Mapping Trail Networks From Crowdsourced Hiking GPS Traces

ABSTRACT

Reconstructing trail networks from GPS data is a challenging task due to the heterogeneity, sparsity, and noise inherent in crowdsourced traces. To address the dichotomy between maintaining topological connectivity and handling severe geometric noise, we present two complementary approaches for automatic trail map generation from large‐scale GPS datasets. The first is an unsupervised multigrid method that aggregates traces using a grid‐based representation, specifically addressing the preservation of structural connectivity without requiring annotated data. The second is a supervised deep learning approach based on the HRNet architecture, trained with automatically generated annotations derived from OpenStreetMap (OSM), designed to filter erratic GPS behaviors and handle complex spatial variations. Both methods are evaluated on real‐world hiking data provided by Visorando, alongside multiple benchmark datasets. We also propose dedicated evaluation metrics, including a novel intersection F ‐score (). Quantitatively, the multigrid method demonstrates high accuracy and connectivity preservation on structured or moderately noisy datasets (achieving an of 0.82 on Blaesheim and up to 0.89 on baseline pedestrian data). However, in highly complex, unconstrained mountain environments, the supervised HRNet model demonstrates superior robustness, reaching an of 0.520 and a topological intersection score () of 0.805 when trained with OSM‐augmented data. Overall, our results highlight the specific trade‐offs of each approach, providing scalable solutions for automatic trail mapping in natural environments.