Depth Estimation in Monocular Images of Inside the Sewer Pipes for Navigation of Inspection Robots

Estimating the depth of a scene from single monocular color images is a fundamental problem in image understanding. Depth estimation is a critical function for robotic tasks such as localization, mapping, and obstacle detection. Recent works based on the development of deep convolutional neural networks provide reliable results. Due to the low cost and relatively small size of monocular cameras, the performance of neural networks for depth estimation from single RGB image has increased significantly. Inspector robots move inside the sewer pipe in an ambiguous environment that has various contaminations and obstacles. As a result of understanding the environment inside the pipe and analyzing the images from the monocular camera, the robot can move more safely and perform the mission more effectively. This paper presents a new deep neural network, called SepiDepth-ASPP. Our approach uses the integration of ASPP and adaptive bins to extract strong global and local contextual features at multiple scales, and then translate them to higher resolutions for clearer depth maps. This network is specially designed for images inside the sewer pipe to more accurately estimate the details of the images in the depth map.This network runs on the dataset inside the sewer pipe and helps the robot comprehend the inside of the pipe environment.