Probabilistic Height Grid Terrain Mapping for Mining Shovels using LiDAR

TL;DR

The approach extends the height grid representation of terrain to include a Hidden Markov Model in each cell, enabling confidence-based mapping of constantly changing terrain, in an attempt to support autonomous machine operation.

Abstract

This paper explores the question of creating and maintaining terrain maps in environments where the terrain changes. The specific example explored is the construction of terrain maps from 3D LiDAR measurements on an electric rope shovel. The approach extends the height grid representation of terrain to include a Hidden Markov Model in each cell, enabling confidence-based mapping of constantly changing terrain. There are inherent difficulties in this problem, including semantic labelling of the LiDAR measurements associated with machinery and determining the pose of the sensor. Solutions to both of these problems are explored. The significance of this work lies in the need for accurate terrain mapping to support autonomous machine operation.

Figures (5)

• Figure 1: A terrain map generated using the proposed approach with point clouds from a 3D LiDAR mounted to the shovel's house represented by the red coordinate frame. An example LiDAR scan is shown in green, with models of the shovel and truck.
• Figure 2: The conversion of the voxelized map (left) to a height grid (right) constructed using the voxelized measurements, $P_{\mathcal{M},k}^{'}$, and the raycast result to find all the observed voxels, $P_{\mathcal{M},k}^{'obs}$. For each observed cell, $c_{i}$, the maximum of consecutive measurements is used to update the global map, $M_{\mathcal{M},k}$. Measurements in $P_{\mathcal{M},k}^{'}$ that are above observed free voxels in $P_{\mathcal{M},k}^{'obs}$ are disregarded as they are likely to be noisy measurements from dust or missed in the semantic labelling process.
• Figure 3: A summary of the map update process. The point cloud is first semantically labelled to estimate points associated with the terrain and then voxelized to find the observed space. This information is used to convert to a height grid and compute the likelihood of each cell being in the $n$ states. The observation is fused with the existing map using the HMM filter.
• Figure 4: The net change in observed volume calculated for two datasets. The datasets are manually labelled to provide a reference of the shovel's behaviour during the dataset collection. Red squares indicate the beginning of a dig and blue diamonds are the dumping of material in a haul truck.
• Figure 5: Two examples visualizing the change in the state of the terrain over a receding window of 1000 scans. Red cells indicate the removal of material $(h_{new,i} < h_{old,i})$ and blue cells represent the addition of material $(h_{new,i} > h_{old,i})$. Both instances show areas of excavation and material spillage.