基于单目视觉的车辆3D空间检测方法  

Vehicle 3D Space Detection Method Based on Monocular Vision

作　　者：顾德英 张松 孟范伟 GU De-ying;ZHANG Song;MENG Fan-wei(School of Control Engineering,Northeastern University at Qinhuangdao,Qinhuangdao 066004,China)

机构地区：[1]东北大学秦皇岛分校控制工程学院,河北秦皇岛066004

出　　处：《东北大学学报（自然科学版）》2022年第3期328-334,共7页Journal of Northeastern University(Natural Science)

基　　金：河北省自然科学基金资助项目(F2019501012).

摘　　要：针对基于单目车辆检测的3D包围框检测精确率比较低的问题,提出了一种基于改进的FPN特征融合、ResNet残差单元、全连接层组合而成的新网络方法.在训练阶段,回归车辆的三维尺寸、残差角度和置信度;在推理阶段,检测出所属类别车辆的三维尺寸和局部角度(α).由车辆的3D包围框中心点坐标、车辆的三维尺寸、车辆偏航角(θ)和相机内参矩阵复原绘制出车辆的3D包围框.所提方法在KITTI验证集上进行了实验,与原方法的检测结果相比,改进的方法在容易、适中、困难三个检测等级下提升了车辆3D包围框平均精确率(AP_(3D))为0.60%,1.37%,1.41%.Aiming at the problem of low detection precision of 3D bounding box based on monocular vehicle detection,a new network method based on improved FPN(feature pyramid networks)feature fusion,ResNet residual unit,and fully connected layer was proposed.In the training phase,the three-dimensional size of vehicles,residual angle and confidence are regressed.In the reasoning phase,the three-dimensional size and local angle(α)of vehicles are detected.The 3D bounding box of vehicles are reconstructed and drawn from the center coordinates,the three-dimensional size of vehicles,the yaw angle(θ),and the camera intrinsic matrix.The proposed method is tested on the KITTI verification set.Compared with the results of the original method,the proposed method improves the average precision of 3D bounding box of vehicles(AP_(3D))to 0.60%,1.37%,and 1.41%,respectively,under the three detection levels of easy,moderate and difficult.

关 键 词：特征融合 单目车辆检测 计算机视觉 深度学习 KITTI 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]