复杂交通环境下的交通参与者红外目标轻量化检测模型  

Lightweight detection model for infrared targets of traffic participants in complex environments

作　　者：周军超 陈鑫 高建杰[3] 唐永清 ZHOU Junchao;CHEN Xin;GAO Jianjie;TANG Yongqing(School of Mechanical Engineering,Sichuan University of Science&Engineering,Zigong 643000,Sichuan,China;Chengdu-Chong qing Economic Circle(Luzhou)Advanced Technology Research Institute,Luzhou 646000,Sichuan,China;Intelligent Policing Key Laboratory of Sichuan Province,Sichuan Police College,Luzhou 646000,Sichuan,China)

机构地区：[1]四川轻化工大学机械工程学院,四川自贡643000 [2]成渝地区双城经济圈(泸州)先进技术研究院,四川泸州646000 [3]四川警察学院智能警务四川省重点实验室,四川泸州646000

出　　处：《安全与环境学报》2025年第4期1400-1411,共12页Journal of Safety and Environment

基　　金：四川省科技计划项目(2024ZHCG0063);泸州市科技计划项目(2023JYJ066)。

摘　　要：针对交通监测系统和智能网联汽车在多类别交通参与者红外目标检测中存在的识别精度低、实时性差和部署性难的问题,提出一种改进YOLOv9m的CNDS-YOLO交通参与者红外目标轻量化检测模型。首先,设计轻量级CE-MobileNetv3结构以替换整个主干网络,旨在优化参数量与检测精度之间的平衡;其次,引入深度可分离卷积(Depthwise Separable Convolution,DSConv)替换特征融合层、辅助可逆分支CBL模块和DS-SPPF模块中的常规卷积,旨在适配实时检测和轻量化部署需求;然后,在颈部融入经归一化改进的N-Swin Transformer模块,用于增强网络对红外目标特征的适应性和表达能力;最后,在检测端开发出尺寸为160×160的小目标检测头,以满足小目标和遮挡目标的检测敏感度需求。结果表明:与基线模型YOLOv9m相比,CNDS-YOLO模型在参数量和浮点计算数上分别降低了31.0%和35.5%,检测速度和平均检测精度分别提升了24.2%和7.0%;与主流模型相比,CNDS-YOLO模型在实时检测、精度和轻量化部署性等方面明显更优。To tackle the challenges of low recognition accuracy,subpar real-time performance,and difficulties in lightweight deployment for multi-class infrared target detection of traffic participants In Intelligent Connected Vehicles(ICVs)and traffic monitoring systems,we propose a lightweight detection model called CNDS YOLO,which is based on an enhanced YOLOv9m architecture.First,we designed a lightweight CE MobileNetv3 structure that incorporates the Generalized Efficient Layer Aggregation Network(C ELAN)to replace the entire backbone network.This approach significantly reduces the parameter count while enhancing detection accuracy.Second,Depthwise Separable Convolution(DSConv)was integrated to replace conventional convolution in the neck feature fusion layer,auxiliary reversible CBL module,and DS SPPF module.This modification further improves real-time performance and enables lightweight deployment capabilities.Additionally,a normalized N Swin Transformer module was incorporated into the neck feature fusion layer,significantly enhancing the model s ability to capture and fuse features related to infrared targets,which in turn improves detection accuracy and stability.Furthermore,a new 160×160 small target detection head was developed at the detection end to increase sensitivity to small and occluded targets.Finally,experiments conducted on the FLIR2 dataset demonstrated that the CNDS YOLO model outperformed the baseline YOLOv9m model,reducing parameters and FLOPs by 31.0%and 35.5%,respectively,while enhancing detection speed(F S)and mean Average Precision(AP)by 24.2%and 7.0%,respectively.Compared to mainstream models,the CNDS YOLO model achieved the highest values in both Average Precision(AP)and detection speed(F S).In the ablation experiments,each of the four improved modules within CNDS YOLO was quantitatively assessed.For example,the DSConv module resulted in a reduction of parameters and FLOPs by 24.5%and 37.6%,respectively,while the N Swin Transformer module improved Recall(R)and AP by 4.1%and 4.3%,respectively

关 键 词：安全工程 交通参与者 红外目标 CNDS-YOLO模型 轻量化 小目标检测 

分 类 号：X924[环境科学与工程—安全科学]