基于改进YOLOv5算法的车辆目标检测
Vehicle Target Detection Based on Improved YOLOv5 Algorithm
朱凯斌 1吕红明 1秦彦彬1
作者信息
- 1. 盐城工学院 汽车工程学院,盐城 224051
- 折叠
摘要
近年来,车辆目标检测已经逐渐成为智能交通领域的研究热点,然而,现实道路场景的复杂性常常给目标检测带来了诸多挑战,尤其是频繁出现的目标遮挡、重叠以及小目标检测问题,该文提出一种基于改进YOLOv5算法的车辆目标检测模型.首先将CA(coordinate attention,坐标注意力)机制集成到检测网络中,以增强在高密度目标环境中对车辆的识别精度;接着,采用Focal-EIOU Loss作为替代损失函数,以实现更精确的定位精度和更快的收敛速度;最后,引入ODConv(omni-dimensional dynamic convolution,全维度动态卷积),通过结合多个卷积核以及执行多维特征关注,提升车辆特征提取的效果.在自制车辆数据集上的实验表明,与原始算法相比,改进算法的平均精度提高了1.5%,证实了其在车辆目标检测方面的有效性和优越性.
Abstract
In recent years,vehicle detection has emerged as a pivotal area of investigation within intelligent transporta-tion systems.The complexities of real-world road conditions often pose considerable challenges to target detection ef-forts,particularly concerning the phenomena of target occlusion,overlap,and small targets.This paper introduces an enhanced vehicle detection model based on the YOLOv5 algorithm.Initially,the integration of the coordinate attention(CA)mechanism fortifies the detection network's capacity to accurately discern vehicle targets amidst densely popu-lated scenes.Subsequently,the implementation of Focal-EIOU Loss as an alternative loss function is delineated,con-tributing to refined bounding box localization and expedited loss function convergence.The culmination of this series of enhancements is the incorporation of omni-dimensional dynamic convolution(ODConv),an innovative approach that significantly refines feature extraction by harnessing the synergy of multiple convolutional kernels and multi-dimen-sional attention to vehicle characteristics.Experimentation on an self-made vehicle dataset elucidates that the pro-posed algorithmic enhancements result in a mean average precision(mAP)increment of 1.5%over the original algo-rithm,thereby proving its efficacy and superior performance in the domain of vehicle target detection.
关键词
深度学习/YOLO/注意力机制/损失函数/ODConvKey words
deep learning/YOLOv5/attention mechanism/loss function/omni-dimensional dynamic convolution(ODConv)引用本文复制引用
基金项目
国家自然科学基金面上项目(51875494)
江苏省研究生实践创新计划项目(SJCX_XY045)
出版年
2024
NETL
NSTL
维普
万方数据