[Paper Review] A Statistical Update of Grid Representations from Range Sensors

Luis Roldão, Raoul de Charette+1 more|arXiv (Cornell University)|Jul 23, 2018

Robotics and Sensor-Based Localization14 references3 citations

TL;DR

This paper proposes a statistical update method for 3D grid maps from range sensors that improves reconstruction accuracy by incorporating ray traversal distances and sensor density via a distance-weighted function. By weighting measurements based on cell distance from the sensor and considering all ray interactions per cell, the approach reduces discretization errors—particularly in ground regions—outperforming standard OctoMap in qualitative results with fewer holes.

ABSTRACT

In a wide range of robotic applications, being able to create a 3D model of the surrounding environment is a key feature for autonomous tasks. In this research report, we present a statistical model to perform 3D reconstructions of the environment from range sensors using an occupancy grid. To do so, we take into account all the available information obtained from the sensor, considering the distances traversed by the rays in each cell and seeking to reduce reconstruction errors caused by discretization. The approach has been validated qualitatively using the KITTI dataset.

Motivation & Objective

• To address discretization inaccuracies in 3D occupancy grid maps caused by binary cell updates from single sensor observations.
• To reduce reconstruction errors in ground and partially occupied regions by leveraging ray path information and measurement density.
• To improve map fidelity by statistically weighting occupancy updates based on sensor range and ray traversal distances.
• To evaluate the impact of weighting only miss measurements (traversed rays) versus both hits and misses on map quality.
• To provide a more accurate, statistically grounded update mechanism for volumetric grids in robotic 3D reconstruction.

Proposed method

• The method models the expected number of rays traversing a voxel at distance d using angular resolution and voxel size, defining a density function ρ(d) to estimate measurement frequency.
• A weighting function w(d) is applied to measurements based on distance from the sensor, reducing influence of distant, sparse observations.
• For hits (impacts), the occupancy probability is updated using a modified Bayes filter with P_occ = 0.7, while misses use P_free = 0.4, both adjusted by w(d).
• The method applies w(d) to miss measurements only in the second approach, reducing the impact of long-traveling rays on distant cells.
• The approach integrates with OctoMap’s octree structure, using a 0.2m voxel size and clamping probabilities between 0.12 and 0.97.
• The weighting function γ = 32 is tuned using sensor specs (φ_s = 0.4°, θ_s = 0.16°) from the Velodyne HDL-64E.

Experimental results

Research questions

• RQ1 How does incorporating ray traversal distance and sensor measurement density improve 3D grid map accuracy?
• RQ2 What is the effect of weighting only miss measurements versus both hits and misses on map reconstruction quality?
• RQ3 Can a statistical update model reduce discretization-induced errors in ground and partially occupied regions?
• RQ4 How does the proposed method compare to standard OctoMap in terms of hole formation and map density?
• RQ5 To what extent does the weighting function mitigate errors from sparse, long-range ray observations?

Key findings

• The method with weighting applied only to miss measurements (Fig. 4c) produced fewer holes in the ground region compared to standard OctoMap (Fig. 4a) and the full-weighting approach (Fig. 4b).
• The qualitative results show that weighting misses reduces the influence of distant, sparsely observed cells, leading to a denser and more accurate reconstruction of the environment.
• Applying weights to both hits and misses (Fig. 4b) did not significantly improve results over standard OctoMap, suggesting that hit weighting may not resolve discretization errors effectively.
• The approach successfully reduces discretization artifacts by prioritizing measurement density and ray path information, especially in areas with partial occupancy.
• The method maintains computational feasibility by using a 0.2m voxel size and integrating with the existing OctoMap octree structure.
• The results indicate that weighting miss measurements is more effective than weighting hits, as it prevents distant, low-confidence traversals from distorting occupancy probabilities.