Cross-View Urban Traffic Dataset: Drone-Supervised Ground Truth for Monocular Bird's-Eye View Localization

Researchers introduce a new cross-view urban traffic dataset combining synchronized drone and bicycle-mounted camera footage from real intersections. The benchmark enables two computer vision tasks: matching identical objects across street and aerial views, and predicting bird's-eye-view layouts from ground-level cameras with drone supervision.

This dataset addresses a critical gap in autonomous vehicle and traffic monitoring research by providing synchronized multi-viewpoint data from urban intersections. Current autonomous driving datasets typically focus on single egocentric perspectives, limiting development of systems that must reason about global traffic patterns and object identity across radically different viewpoints. The benchmark's identity-level alignment across street-view and drone-view presents a genuine technical challenge: machines must recognize the same pedestrian or vehicle from both ground and aerial perspectives simultaneously, a problem absent from existing datasets like nuScenes or Argoverse.

The motivation reflects real-world deployment scenarios where autonomous vehicles and smart city infrastructure increasingly need to fuse information from multiple sensor types and perspectives. Traffic analysis at intersections demands understanding both local interactions (how nearby objects relate) and global structure (where everything sits in the larger scene), requiring models that bridge ego-centric and bird's-eye-view reasoning. The benchmark's standardized evaluation metrics—including temporal stability and cross-view consistency—establish concrete measurement standards that were previously missing.

For the computer vision and autonomous vehicle industries, this dataset accelerates development of perception systems for complex urban scenarios. The baseline results reveal significant challenges: cross-view matching suffers from false positives and temporal inconsistency, while monocular bird's-eye-view prediction remains far from practical deployment levels despite aerial supervision. These gaps represent immediate research priorities and potential commercialization opportunities for companies developing intersection monitoring or autonomous delivery systems.

• →New benchmark combines synchronized drone and ground-level video from real intersections, enabling cross-view object identity matching and bird's-eye-view prediction research.
• →Baseline results show cross-view matching is feasible but limited by over-assignment errors and temporal instability in tracking.
• →Current monocular bird's-eye-view prediction methods remain far from practical deployment despite access to aerial supervision data.
• →Standardized evaluation metrics for temporal stability and cross-view consistency address gaps in existing autonomous driving datasets.
• →Dataset targets intersection-centric traffic analysis where understanding global spatial structure and local interactions jointly across views is essential.