Map Construction With Multiobservation on Geographic Perspective

Abstract

Online vectorized map construction is essential for autonomous driving, yet camera-based approaches suffer significant performance degradation under adverse conditions due to limited sensing range and perspective constraints. We propose MapGeoTR, a multisource fusion framework that augments onboard perception with satellite imagery and OpenStreetMap (OSM) data as geographic priors. A lightweight ResNet-UNet encoder extracts top-down spatial features from auxiliary sources, which are spatially aligned and fused into the onboard bird’s-eye-view (BEV) representation via cross attention. Experiments on nuScenes show MapGeoTR achieves mAP of 46.1% and 68.0% at strict and relaxed thresholds, respectively, validating the effectiveness of multisource geographic priors for robust online map construction.