基于改进YOLOv7-Tiny的交通多目标检测方法  

Traffic Multi-target Detection Method Based on Improved YOLOv7-Tiny

作　　者：许文娟 李野[1] 江晟 王博文 XU Wenjuan;LI Ye;JIANG Sheng;WANG Bowen(School of Physics,Changchun University of Science and Technology,Changchun 130022)

机构地区：[1]长春理工大学物理学院,长春130022

出　　处：《长春理工大学学报（自然科学版）》2024年第2期75-83,共9页Journal of Changchun University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金(12274041,U2031113)。

摘　　要：在复杂的多目标交通环境中存在检测种类多、背景信息繁杂、图像分辨率低不能有效检测等问题,使用常见的目标检测算法不能达到高精度的实时检测效果,因此提出一种改进YOLOv7-Tiny的交通多目标检测算法。改进算法中首先使用部分卷积——PConv替换原始卷积,优化模型参数量和运行速度;其次采用轻量级算子CARAFE替换原有上采样部分的最临近插值,提升特征融合能力;最后采用EfficiCLoss替换原有损失函数,提高边界框的定位精度改善检测目标因遮挡而漏检问题。此外创建一个基于交通复杂场景的多目标数据集,在此数据集上进行实验,结果表明改进后的检测算法相较于原YOLOv7-Tiny网络的mAP提高了4.3%,检测速度提高了12.5%,参数量减少了30%,满足智慧交通实时检测的要求。In the complex multi-object traffic environment,there are challenges such as diverse detection categories,i ntricate background information,and ineffective detection due to low image resolution.Commonly used object detection algorithms fail to achieve high-precision real-time detection.To address these issues,we propose an improved algorithm for multiobject traffic detection based on YOLOv7-Tiny.In the enhanced algorithm,we first employ Partial Convolution(PConv)to replace the original convolution,optimize model parameters and runtime speed.Next,we integrate the lightweight operator,Context-Aware Reassembly Feature(CARAFE),to replace the previous nearest-neighbor interpolation in the upsampling section,enhancing feature fusion capabilities.Lastly,Efficient Classification Loss(EfficiCLoss)is introduced to replace the original loss function,i mproving the localization accuracy of bounding boxes and mitigating the issue of missed detections caused by occlusion.Additionally,we create a multi-object dataset based on complex traffic scenarios for experimentation.The results demonstrate that the enhanced detection algorithm achieves a 4.3%improvement in mean Average Precision(mAP)compared to the original YOLOv7-Tiny network.Furthermore,the detection speed increases by 12.5%,and the parameter count decreases by 30%,meeting the requirements for real-time detection in intelligent transportation systems.

关 键 词：交通目标检测 YOLOv7-Tiny Faster-Net EfficiCLoss 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]