面向交通场景基于双注意力机制和自适应代价卷的自监督单目深度估计

Self-Supervised Monocular Depth Estimation for Traffic Scenes Based on Dual Attention Mechanism and Adaptive Cost Volume

针对当前交通场景下自监督单目深度估计存在特征表达能力弱、深度图局部细节模糊、深度估计精度低的问题,提出一种基于双注意力机制和自适应代价卷的自监督单目深度估计方法.该方法首先利用双注意力机制的特征提取网络,结合通道注意力和空间注意力,对提取的场景特征进行自适应加权,增强特征表达能力.其次,根据提取的全局特征自适应的构建代价卷,引导网络学习精细的深度特征,提升网络模型对深度图局部细节的学习能力,解决现有方法深度估计精度低的问题.在自动驾驶公开数据集KITTI、Cityscapes上的实验结果表明,本文方法优于目前主流方法.

Aiming at the problems of self-supervised monocular depth estimation in current traffic scenarios,such as weak feature expression ability,fuzzy local details of depth map and low accuracy of depth estimation,a self-supervised monocular depth estimation method based on dual attention mechanism and adaptive cost volume is proposed.Firstly,a dual attention mechanism combining channel attention and spatial attention is used to adaptively weight the extracted scene features to enhance the feature expression ability of the feature extraction network.Secondly,according to the adaptively constructed cost volume of extracting global features,the network is guided to learn fine depth features,which improves the learning ability of the network model for the local details of the depth map and solves the problem of low accuracy of existing depth estimation methods.Experimental results on public datasets KITTI and Cityscapes show that the proposed method is superior to the current mainstream methods.