改进YOLOv4行人和车辆目标检测算法  

An improved YOLOv4 target detection algorithm for pedestrian and vehicle detection

作　　者：王玉莹 朱福珍[1] WANG Yuying;ZHU Fuzhen(College of Electronic Engineering,Heilongiang University,Harbin 150080,China)

机构地区：[1]黑龙江大学电子工程学院,哈尔滨150080

出　　处：《黑龙江大学自然科学学报》2023年第1期120-126,共7页Journal of Natural Science of Heilongjiang University

基　　金：国家自然科学基金(61601174);黑龙江省博士后科研启动金项目(LBH-Q17150);黑龙江省普通高等学校电子工程重点实验室(黑龙江大学)开放课题及省高校科技创新团队项目(2012TD007);黑龙江省省属高等学校基本科研业务费基础研究项目(KJCXZD201703);黑龙江省自然科学基金(F2018026)。

摘　　要：针对传统行人和车辆检测方法中小目标检测精度低、识别效果差以及遮挡的行人目标漏检问题,提出了一种基于YOLOv4改进的行人车辆检测算法。在主干网络与特征融合模块之间增加卷积层,3×3的卷积增大感受野,随后1×1的卷积降维。多层卷积学习到更多的纹理信息,提高了网络对特征的感知能力。为了解决YOLOv4样本不平衡的问题,利用焦点损失函数在解决正负样本分布不平衡问题上的优势,在分类损失函数中添加了一个调制系数λ,使得算法能够减少分类样本的损失。在Udacity数据集上的实验结果表明,相较于YOLOv4,改进的算法在主观视觉上能够检测出更多的目标,并且mAP提升了3.16%。An improved pedestrian and vehicle detection algorithm is proposed, based on YOLOv4 to address the issues of low accuracy of small target detection, poor recognition effect, and missed detection of occluded pedestrian targets in the conventional pedestrian and vehicle detection approaches. Convolutional layers are added between the backbone network and the feature fusion module, and the 3×3 convolution increases the perceptual field, followed by the 1×1 convolution to reduce the dimensionality. The multi-layer convolution learns more texture information and improves the network’s ability to detect features. To solve the YOLOv4 sample imbalance problem, the advantage of the focal loss function in solving the positive and negative sample distribution imbalance problem is exploited and adds a modulation coefficient λ to the Classification Loss function so that the proposed algorithm can reduce the loss of classification samples. Testing results on the Udacity dataset demonstrate that, in comparison to YOLOv4, the enhanced algorithm can detect more targets in subjective vision and improve mAP by 3.16 %.

关 键 词：目标检测 YOLOv4 CSPDarknet53 交叉熵损失 

分 类 号：TP344.1[自动化与计算机技术—计算机系统结构]