单目路侧视角下的行人三维形态检测算法  

Pedestrian 3D Form Detection Algorithm from Monocular Roadside View

作　　者：王伟[1] 张旭[1] 刘昌法 徐悦 WANG Wei;ZHANG Xu;LIU Changfa;XU Yue(School of Information Engineering,Chang’an University,Xi’an 710064,Shaanxi;Xi’an Soar Electromechanical Technology Co.,Ltd.,Xi’an 710005,Shaanxi)

机构地区：[1]长安大学信息工程学院,陕西西安710064 [2]西安翱翔机电科技有限公司,陕西西安710005

出　　处：《火箭军工程大学学报》2024年第6期31-43,共13页Journal of Rocket Force University of Engineering

基　　金：陕西省自然科学基础研究计划项目(2023-JC-YB-600);长安大学创新训练计划项目(S202410710236)。

摘　　要：针对现有智能监控系统存在检测精度差、标定速度慢等问题,以三维目标检测网络SMOKE为框架,提出了一种基于单目路侧视角下的行人三维形态检测算法(MonoNet),以实现更精准、更隐蔽、更高效的行人无感三维形态建视。首先,利用道路场景自动标定算法构建二维到三维场景的映射矩阵,然后将已知的行人二维标注信息转换为行人三维圆柱型标注信息,构造行人检测三维数据集;其次,通过改进的骨干特征提取网络和特征增强模块进一步提取目标特征,获取更加丰富的浅层和深层特征信息;最后,添加新的损失函数用以提高模型训练的收敛性和检测精度。实验结果表明:改进后的行人三维检测算法的平均精度相比于现有的检测算法提高了9.21%,与目前大部分三维目标检测算法相比,该算法取得了更高的检测精度和较好的检测速度,能够有效提升智能监控系统在安全监测、异常行为预警等方面的性能。To explore the problems such as poor detection precision and low calibration speed in current intelligent monitoring systems,a pedestrian 3D form detection algorithm from monocu-lar roadside view was proposed with the 3D object detection network SMOKE as the framework.Thereby,a more precise,more concealed and more efficient pedestrian unconscious 3D form vi-sualization could be realized.Firstly,road scene automatic calibration algorithms were used to construct a mapping matrix from 2D to 3D scenes,by which the known 2D pedestrian annotation information was converted into 3D cylindrical annotation information to construct a 3D pedestrian detection dataset.Secondly,through the improved backbone feature extraction network and fea-ture enhancement module,target features were further extracted to obtain more shallow and deep feature information.Finally,a novel loss function was added to improve the convergence and de-tection precision of model training.Experimental results showed that the average precision(AP)of the improved pedestrian 3D detection algorithm increased by 9.21%compared with the current detection algorithms.And compared with most current 3D object detection algorithms,the pro-posed algorithm achieved higher detection precision and speed,which can effectively improve the performance of intelligent monitoring systems in safety monitoring,abnormal behavior warning and other aspects.

关 键 词：计算机视觉 行人三维形态检测 路侧视角 特征增强模块 圆柱型标注 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]