Graph-Loc: Robust Graph-Based LiDAR Pose Tracking with Compact Structural Map Priors under Low Observability and Occlusion

Map-based LiDAR pose tracking is essential for long-term autonomous operation, where onboard map priors need be compact for scalable storage and fast retrieval, while online observations are often partial, repetitive, and heavily occluded. We propose Graph-Loc, a graph-based localization framework that tracks the platform pose against compact structural map priors represented as a lightweight point-line graph. Such priors can be constructed from heterogeneous sources commonly available in practice, including polygon outlines vectorized from occupancy/grid maps and CAD/model/floor-plan layouts. For each incoming LiDAR scan, Graph-Loc extracts sparse point and line primitives to form an observation graph, retrieves a pose-conditioned visible subgraph via LiDAR ray simulation, and performs scan-to-map association through unbalanced optimal transport with a local graph-context regularizer. The unbalanced formulation relaxes mass conservation, improving robustness to missing, spurious, and fragmented structures under occlusion. To enhance stability in low-observability segments, we estimate information anisotropy from the refinement normal matrix and defer updates along weakly constrained directions until sufficient constraints reappear. Experiments on public benchmarks, controlled stress tests, and real-world deployments demonstrate accurate and stable tracking with KB-level priors from heterogeneous map sources, including under geometrically degenerate and sustained occlusion and in the presence of gradual scene changes.

💡 Research Summary

Graph‑Loc addresses the long‑standing challenge of map‑based LiDAR pose tracking under strict onboard memory constraints, partial observability, and heavy occlusion. The authors propose to represent both the prior map and each incoming LiDAR scan as lightweight point‑line graphs. Nodes correspond to sparse geometric primitives—corner points and line segments—while edges encode local k‑nearest‑neighbor relationships, providing contextual information without inflating the data size. Such graphs can be generated from heterogeneous sources, including vectorized occupancy‑grid outlines, CAD models, or floor‑plan layouts, making the approach applicable to a wide range of practical deployments.

During operation, a constant‑velocity motion model predicts the current pose. Using this prediction, the system performs a ray‑casting simulation to extract a pose‑conditioned visible subgraph from the global map, thereby limiting the matching problem to structures that are actually observable from the current viewpoint. This step dramatically reduces the number of candidate correspondences and eliminates spurious matches caused by occluded geometry.

The core of the matching algorithm is an unbalanced optimal transport (UOT) formulation applied to the nodes of the observation and visible‑map graphs. Unlike classic nearest‑neighbor pipelines that enforce strict one‑to‑one assignments, UOT relaxes mass conservation, allowing the algorithm to gracefully ignore missing or noisy primitives. To further enforce structural consistency, a graph‑context regularizer couples the transport plan across neighboring nodes, ensuring that matched pairs respect the local topology of both graphs. The resulting soft assignment is globally optimal, robust to repetitive patterns and fragmented observations, and can be solved efficiently using entropic regularization techniques.

Even with robust matching, low‑observability segments (e.g., long straight corridors, repetitive wall patterns) can lead to ill‑conditioned pose updates because the information matrix becomes anisotropic. Graph‑Loc tackles this by estimating information anisotropy from the normal matrix of the refinement step. Eigen‑analysis identifies motion directions that are weakly constrained; updates along these directions are temporarily deferred. The system accumulates additional measurements until sufficient constraints appear, at which point a full 3‑DoF (x, y, yaw) update is released. This degeneracy‑aware delayed optimization prevents drift accumulation and stabilizes tracking during prolonged occlusion or geometric degeneracy.

The authors evaluate the method on several fronts. On public benchmarks such as KITTI and the ERPoT dataset, Graph‑Loc achieves sub‑meter accuracy while using map representations on the order of a few hundred kilobytes—orders of magnitude smaller than dense point‑cloud maps. Controlled simulations based on CMU‑EXPLORATION introduce increasing pedestrian occlusion; Graph‑Loc maintains low positional error even when occlusion exceeds 70 %, outperforming LOAM‑style, KISS‑MCL, and HDL‑LOC baselines. Real‑world experiments in an outdoor parking lot and indoor corridor demonstrate stable long‑duration tracking despite dynamic obstacles, temporary structural changes (parked vehicles), and sustained low‑visibility periods. The delayed optimization component alone reduces drift by more than 30 % in degenerate scenarios.

In summary, Graph‑Loc combines three novel contributions: (1) a compact point‑line graph representation that unifies heterogeneous map sources, (2) an unbalanced optimal transport matcher with graph‑context regularization for globally consistent scan‑to‑map association, and (3) an anisotropy‑driven delayed update scheme that safeguards pose estimation under weak geometric constraints. The system delivers accurate, robust, and memory‑efficient LiDAR localization suitable for long‑term autonomous operation in structured environments such as buildings, corridors, and industrial sites.