融合图像显著性特征的轻量级目标检测算法  

Lightweight Object Detection Algorithm Based on Image Saliency Feature Fusion

作　　者：马雷[1] 杨顺清 王欢欢 翟家琛 徐健傲 Ma Lei;Yang Shunqing;Wang Huanhuan;Zhai Jiachen;Xu Jianao(College of Vehicle and Energy,Yanshan University,Qinhuangdao 066004)

机构地区：[1]燕山大学车辆与能源学院,秦皇岛066004

出　　处：《汽车工程》2024年第1期84-91,共8页Automotive Engineering

基　　金：国家重点研发计划项目(2021YFB3202204)资助。

摘　　要：针对智能车辆在实际交通环境中面临的目标密集、边缘严重遮挡和前景背景模糊的问题,本文提出了一种融合图像显著性特征的轻量级目标检测算法。首先基于灰度图像提取出显著性特征图,和彩色图像分别输入卷积神经网络。其次采用轻量化模块(ghost model)搭建轻量级融合网络,并使用EIoU优化模型的边框定位损失。在网络后端将非极大值抑制算法进行改进,以此提高网络对同类别遮挡目标的检测准确率。最后在KITTI数据集上进行训练和测试。实验表明,改进后的网络mAP达到92.7%,相比原始网络YOLOv5平均精度提高3.8%,精确率和召回率分别提高3%和6.2%。For the problems of dense targets,severe edge occlusion,and blurred foreground and back⁃ground that intelligent vehicles face in actual traffic environments,a lightweight object detection algorithm based on image saliency feature fusion is proposed in this paper.Firstly,salient feature maps are extracted based on gray⁃scale images,and input into convolutional neural networks with color images.Secondly,a lightweight fusion net⁃work is constructed using the Ghost Model,and the EIoU is used to optimize the model's border localization loss.In order to enhance the detection accuracy of similar occluded targets,non-maximum suppression algorithm is im⁃proved on the backend of the network.Finally,the KITTI dataset is used for training and testing.The experiment shows that the improved detection mAP value of the network reaches 92.7%,with an average accuracy improvement of 3.8%compared to the original network YOLOv5.The accuracy and recall rates are increased by 3%and 6.2%.

关 键 词：目标检测 多特征融合 轻量级网络 YOLOv5 

分 类 号：U495[交通运输工程—交通运输规划与管理] TP391.41[交通运输工程—道路与铁道工程]