Automatic 3-D Point Cloud Classification of Urban Environments

摘要

This paper addresses the problem of assigning a label to three- dimensional data points collected from laser scanners. We are specifically interested in the application of environment modeling for autonomous robot navigation in natural and urban terrains. To capture contextual information, we choose to work within the Markov Random Field framework. The approach used in this paper is a variant of the Associative Markov Network (AMN), extended to learn directionality in the clique potentials, resulting in a new anisotropic model that can be efficiently learned using a gradient-based method for non- differentiable function. We validate the proposed approach using data collected from different range sensors.