轻量级行人检测算法研究  

Research on Lightweight Pedestrian Detection Algorithm

作　　者：苏喆 徐成也 吴林煌[1] SU Zhe;XU Chengye;WU Linhuang(School of Physics and Information Engineering,Fuzhou University,Fuzhou 350108,China)

机构地区：[1]福州大学物理与信息工程学院,福建福州350108

出　　处：《电视技术》2025年第2期30-36,共7页Video Engineering

摘　　要：针对当前基于深度学习的行人检测算法检测精度低、模型复杂、对设备要求较高的问题,提出一种基于YOLOv8的轻量级行人检测模型YOLOv8-PGL。首先,将卷积神经网络(Convolutional Neural Networks,CNN)和Transformer的思想结合,设计出C2f_PTB模块,增强特征提取能力,降低计算量和参数量;其次,引入改进的BiFPN网络替换原模型中的特征提取网络,提高对不同尺度特征的融合效率;再次,采用一种轻量级非对称检测头LADH,以微小性能损失大幅度减少模型的计算量和参数量;最后,使用PIoU(Powerful-IoU)替换CIoU(Complete-IoU)作为损失函数,以更准确地优化模型的预测结果,进一步提高模型的检测精度。实验结果表明,所提出的模型在多个指标上有明显的提升。相较于基准模型,YOLOv8-PGL的mAP50%提升1.9个百分点,参数量降低50%,计算量降低33%,模型大小降低48%。Aiming at the problems of low detection accuracy,complex model and high equipment requirements of current pedestrian detection algorithms based on deep learning,a lightweight pedestrian detection model YOLOV8-PGL based on YOLOv8 is proposed in this paper.Firstly,the Convolutional Neural Networks(CNN)and Transformer are combined to design the C2f_PTB module,which can enhance the feature extraction ability and reduce the calculation and parameter number.Secondly,the improved BiFPN network is introduced to replace the feature extraction network in the original model to improve the fusion efficiency of features of different scales.Thirdly,LADH,a lightweight asymmetric detection head,is used to greatly reduce the calculation and parameter quantity of the model with small performance loss.Finally,PloU(Powerful-loU)is used to replace CloU(Complete-IoU)as a loss function to optimize the prediction results of the model more accurately and further improve the detection accuracy of the model.The experimental results show that the proposed model has obvious improvement on several indexes.Compared with the benchmark model,the mAP50%of YOLOv8-PGL is increased by 1.9 percentage points,the number of parameters is reduced by 50%,the calculation amount is reduced by 33%,and the model size is reduced by 48%.

关 键 词：深度学习 YOLOv8 轻量化 行人检测 

分 类 号：TP311.1[自动化与计算机技术—计算机软件与理论]