Geometric Late-Fusion of GPS and Satellite-based Road Network Reconstructions.

Abstract

This master's thesis presents a late-fusion algorithm for multi-modal road network reconstruction that combines GPS trajectory data and satellite imagery-based road maps. GPS-based methods like Roadster capture road connectivity but produce geometrically imprecise networks with sparse coverage, while satellite-based methods such as Sat2Graph provide geometric detail and complete spatial coverage but have connectivity limitations. This work develops a geometric algorithm to resolve road continuation conflicts — disagreements between maps about whether roads connect — by combining the complementary strengths of unimodal map reconstructions.

The main research question examines whether map fusion improves map similarity by resolving conflicts in road continuation between two uni-modal map reconstructions. The thesis provides four primary contributions: introducing road continuation as an edge property describing connectivity; developing edge coverage using Fréchet distance to identify road continuation conflicts; designing a three-step fusion algorithm (edge injection, deletion, and reconnection); and introducing TOPO* and APLS* metrics to handle comparisons between graphs with different edge densities.

The methodology employs a late-fusion approach operating on fully reconstructed unimodal maps from Roadster (GPS-based) and Sat2Graph (satellite-based). The fusion algorithm uses edge coverage computations with Fréchet distance to assess road continuation. The approach is validated using datasets from Berlin and Chicago, with GPS trajectory data from 2018 and satellite imagery from 2024, evaluated using TOPO and APLS similarity metrics alongside the newly introduced TOPO* and APLS* variants.

The experimental results showed that map fusion did not systematically improve reconstruction performance over the best-performing unimodal methods. Fusion effectiveness varied between regions and input map quality, with Berlin favoring GPS-based approaches and Chicago favoring satellite-based approaches. This indicated that fusion performance depends on the relative quality of input maps rather than the fusion process itself. Performing only edge injection often outperformed the complete fusion pipeline, suggesting that duplicated road infrastructure performs better than single infrastructure with potential misalignment. A key limitation emerged in distinguishing road continuation disagreement from road existence disagreement, preventing definitive answers about road continuation conflict resolution.

The map similarity metrics proved inadequate for comparing maps with different edge densities. Future work should focus on improving road continuation detection by distinguishing it from road existence disagreement, potentially through Hausdorff distance integration and bidirectional edge coverage requirements. Enhanced evaluation methods and more extensive datasets are also needed to better understand fusion performance across different urban contexts.

The main finding was that fusion performance depends on retaining edges from higher-quality source maps rather than true integration, with duplicated road infrastructure often outperforming attempts to reduce graph complexity. Despite measurement challenges preventing definitive answers about road continuation conflict resolution, this work contributed a complete map fusion pipeline, introduced road continuation as a distinct map disagreement type, and developed techniques to measure it using Fréchet distance. The findings suggest that future multi-modal reconstruction efforts should prioritize addressing road existence and geometry alignment before tackling continuation issues.