Lifelong 3D Mapping Framework for Hand-held & Robot-mounted LiDAR Mapping Systems

TL;DR

The paper tackles the challenge of maintaining useful 3D maps over time in dynamic environments by introducing a modular, cloud-native Lifelong 3D Mapping Framework that supports both hand-held and robot-mounted LiDAR systems. The approach combines dynamic point removal, automatic two-stage multi-session map alignment, BEV-based map change detection, and a Git-like map version control that stores only changes and boundaries, enabling reconstruction of past sessions without raw inputs. Key contributions include sensor-setup agnostic dynamic removal, automatic parameter-free alignment, robust change detection across sessions, and memory-efficient map version control. The framework enables retrieval of clean session maps and querying of inter-session changes, offering significant practical value for long-term robot operation, digital twins, and semantic analysis of evolving environments.

Abstract

We propose a lifelong 3D mapping framework that is modular, cloud-native by design and more importantly, works for both hand-held and robot-mounted 3D LiDAR mapping systems. Our proposed framework comprises of dynamic point removal, multi-session map alignment, map change detection and map version control. First, our sensor-setup agnostic dynamic point removal algorithm works seamlessly with both hand-held and robot-mounted setups to produce clean static 3D maps. Second, the multi-session map alignment aligns these clean static maps automatically, without manual parameter fine-tuning, into a single reference frame, using a two stage approach based on feature descriptor matching and fine registration. Third, our novel map change detection identifies positive and negative changes between two aligned maps. Finally, the map version control maintains a single base map that represents the current state of the environment, and stores the detected positive and negative changes, and boundary information. Our unique map version control system can reconstruct any of the previous clean session maps and allows users to query changes between any two random mapping sessions, all without storing any input raw session maps, making it very unique. Extensive experiments are performed using hand-held commercial LiDAR mapping devices and open-source robot-mounted LiDAR SLAM algorithms to evaluate each module and the whole 3D lifelong mapping framework.

Figures (7)

• Figure 1: Users can upload multi-session 3D maps collected from any source to our proposed lifelong 3D mapping framework. Our system performs dynamic object removal, multi-session map alignment, map change detection and map version control. It allows users to retrieve any clean session map or query changes between any two sessions, all without storing the memory heavy input session maps.
• Figure 2: Our proposed sensor-setup agnostic dynamic point removal pipeline. We create a submap, regress multiple planes and add them back based on a ratio check to fill the residual holes from OctoMap. We perform radial search based post-processing to further improve the quality of static maps.
• Figure 3: Map Change Detection - Given $base\ map($t$)$ and $session\ map(t+1)$, the map change detection identifies negative differences $base\ ND(t)$ and positive differences $session\ PD(t+1)$. Negative difference is the objects that were present in base map but disappeared in session map while the positive difference is the map data that newly appeared in session map.
• Figure 4: Illustration of Complete Workflow. (a) Initialization of base map. Grey points are static and red points are dynamic. (b)$session\ map(t+1)$ first goes through dynamic object removal and is then aligned to $base\ map(t)$. The map change detection detects negative changes $base\ ND(t)$ and positive changes $session\ PD(t+1)$ while the map version control module updates the $base\ map(t)$ to $base\ map(t+1)$. The rectangular boxes highlight regions with changes and how the changes in $session\ map(t+1)$ get reflected in $base\ map(t+1)$. Specifically, the partly occluded region between parked cars, as highlighted in rightmost box in session map ($t+1$) is maintained as before, in updated $base\ map(t+1)$.
• Figure 5: Multi-Session map alignment results on multiple datasets with different local coordinate frames. Height difference is added to visualize & distinguish different trajectories.