改进YOLOv5s的轻量化车辆行人目标检测方法  

A LIGHTWEIGHT VEHICLE-PEDESTRIAN DETECTION METHOD BASED ON IMPROVED YOLOV5S

作　　者：李永潭 张晋玮 王俊杰 胡楠楠 赵曰峰[1] Li Yongtan;Zhang Jinwei;Wang Junjie;Hu Nannan;Zhao Yuefeng(School of Physics and Electronics,Shandong Normal University,250358,Jinan,China)

机构地区：[1]山东师范大学物理与电子科学学院,济南250358

出　　处：《山东师范大学学报(自然科学版)》2025年第1期71-80,共10页Journal of Shandong Normal University(Natural Science Edition)

基　　金：国家自然科学基金资助项目(42271093)。

摘　　要：针对当前车辆与行人目标检测网络设计复杂、计算资源消耗大的问题,探索高效的轻量化算法设计。首先以YOLOv5s模型为基础,改用轻量级网络Shufflenetv2作为主干结构,并用ReLU激活函数替换原函数,成功减少参数和计算量。同时,引入空间金字塔池化快速跨级部分连接模块,该模块通过金字塔池化结构保留不同尺度层次的特征,并通过跨阶段连接捕获底层与高层间的特征关系,以提升识别精度。然后在公开数据集KITTI上进行了大量对比测试及消融实验,结果显示改进后的模型识别精确率提升了2%,且参数数量、计算量、模型大小均显著减少。最后将改进后的算法在资源受限的移动设备嵌入式开发板RDK X3上进行实时推理验证,结果显示在RDK X3上帧率稳定接近设备的最大限制(30 FPS),以快速准确的识别车辆与行人。This paper delves into vehicle-pedestrian detection in the context of autonomous driving.Existing target-detection networks are often plagued by complex architectures and high demands for computational resources.To surmount these limitations,we present an enhancement to the YOLOv5s model.Specifically,we reengineer the YOLOv5s and integrate the lightweight Shufflenetv2network as the backbone and replace the activation function with ReLU.This redesign effectively reduces both the number of parameters in the model and its computational load,thus optimizing its performance.Moreover,we incorporate the Spatial Pyramid Pooling Faster Cross Stage Partial Channel(SPPFCSPC)structure into the improved YOLOv5s model.The SPPFCSPC module first acquires multi-scale feature information through pyramid pooling.Subsequently,it uses cross-stage connections to model the feature relationships between lower and higher layers,which significantly enhances the recognition accuracy.We assess the performance of the proposed model on the KITTI dataset.Experimental results show a 2%improvement in accuracy,along with substantial reductions in the number of parameters,computational complexity,and model size.Finally,the optimizedYOLOv5s model is deployed on the edge device RDK X3 for real-time inference validation.The results reveal that the frame rate on the RDK X3 remains stable,approaching the device's maximum capacity of 30 FPS.This enables rapid and precise vehicle and pedestrian detection.

关 键 词：YOLOv5s 目标检测 轻量化 Shufflenetv2 SPPFCSPC 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]