基于改进YOLOv4的多目标车辆检测算法
Multi-object vehicle detection algorithm based on improved YOLOv4
江屾 1殷时蓉 1罗天洪 2郑讯佳 3张洪杰1
作者信息
- 1. 重庆交通大学机电与车辆工程学院,重庆 400074
- 2. 重庆文理学院智能制造工程学院,重庆 402160
- 3. 重庆文理学院智能制造工程学院,重庆 402160;中国汽车工程研究院股份有限公司,重庆 401122
- 折叠
摘要
针对现有检测方法存在小目标车辆漏检率高以及夜间车辆误检率高的问题,提出一种基于改进YOLOv4的多 目标检测算法.引入深度可分离卷积代替标准卷积,减少模型的参数量与计算量.在保留YOLOv4输出层的同时,增加一层网格为104×104的输出层,提升算法对小目标车辆的检测性能.在Head部分引入Inceptionv3结构,采用K-means++聚类算法重新确定锚框,进一步提高算法对小目标车辆的检测性能.实验结果表明,算法相比改进前,在不降低检测速度的同时,其mAP增加2.44％,模型大小减少1/3,具有良好的鲁棒性.
Abstract
Aiming at the problems of high missing detection rate of small target vehicles and high false detection rate of vehicles at night in existing detection methods,a multi-object detection algorithm based on improved YOLOv4 was proposed.The depth separable convolution was introduced to replace the standard convolution,which reduced the amount of parameters and calcula-tion of the model.While retaining the YOLOv4 output layer,an output layer with grid of 104 × 104 was added to improve the detection performance of the algorithm for small target vehicles.The Inceptionv3 structure was introduced in the Head part,and the anchor frame was redefined using K-means++ clustering algorithm,which further improved the detection performance of the algorithm for small target vehicles.The results show that the improved algorithm has good robustness,while its mAP increases by 2.44％and the model size decreases by one third without reducing the detection speed.
关键词
车辆检测/深度学习/YOLOv4/深度可分离卷积/Inceptionv3/K-means++/多目标识别Key words
vehicle detection/deep learning/YOLOv4/depth separable convolution/Inceptionv3/K-means++/multiple object detection引用本文复制引用
基金项目
国家自然科学基金青年科学基金项目(52102454)
中国博士后科学基金面上基金项目(2021M700169)
重庆市自然科学基金面上基金项目(cstc2021jcyjmsxmX0395)
重庆市博士后研究项目特别基金项目(2021XM3069)
出版年
2024
NETL
NSTL
万方数据