is critical for rescue operations. For the challenge of detecting passengers in vehicles during rescue, a detection model based on YOLOv5s was proposed to achieve two-stage detection of the vehicle and passengers inside. In the first stage, the vehicle’s location was determined, followed by focusing on the Regions Of Interest (ROIs) through dynamic scaling. After that, histogram equalization was utilized to enhance the contrast of underexposed images, and non-local means filtering was employed to remove noise, thereby improving the image quality. In the second stage of the model, the existence of trapped passengers inside the vehicle was determined accurately, and the exact locations of the passengers were identified. Experimental results on BIT-Vehicle and UA-DETRAC datasets show that compared to models such as Faster-RCNN and YOLOv7-tiny, the proposed model has the best performance in terms of precision, recall, and other metrics, demonstrating stronger robustness and higher accuracy. Besides, the real-time performance of the proposed model can meet the needs of intelligent rescue scenarios.

Key words: image processing, image enhancement, object detection, deep learning, YOLOv5s, location of trapped passengers inside vehicle

中图分类号: