基于改进YOLOv8的行人和车辆检测算法研究  

Research on an improved pedestrian and vehicle detection algorithm based on YOLOv8

作　　者：刘佳琳 殷丽凤[1] LIU Jia-lin;YIN Li-feng(School of Software,Dalian Jiaotong University,Dalian 116028,China)

机构地区：[1]大连交通大学软件学院,辽宁大连116028

出　　处：《云南民族大学学报(自然科学版)》2025年第2期197-205,共9页Journal of Yunnan Minzu University:Natural Sciences Edition

基　　金：国家自然科学基金(61771087).

摘　　要：针对目前主流的目标检测算法在复杂的交通环境下对行人和车辆检测精度不高的问题,提出一种基于YOLOv8模型改进的目标检测算法.首先,在主干网络前增加一种可学习增强网络,该网络是通过拉普拉斯分解残差学习方式来构建的,不仅可以充分提取到目标的特征信息,还能增强算法的准确性和鲁棒性,从而减少不同强度的光照对图像目标检测的干扰;其次,在中等目标检测层之前添加提出的KSA注意力机制,检测时可以将注意力集中在图像中重要信息的区域,从而更准确地定位并识别出中等目标,同时也可以减少复杂的背景噪声对检测的干扰;最后,提出了基于像素点的Transformer结构,即Pixel Transformer结构.将该结构添加到主干网络中,从而进一步增强算法对图像全局特征的提取能力,使其能学习到更丰富、更全面的目标特征.使用KITTI数据集来进行消融实验以及各算法的对比实验,实验结果表明,设计的算法在相关指标上取得了一定程度的提升,其中mAP@0.5值提升了3.2个百分点,达到了96.7%,这充分体现了该算法的优越性.In view of the problem that the mainstream target detection algorithms currently have low detection accuracy for pedestrians and vehicles in complex traffic environments,this paper proposes an improved target detection algorithm based on the YOLOv8 model.Firstly,a learnable enhancement network is added in front of the backbone network.This network is constructed through Laplacian decomposition residual learning method,which not only allows for full extraction of target feature information,but also enhances the accuracy and robustness of the algorithm,thus reducing the interference of different intensities of illumination on image target detection.Secondly,the proposed KSA attention mechanism is introduced before the medium target detection layer in this paper.After adding this attention mechanism,during detection,the attention can focus on the areas of important information in the image,so as to more accurately locate and identify medium targets,and at the same time,it can also reduce the interference of complex background noise on detection.Finally,this paper proposes a transformer structure based on pixel points,namely the Pixel Transformer structure.By adding this structure to the backbone network,the ability of the algorithm to extract global features of the image is further enhanced,enabling it to learn richer and more comprehensive target features.This paper uses the KITTI dataset to conduct ablation and comparison experiments of various algorithms.The experimental results show that the algorithm designed in this paper has achieved a certain degree of improvements in relevant indicators.Among them,the mAP@0.5 value has increased by 3.2 percentage points and reached 96.7%,which fully reflects the superiority of the algorithm.

关 键 词：目标检测 YOLOv8 注意力机制 全局特征 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]