EN FR
EN FR
RITS - 2018
New Software and Platforms
Bilateral Contracts and Grants with Industry
Bibliography
New Software and Platforms
Bilateral Contracts and Grants with Industry
Bibliography


Section: New Results

A Statistical Update of Grid Representations from Range Sensors

Participants : Luis Roldao, Raoul de Charette, Anne Verroust-Blondet.

An accurate 3D model of the surrounding environment is a fundamental feature for autonomous vehicles to perform different tasks such as obstacle detection, localization and mapping. While continuous representations are widely used in the literature, we prefer to use a three dimensional discrete grid representation in this work in order to reduce memory an computational complexity. In this case, each grid cell represents the occupancy state of a portion of the environment in a probabilistic manner.

By definition, a discretized representation inhibits a completely accurate reconstruction. Therefore, grid models are unable to create a perfect model of the surroundings. In the literature, it is usually considered that within a single scan, the state of each cell is binary (free or occupied). Hence, a cell is set occupied if at least one impact occurred within, and free if it has been traversed by any ray. The problem of such an approach is that the complete state of the cell is updated from a single partial observation, neglecting the contribution of multiple measurements and their validity. Moreover, the traversed distance of the rays within each cell is usually ignored.

Towards the goal of achieving a more accurate representation, we propose a different way to update the occupancy probability of each cell according to the observations; considering the traversed distance of the rays within each cell (ray-path information), the contribution of the complete set of observations within the cell, and the density of observations that can be obtained at such cell according to its distance from the sensor. Proposed method was evaluated in both simulated and real data. Reconstruction results show an improvement on the representation of the surroundings with less occupancy state errors in the cells of the grid. Future works will include the comparison against a continuous representation to test the accuracy along with the time and computation needs for both representations.

More details can be found in [38] and [30]. This research is partially funded by AKKA Technology.