基于影像视锥的二维三维一体雷视融合车速测量  

Integrated 2D-3D LiDAR-vision fusion vehicle speed estimation based on image frustum

作　　者：周奎宇 黄玉春[1] 杨鹤 李娜[3] ZHOU Kuiyu;HUANG Yuchun;YANG He;LI Na(School of Remote Sensing and Information Engineering,Wuhan University,Wuhan 430079,China;Transportation Development Center of Henan Province,Zhengzhou 450000,China;Troops 61175 of PLA,Nanjing 210049,China)

机构地区：[1]武汉大学遥感信息工程学院,湖北武汉430079 [2]河南省交通事业发展中心,河南郑州450000 [3]中国人民解放军61175部队,江苏南京210049

出　　处：《光学精密工程》2024年第19期2945-2956,共12页Optics and Precision Engineering

基　　金：国家自然科学基金资助项目(No.41671419);河南省交通运输厅科技计划项目(No.2022-3-2)。

摘　　要：为克服激光、相机在雷视监控设备测速任务中的采集频率、分辨率、视角等差异,提出一种影像优先与二维三维视图一体的多目标连续追踪测速方法。首先,均匀选取场景几何特征同名点对,采用直接线性变换法计算外参参数,在线完成相机与激光雷达间标定;然后,提出视觉引导的视锥体空间下二维三维一体目标定位方法,以高分辨率地面点计算区域界限初步去除背景点,将视锥三维视图转换为二维俯视图自适应计算目标聚类参数,以二维最密体素指引恢复至三维选取起始种子点完成聚类定位以消除无关点,解决视角导致的不同分辨率点云目标检测混淆问题;最后,基于卡尔曼滤波,考虑车辆运动状态,根据离散同步帧点云定位结果将测速过程建模为相应观测方程,并以点云分辨率估计观测噪声协方差矩阵,实现对离散测速结果的连续最优估计,解决观测噪声以及两设备时间异步的误差影响。在常见交通场景的图像+点云数据下进行实验,结果表明,本文方法速度测量的平均绝对误差为0.2764 m/s,均方根误差为0.3251 m/s,最远探测距离达到103.211 m,具有较高的准确性和实用性。This paper introduces an image-priority approach with integrated 2D-3D views for continuous multi-target tracking and speed estimation in camera-LiDAR surveillance systems.It addresses differences in acquisition frequency,resolution,and viewing angles between camera and LiDAR-based speed estima⁃tion tasks.Geometrically matched feature points are selected,and external parameters are computed using the Direct Linear Transformation method for online calibration between devices.A vision-guided,frus⁃tum-based spatial method combines 2D and 3D localization,uses high-resolution ground points to define area boundaries,and adapts clustering parameters in a 2D top-down view transformed from a 3D frustum perspective.This approach helps eliminate irrelevant points and address mixed-resolution point cloud de⁃tection due to varying viewing angles.The speed estimation process employs Kalman filtering and vehicle motion states,modeling speed estimation as observation equations using discrete synchronized frame point cloud data.The observation noise covariance matrix is calculated based on point cloud resolution,allow⁃ing continuous optimal estimation and reducing observation noise and asynchronous timing effects.Experi⁃ments on traffic scene datasets show that the method achieves an average absolute error of 0.2764 m/s and a root mean square error of 0.3251 m/s,with a maximum detection range of 103.211 m,demonstrat⁃ing high accuracy and practicality.

关 键 词：智能交通 速度测量 影像视锥 数据融合 卡尔曼滤波 目标检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] U491[自动化与计算机技术—计算机科学与技术]