【作者】 李响；

【导师】 康崇； 王建生；

【作者基本信息】 哈尔滨工程大学 ， 工程硕士（专业学位）， 2023， 硕士

【摘要】 单目深度估计是目前的一个热门研究方向,尤其是在自动驾驶领域,获取准确的深度信息有着举足轻重的作用。基于无监督学习的单目深度估计是当前深度估计领域的研究热点,可以快速获取图像中所有点的相对深度。目前该方法主要是通过可见光数据集进行训练,因此受环境影响较大,例如在夜间或沙尘、暴雨等光线较暗的恶劣天气中,得到准确的深度估计是比较困难的,实用性大大降低。因此,本文针对该问题,研究如何将基于无监督学习的单目深度估计算法应用于受光照影响相对更小的红外图像,并根据红外图像的特性和实际应用需求对算法进行改进,以使模型可以适用于夜间或光线较暗的恶劣天气行车,提高行车安全性。本文旨在解决在夜间或恶劣天气行车时,较难准确获取行车过程中行人和车辆目标真实距离的问题。所做的主要工作包括:1.红外数据集的制作。使用烟台艾睿光电科技有限公司所研发的HD1280单目红外相机拍摄红外数据集28000张,分辨率为1280×1024。该数据集主要在户外车辆行人较多的道路上进行采集,该数据集包含了大量的道路、车辆、行人目标以及树木房屋等场景,适用于自动驾驶研究。2.将单目无监督深度估计方法应用于红外图像,以获取红外图像的相对深度。由于目前已有的单目无监督深度估计算法大多基于可见光图像设计,针对上述情况,需要对无监督单目深度估计算法进行改进以使其适用于红外图像,我们提出采用单应性估计网络改进原模型中的姿态估计网络,进而使单目深度估计结果更加精确;由于训练所用数据集中不能包含动态场景,我们提出自动屏蔽掩码用以屏蔽图像中的动态场景。3.提出红外图像绝对深度估计算法,该算法可用于获取红外图像中行人和车辆目标相对于相机的距离。结合YOLOv5目标检测算法,可以获取红外图像中的行人和车辆目标,通过算法计算相对深度与绝对深度之间的尺度因子,将深度估计模型中得到的行人和车辆目标的相对深度转换为绝对深度。通过在自行构建的红外图像数据集上训练本文的算法模型,实验结果表明,该方法在20米范围内测距平均相对误差为7.32%,在10米内相对误差基本可以控制在7%以下,而超过10米后的相对误差较大,由此可以实现短距离较为精确的测距,同时单幅图像的推理速度也可以满足工程实时性要求。更多还原

【Abstract】 Monocular depth estimation is currently a hot research direction,especially in the field of automatic driving,where obtaining accurate depth information plays a decisive role.Monocular depth estimation based on unsupervised learning is a research hotspot in the field of depth estimation,which can quickly obtain the relative depth of all points in the image.At present,this method is mainly trained through visible light data sets,so it is greatly affected by the environment.For example,it is difficult to obtain accurate depth estimation at night or in bad weather such as sand,dust,and rainstorms,and the practicability is greatly reduced..Therefore,this thesis aims at this problem and studies how to apply the monocular depth estimation algorithm based on unsupervised learning to infrared images that are relatively less affected by illumination,and improve the algorithm according to the characteristics of infrared images and practical application requirements,so that the model It can be applied to driving at night or in bad weather with low light to improve driving safety.This thesis aims to solve the problem that it is difficult to accurately obtain the real distance between pedestrians and vehicle targets during driving at night or in bad weather.The main work done includes:1.Production of infrared data sets.The HD1280 monocular infrared camera developed by Yantai Iray Optoelectronics Technology Co.,Ltd.was used to shoot 28,000 infrared data sets with a resolution of 1280×1024.This data set is mainly collected on roads with many outdoor vehicles and pedestrians.This data set contains a large number of scenes such as roads,vehicles,pedestrian targets,trees and houses,and is suitable for autonomous driving research.2.Apply the monocular unsupervised depth estimation method to the infrared image to obtain the relative depth of the infrared image.Since most of the existing monocular unsupervised depth estimation algorithms are designed based on visible light images,in view of the above situation,it is necessary to improve the unsupervised monocular depth estimation algorithm to make it suitable for infrared images.We propose to use the homography estimation network to improve The pose estimation network in the original model makes the monocular depth estimation results more accurate;since the data set used for training cannot contain dynamic scenes,we propose an automatic masking mask to shield the dynamic scenes in the image.3.An infrared image absolute depth estimation algorithm is proposed,which can be used to obtain the distance of pedestrian and vehicle targets relative to the camera in infrared images.Combined with the YOLOv5 target detection algorithm,pedestrian and vehicle targets in infrared images can be obtained,and the scale factor between relative depth and absolute depth can be calculated through the algorithm,and the relative depth of pedestrian and vehicle targets obtained in the depth estimation model can be converted into absolute depth.By training the algorithm model on the self-constructed infrared image dataset,the experimental results show that the average relative error of the method in the range of 20 meters is 7.32%,and the relative error can be basically controlled below 7% within 10 meters,and the relative error after more than 10 meters is large,which can achieve more accurate ranging in short distances,and the inference speed of a single image can also meet the real-time requirements of engineering.更多还原