Discrete-continuous clustering for obstacle detection using stereo vision

Efficient obstacle detection is a key requirement for safe robot navigation. We consider the operation of autonomous vehicles in structured industrial environments. In such scenarios, an usual way to perform obstacle detection is to generate an estimate of the ground and detect elements that are on the path of the vehicle, using the ground as a spatial reference. For this task, 3D occupancy grids are a well-know solution. In this work we extrapolate the concept of 3D grids by considering a discrete-continuous representation of the environment. The discrete nature lies in a 2D grid parallel to the ground whereas the continuous aspect represents the height of each cell in the grid. This framework allows for very efficient clustering, for which we also propose a novel algorithm to cluster potential obstacles. Experiments on an autonomous ground vehicle illustrate the applicability of the method.