文献类型：期刊文献

中文题名：基于红外与雷达的夜间无人车场景深度估计

英文题名：Depth Estimation of Night Driverless Vehicle Scene Based on Infrared and Radar

作者：姚广顺[1,2];孙韶媛[1,2];方建安[1,2];赵海涛[3]

机构：[1]东华大学信息科学与技术学院,上海201620;[2]东华大学数字化纺织服装技术教育部工程研究中心,上海201620;[3]华东理工大学信息科学与工程学院,上海200237

年份：2017

卷号：54

期号：12

起止页码：158

中文期刊名：激光与光电子学进展

外文期刊名：Laser & Optoelectronics Progress

收录：CSTPCD;;Scopus;北大核心:【北大核心2014】；CSCD:【CSCD2017_2018】；

基金：国家自然科学基金(61375007);上海市科委基础研究项目(15JC1400600)

语种：中文

中文关键词：图像处理;红外图像;深度估计;卷积神经网络;反卷积

外文关键词：image processing; infrared image; depth estimation; convolutional neural networks; deconvolution

摘要：单目红外图像的深度估计是夜间无人车场景理解的关键,针对夜间无人车场景的深度估计,提出一种基于深度卷积-反卷积神经网络的深度估计方法。将红外图像和雷达距离数据作为深度卷积-反卷积神经网络的输入,并将深度估计问题转化为像素级分类任务进行深度估计模型的训练。将雷达的距离数据根据深度值的范围量化为与红外图像像素一一对应的离散值并对其做标记,然后训练过程采用分类的思想解决深度估计问题。实验结果表明,利用训练得到的深度估计模型对夜间无人车获取的红外图像进行深度估计的时间为0.04s/frame,达到了实际应用中的实时性要求。
Depth estimation of monocular infrared image is a key to scene understanding of night driverless vehicle. Aiming at the depth estimation of night driverless vehicle scene; a depth estimation method based on the deep convolution-deconvolution neural network is proposed. Infrared images and radar depth data are fed to the deep convolution-deconvolution neural network. The depth estimation problem is transformed to a pixel-wise classification task in the training of the depth estimation model. The radar depth values are quantized into discrete bins corresponding to the pixels of infrared image and the bins are labeled according to their depth range. The deep convolution-deconvolution neural network based depth estimation model is trained by classifying each pixel to the corresponding depth. The experimental results show that the depth estimation time is 0.04 s/frame, which use the depth estimation model to estimate the scene depth information of infrared image captured by the night driverless vehicle, and the real-time requirement in practical applications is reached.