Object-Oriented Grid Mapping in Dynamic Environments

TL;DR

This work relaxes the standard independence assumption of grid-map cells by introducing latent variables $\,\mathcal{D}_t$ that encode object-level dependencies across cells. It augments the inverse sensor/update model with these latent dependencies and estimates both map occupancies and latent memberships, using a region-growing clustering guided by semantic labels and an extension called C-NDT-OM. Empirical results on real-world datasets show fewer residual dynamic cells and faster occlusion removal, improving map quality in dynamic environments and offering modest localization gains. The approach advances object-aware occupancy mapping, with practical implications for planning and long-term map maintenance in changing environments.

Abstract

Grid maps, especially occupancy grid maps, are ubiquitous in many mobile robot applications. To simplify the process of learning the map, grid maps subdivide the world into a grid of cells whose occupancies are independently estimated using measurements in the perceptual field of the particular cell. However, the world consists of objects that span multiple cells, which means that measurements falling onto a cell provide evidence of the occupancy of other cells belonging to the same object. Current models do not capture this correlation and, therefore, do not use object-level information for estimating the state of the environment. In this work, we present a way to generalize the update of grid maps, relaxing the assumption of independence. We propose modeling the relationship between the measurements and the occupancy of each cell as a set of latent variables and jointly estimate those variables and the posterior of the map. We propose a method to estimate the latent variables by clustering based on semantic labels and an extension to the Normal Distributions Transform Occupancy Map (NDT-OM) to facilitate the proposed map update method. We perform comprehensive map creation and localization experiments with real-world data sets and show that the proposed method creates better maps in highly dynamic environments compared to state-of-the-art methods. Finally, we demonstrate the ability of the proposed method to remove occluded objects from the map in a lifelong map update scenario.

Figures (6)

• Figure 1: Occlusions and dynamic objects constitute a major source of mapping errors. When not treated properly, partial objects and other residual noise remain in the map.
• Figure 2: Inverse sensor model of the generic mapping method (a) and the proposed method (b). The vertices in gray are assumed to be known, while the vertices in white, i.e., the map cells $m^i$ and the latent variables $d^{(i,j)}$, are unknown. We present the inverse sensor model instead of the usual likelihood model for clarity.
• Figure 3: The initial map (a) is updated in the lifelong scenario with (b) - (d), and our method (e) and (f). Our method outperforms in update speed.
• Figure 4: The initial map (a) is updated in the lifelong scenario with NDT-OM (b) while our method (c) removes the occluded parts of the object.
• Figure 5: The initial map (a) is updated in the lifelong scenario with our method (b). The correct state after the update is shown in (c). Our method erroneously clusters all of the vehicles into a single cluster and removes the whole cluster.