文献类型：期刊文献

中文题名：基于改进YOLOv4的小目标行人检测算法

英文题名：Small-Target Pedestrian-Detection Algorithm Based on Improved YOLOv4

作者：王程[1,2];刘元盛[1,2];刘圣杰[1,2]

第一作者：王程

机构：[1]北京联合大学北京市信息服务工程重点实验室,北京100101;[2]北京联合大学机器人学院,北京100101

第一机构：北京联合大学北京市信息服务工程重点实验室

年份：2023

卷号：49

期号：2

起止页码：296-302

中文期刊名：计算机工程

外文期刊名：Computer Engineering

收录：CSTPCD;;北大核心:【北大核心2020】；CSCD:【CSCD_E2023_2024】；

基金：国家自然科学基金“无人车多视视频信息获取与定位关键技术”(61871038);国家自然科学基金“基于视觉计算的智能驾驶实时城市道路场景理解”(61871039);北京联合大学研究生科研创新项目(YZ2020K001);北京联合大学人才强校优选-拔尖计划“无人驾驶车复杂场景中可靠性定位技术研究”(BPHR2020BZ01)。

语种：中文

中文关键词：无人驾驶;小目标行人;深度可分离卷积;scSE注意力模块;特征金字塔网络

外文关键词：driverless vehicle;small-target pedestrian;deeply separable convolution;scSE attention module;Feature Pyramid Network(FPN)

摘要：行人检测在无人驾驶环境感知领域具有重要应用。现有行人检测算法多数只关注普通大小的行人目标,忽略了小目标行人特征信息过少的问题,从而造成检测精度低、应用于嵌入式设备中实时性不高等情况。针对该问题,提出一种小目标行人检测算法YOLOv4-DBF。引用深度可分离卷积代替YOLOv4算法中的传统卷积,以降低模型的参数量和计算量,提升检测速度和算法实时性。在YOLOv4骨干网络中的特征融合部分引入scSE注意力模块,对输入行人特征图的重要通道和空间特征进行增强,促使网络学习更有意义的特征信息。对YOLOv4颈部中特征金字塔网络的特征融合部分进行改进,在增加少量计算量的情况下增强对图像中行人目标的多尺度特征学习,从而提高检测精度。在VOC07+12+COCO数据集上进行训练和验证,结果表明,相比原YOLOv4算法,YOLOv4-DBF算法的AP值提高4.16个百分点,速度提升27%,将该算法加速部署在无人车中的TX2设备上进行实时测试,其检测速度达到23FPS,能够有效提高小目标行人检测的精度及实时性。
Pedestrian detection is vital to applications in unmanned environment perception.Most existing pedestriandetection algorithms focus only on ordinary pedestrian targets and do not consider the low accuracy caused by the insufficient pedestrian feature information of small targets;furthermore,they do not offer favorable real-time performance when applied to embedded devices.Hence,a small-target pedestrian-detection algorithm,YOLOv4-DBF,is proposed herein.The conventional convolution is replaced with deeply separable convolution in the YOLOv4 algorithm,which reduces the number of parameters and the computation time of the model,as well as improves the detection speed and real-time performance of the algorithm.Additionally,the concurrent spatial and channel Squeeze&Excitation(scSE)attention module is introduced into the feature fusion component of the YOLOv4 backbone network to enhance the important channels and spatial features of the input pedestrian feature map as well as to enable the network to learn more meaningful feature information.The feature fusion component of the Feature Pyramid Network(FPN)in the YOLOv4 neck is improved to enhance the multiscale feature learning of the pedestrian target in the image,which improves the detection accuracy but increases the amount of computation.After training and verification based on the VOC07+12+COCO dataset,the results show that compared with the original YOLOv4 algorithm,YOLOv4-DBF increases the Average Precision(AP)by 4.16 percentage points and the speed by 27%.Finally,YOLOv4-DBF is accelerate deployed on the TX2 equipment of an unmanned vehicle for real-time testing,where the maximum speed reaches 23FPS.The algorithm proposed herein can effectively improve the accuracy and real-time performance of small-target pedestrian detection.