摘要
针对由一般卷积神经网络预测的粗糙特征生成的深度图质量低、监督学习处理任务限制数据量等问题,提出一种新颖的融合扩张卷积神经网络和同时定位与建图(SLAM)的无监督单目深度估计方法。该方法采用视图重构的思想估计深度,利用光学一致性误差约束网络训练,扩大感受野,考虑图片细节特征。同时采用SLAM算法优化相机姿态,并将其嵌入视图重构框架中,实现单目图片与其深度图的直接映射。利用该方法在公开的KITTI数据集上进行实验,结果表明,与经典的sfmlearner方法相比,误差度量指标绝对差、平方差、均方差和对数均方差分别降低了0.032、0.634、1.095和0.026;准确率度量指标δ1、δ2和δ3分别提升了3.8%、2.6%和0.9%。该模型的可用性与稳健性得到验证。
The quality of a depth map generated by coarse features which are predicted by convolutional neural networks(CNNs)is low.Meanwhile,strong-supervised methods strictly limit the data volume due to lack of labeling.To address these problems,an unsupervised monocular depth estimation method by fusing dilated convolutional neural network and simultaneous localization and mapping(SLAM)is proposed.This method adopts the idea of view reconstruction to estimate depth.Photo-consistency error is utilized in the method to constrain training,expand the field of view,and concern the image details.Traditional SLAM algorithm functions to globally optimize the camera pose and incorporate it into the reconstruction framework.Finally the straight correspondence between the input monocular image and its depth map is exploited.The method is evaluated on the public KITTI dataset.The evaluation results show that,compared with the classical sfmlearner method,the error indicators,including absolute relative difference,squared relative difference,root mean squared error,and log root mean squared error,decrease by 0.032,0.634,1.095,and 0.026 respectively,and the accuracy indicators,δ1,δ2 andδ3,increase by 3.8%,2.6%,and 0.9%respectively.The availability and robustness of the proposed method are verified.
作者
戴仁月
方志军
高永彬
Dai Renyue;Fang Zhijun;Gao Yongbin(School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201600,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2020年第6期106-114,共9页
Laser & Optoelectronics Progress
基金
国家自然科学基金(61772328,61802253,61831018)
上海晨光人才计划(17CG59)
江西省经济犯罪调查和预防技术协作创新中心(JXJZXTCX-027)。
关键词
图像处理
扩张卷积神经网络
同时定位与建图
无监督学习
单目视觉
深度估计
image processing
dilated convolutional neural network
simultaneous localization and mapping
unsupervised learning
monocular vision
depth estimation
