PixelLaser: Evaluating monocular range from texture

摘要

This work presents and evaluates the PixelLaser system, designed to estimate range-to-obstacle scans from single images. Its visual pipeline uses nearest-neighbor texture-matching to segment groundplane (traversable) texture from non-groundplane (obstacle) texture. Using the known pose of a webcamera, the system then transforms these segmentations into range-scans. This paper presents a thorough evaluation of the range accuracy obtained in the resulting scans. In addition, it presents three practical robot applications that use PixelLaser's monocular ranging: (a) a safe-wandering agent, (b) a Monte-Carlo localizer, and (c) a mapping system using CoreSLAM. In each case, algorithms that usually depend on laser scans have succeed with PixelLaser data instead.