Towards improving depth estimation techniques based on deep neural networks for autonomous vehicles

Abstract

The depth estimation task aims to generate a dense depth map corresponding to an RGB image and has been extensively studied as a fundamental problem in computer vision. In recent years, many researchers have actively investigated how to apply dense depth maps to various applications such as robotics and autonomous vehicles. An accurate dense depth map provided for a corresponding RGB image is particularly useful for understanding the three-dimensional geometric information of a scene for solving various computer vision tasks. Therefore, the depth estimation task which uses high-quality depth maps as prior information for RGB image processing is essential to academia and industry. A solution to improve the performance of the depth estimation task should be found. This dissertation addresses the depth estimation techniques based on deep neural networks. First, we propose a patch-wise attention module operating in local areas of a feature map to predict a dense depth map from a single RGB image. Next, we describe the following methods for inferring a dense depth map from a single RGB image and LiDAR data. We propose a cross guidance module for efficient multi-modal feature fusion of an RGB image and LiDAR data. Then, multiscale and densely connected locally convolutional layers are proposed. This module effectively learns the neighborhood's relationship in a local area with multiple scales.