摘要
为提高驾驶环境中的车辆目标检测精度与测距稳定性,提出了一种基于深度学习的车辆目标检测与测距方法。以YOLOX-S算法为车辆目标检测框架进行改进:在原算法的基础上引入卷积块注意力模块,增强网络特征表达能力,并将置信度损失函数更换为Focal Loss,降低简单样本训练权重,提高正样本关注度。根据车载相机成像原理和几何关系建立车辆测距模型,并输入测距特征点坐标和相机内参得到测距结果。采用自制Tlab数据集和BDD100K数据集对改进的YOLOX-S算法进行训练与评价,搭建静态测距实验场景对车辆测距模型进行验证。实验结果表明:改进的YOLOX-S算法在实验数据集上检测速度为70.14帧/s,与原算法相比精确率、召回率、F_1值、mAP分别提高了0.86%、1.32%、1.09%、1.54%;在纵向50 m、横向11.25 m的测量范围内,平均测距误差保持在3.20%以内。可见,本文方法在满足车辆检测实时性要求的同时,具有良好的车辆测距准确性与稳定性。
In order to improve the vehicle target detection accuracy and ranging stability in the driving environment,a deep learning-based vehicle target detection and ranging method is proposed.The YOLOX-S algorithm is used as the vehicle target detection framework for improvement:on the basis of the original algorithm,the CBAM attention module is introduced to enhance the network feature expression ability,and the confidence loss function is replaced with Focal Loss,which reduces the training weight of simple samples and improves the positive sample attention.The vehicle ranging model is established according to the imaging principle and geometric relationship of the vehicle camera,and the ranging results are obtained by inputting the coordinates of the ranging feature points and the internal reference of the camera.The improved YOLOX-S algorithm is trained and evaluated using the home-made Tlab dataset and BDD 100K dataset,and the static ranging experimental scenario is constructed to validate the vehicle ranging model.The experimental results show that the improved YOLOX-S algorithm has a detection speed of 70.14 frames/s on the experimental dataset,and the precision,recall,F_1 value,and m AP are improved by 0.86%,1.32%,1.09%,and 1.54%respectively compared with the original algorithm;the average ranging error stays within 3.20%in the longitudinal 50m and lateral 11.25 m measurement range.It can be seen that the method in this paper has good vehicle ranging accuracy and stability while meeting the real-time requirements of vehicle detection.
作者
徐慧智
蒋时森
王秀青
陈爽
XU Hui-zhi;JIANG Shi-sen;WANG Xiu-qing;CHEN Shuang(School of Civil Engineering and Transportation,Northeast Forestry University,Harbin 150040,China)
出处
《吉林大学学报(工学版)》
北大核心
2025年第1期185-197,共13页
Journal of Jilin University:Engineering and Technology Edition
基金
国家自然科学基金青年科学基金项目(71701041)
黑龙江省自然科学基金项目(LH2019E007)。
关键词
交通运输规划与管理
深度学习
环境感知
车载图像
目标检测
单目测距
transportation planning and management
deep learning
environment awareness
in-vehicle imagery
target detection
monocular ranging
