Ephemerality meets LiDAR-based Lifelong Mapping

TL;DR

ELite tackles the problem of long-term LiDAR mapping in dynamic environments by introducing a two-stage ephemerality model that distinguishes transient from persistent map elements. It unifies three modules—multi-session map alignment, dynamic object removal, and long-term map update—under ephemerality-guided weighting, producing lifelong, static, and delta maps. The system leverages local ephemerality ($\epsilon_l$) within sessions and global ephemerality ($\epsilon_g$) across sessions to robustly align data, remove moving objects without voxel discretization, and apply category-specific updates that emphasize meaningful changes. Experimental results across multiple real-world datasets demonstrate improved alignment robustness, effective dynamic object removal, and rich inter-session change representations, with a public implementation to support adoption. This approach enables more reliable long-term operation of robots in changing environments and supports downstream analysis such as change heatmaps and time-domain dynamics.

Abstract

Lifelong mapping is crucial for the long-term deployment of robots in dynamic environments. In this paper, we present ELite, an ephemerality-aided LiDAR-based lifelong mapping framework which can seamlessly align multiple session data, remove dynamic objects, and update maps in an end-to-end fashion. Map elements are typically classified as static or dynamic, but cases like parked cars indicate the need for more detailed categories than binary. Central to our approach is the probabilistic modeling of the world into two-stage $\textit{ephemerality}$, which represent the transiency of points in the map within two different time scales. By leveraging the spatiotemporal context encoded in ephemeralities, ELite can accurately infer transient map elements, maintain a reliable up-to-date static map, and improve robustness in aligning the new data in a more fine-grained manner. Extensive real-world experiments on long-term datasets demonstrate the robustness and effectiveness of our system. The source code is publicly available for the robotics community: https://github.com/dongjae0107/ELite.

Figures (11)

• Figure 1: An example scene from three KAIST sequences kim2020mulranjung2023helipr with newly appeared walls and parked cars. Representing changes in a simple binary manner, existing methods treat both the car and the wall as static objects. Our proposed system leverages two-stage ephemerality to differentiate parked cars as ephemeral objects and walls as persistent changes based on their ephemerality scores.
• Figure 2: Overview of the ELite system pipeline. Given multiple input sessions, ELite updates the map by estimating local ephemerality within each session and updating global ephemerality across sessions. The system operates through three modules: multi-session map alignment, dynamic object removal, and map update. Additionally, the system manages three types of maps: a lifelong map, a delta map, and a static map.
• Figure 3: Illustration of our map alignment module, which begins by aligning poses using the initial transform from the first loop closure candidate. It then refines alignment in two stages: forward and backward. By iterating through scans in both directions, the module updates poses via scan-to-map ICP registration, ensuring global and local consistency in the final pose estimates.
• Figure 4: When voxelizing space, as indicated by the green square, the occupied area is inflated and errors can occur when a single voxel mixes static and dynamic points. Our method updates point-wise ephemerality based on ray information, enabling more precise removal of dynamic objects.
• Figure 5: In our map update module, points are classified into five categories. Coexisting points ($\mathcal{C}_t$) are shown in grey. Deleted points ($\mathcal{D}_t$) are red if they truly disappeared and pink if they belong to previously visited regions only. Emerged points ($\mathcal{E}_t$) are blue if newly added and sky blue if observed only in the current session. Each category follows a specific update strategy for robust map maintenance.