VDAS中基于单目红外图像的深度估计方法  

Depth estimation method based on monocular infrared image in VDAS

作　　者：李旭 丁萌[1] 魏东辉[2] 吴晓舟[1] 曹云峰 LI Xu;DING Meng;WEI Donghui;WU Xiaozhou;CAO Yunfeng(College of Civil Aviation, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China;Science and Technology on Complex System Control and Intelligent Agent Cooperation Laboratory, Beijing Electro-Mechanical Engineering Institute, Beijing 100074, China;College of Astronautics, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China)

机构地区：[1]南京航空航天大学民航学院,江苏南京211106 [2]北京机电工程研究所复杂系统控制与智能协同技术重点实验室,北京100074 [3]南京航空航天大学航天学院,江苏南京211106

出　　处：《系统工程与电子技术》2021年第5期1210-1217,共8页Systems Engineering and Electronics

基　　金：装备预研航天科工联合基金(6141B07090114);南京航空航天大学研究生创新基地(实验室)开放基金(kfjj20190730,kfjj20190723)资助课题。

摘　　要：针对视觉辅助驾驶系统(vision-based driving assistance system,VDAS)对夜间等低能见度场景下的车辆前视深度感知需求,提出一种基于深度学习的单目红外图像深度估计方法。该方法采用端对端的多任务自监督学习框架,利用单目红外视频帧之间的立体几何约束构建损失函数,无须场景的真实深度信息。取前后两帧重投影误差的最小值,解决了单目视频帧之间的遮挡问题,同时削弱了红外图像噪点多的影响。网络解码器将多尺度深度图统一上采样到较高分辨率并计算重投影误差,改善了深度图中的空洞现象。在FLIR车载红外数据集上的定性实验表明,所提方法能利用单目红外图像得到对应的像素级稠密深度;真实道路上的分析试验表明,所提方法能够在夜间情况下利用红外图像有效感知目标物的深度信息,15 m以内的相对误差为13.2%,可以满足多数突发情况下情况下的避撞要求。In view of the demand of vision-based driving assistance system(VDAS)for vehicle forward-looking depth perception in low visibility scenes,a depth learning based monocular infrared image depth estimation method is proposed.In this method,an end-to-end multi-task self-monitoring learning framework is adopted,and the loss function is constructed by using the stereo geometric constraints between monocular infrared video frames,so the real depth information of the scene is not need.The minimum value of the reprojection error between the two frames is taken to solve the occlusion problem,and the influence of infrared image noise is weaken;The network decoder samples the multi-scale depth map to higher resolutions and calculates the reprojection error,which avoids the hole phenomenon in the depth map.Qualitative experiments on FLIR infrared datasets show that the proposed method can obtain pixel-level dense depth from monocular infrared images.Experiments on real roads show that the proposed method can effectively perceive the depth information of the target in the night,and the absolute error is 13.2%within 15 m.It can meet the requirements of collision avoidance in most emergency situations.

关 键 词：红外图像 深度估计 卷积神经网络 自监督学习 

分 类 号：TP39[自动化与计算机技术—计算机应用技术]