融合注意力机制的轻量危险驾驶行为检测模型  

A Lightweight Model for Dangerous Driving Behavior Detection Integrated with Attention Mechanism

作　　者：周叶凡 周正贵[1] ZHOU Yefan;ZHOU Zhenggui

机构地区：[1]安徽商贸职业技术学院信息与人工智能学院,安徽芜湖241002

出　　处：《芜湖职业技术学院学报》2025年第1期24-30,58,共8页Journal of Wuhu Institute of Technology

基　　金：2023年安徽省高校自然科学重点项目“基于深度学习技术的智慧交通机器视觉目标检测应用研究”(项目号:2023AH052296);2024年安徽省高校自然科学重点项目“基于物联网技术的智慧交通系统研究”(项目号:2024AH050528);2024年安徽省职业与成人教育学会教育教学研究规划课题“‘现代产业学院’背景下人工智能技术应用‘现场工程师’人才培养模式研究”(项目号:AZCJ2024129)。

摘　　要：危险驾驶行为(如使用手机、拿取食物、查阅文件等)会导致驾驶员分心,增加交通事故风险。车内复杂的环境以及光照不均的状况,使针对危险驾驶行为的检测工作面临挑战。因此,提出一种改进型YOLOv7-tiny模型——DBNet-YOLO,用于实时危险驾驶行为检测。该模型通过融合通道与空间特征的注意力机制,减少了背景干扰,引入的多尺度信息融合,也有助于增强在复杂光照条件下的特征提取能力;同时,采用结构化剪枝技术降低了模型的计算复杂度和参数量。在CVRR-HANDS 3D数据集上,DBNet-YOLO的mAP@0.75达57.1%,比原模型提升3.1%;参数量减至0.33MB;推理速度达330帧/秒。实验表明,该模型具备实时性强、精确度高和轻量化的优势,适合嵌入式平台的危险驾驶行为检测工作。Dangerous driving behavior(such as using a mobile phone,taking food or checking documents)distract drivers and increase the risk of traffic accidents.The complex in-vehicle environment and uneven lighting conditions pose significant challenges to the detection of dangerous driving behaviors.Hence,an improved YOLOv7-tiny model,named DBNet-YOLO,is proposed for real-time detection of dangerous driving behavior.This model reduces background noise through an attention mechanism incorporating the features of channel and space,and introduces multi-scale information fusion to enhance the model’s ability in extracting features under complex lighting conditions.Additionally,a structured pruning technique is employed to reduce the model’s computational complexity and parameter size.On the CVRR-HANDS 3D dataset,DBNet-YOLO achieves an mAP@0.75 of 57.1%,which increases by 3.1%over the original model;the parameter size is reduced to 0.33 MB,and its inference speed reaches 330 frames per second.Experimental results demonstrate that the model provides strong real-time performance,high precision,and lightweight characteristics,suitable for the embedded platform-based detection of dangerous driving behavior.

关 键 词：危险驾驶行为 YOLOv7-tiny 注意力机制 多尺度信息融合 结构化剪枝 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]