CityTrajBench: A Unified Benchmark for City-Scale Vehicle Trajectory Generation
Summary
CityTrajBench is introduced as a unified benchmark framework and protocol designed to standardize city-scale vehicle trajectory generation research. It addresses the current fragmentation caused by diverse datasets, preprocessing, representations, and evaluation metrics, which hinder systematic comparison of methods. The benchmark standardizes data ingestion, normalization, feature construction, model adaptation, map-aware post-processing, model selection, and multi-level evaluation. It supports heterogeneous generators, including statistical baselines, VAE-based, GAN-based, diffusion-based, and flow-matching-based models, evaluating them across three real-world urban trajectory datasets. CityTrajBench measures global spatial realism, trip-level distribution fidelity, trajectory-level geometric similarity, conditional mobility consistency, and efficiency. Experimental results indicate that DiffTraj performs strongest on trajectory-level geometric fidelity, DiffRNTraj is competitive on structure-sensitive global realism, and TrajFlow achieves a strong balance across multiple criteria, while a simple Markov baseline remains competitive for coarse-grained statistics. These findings underscore the multi-objective nature of urban trajectory generation quality.
Key takeaway
For Machine Learning Engineers and Research Scientists developing or selecting urban trajectory generation models, CityTrajBench provides a critical standardized evaluation framework. You should use this benchmark to accurately compare model performance across diverse metrics, understanding that no single model excels universally. This prevents misinterpreting performance differences due to inconsistent protocols. Use its findings to select models like DiffTraj for geometric fidelity or TrajFlow for balanced quality, aligning with your specific project objectives.
Key insights
CityTrajBench unifies urban trajectory generation evaluation, revealing multi-objective quality and no single model dominance.
Principles
- Urban trajectory generation quality is inherently multi-objective.
- No single model dominates all trajectory generation criteria.
- Standardized benchmarks are crucial for systematic method comparison.
Method
CityTrajBench standardizes data ingestion, normalization, feature construction, model adaptation, map-aware post-processing, model selection, and multi-level evaluation for trajectory generation.
In practice
- Evaluate DiffTraj for high geometric fidelity needs.
- Consider DiffRNTraj for structure-sensitive global realism.
- Use TrajFlow for balanced performance across metrics.
Topics
- CityTrajBench
- Vehicle Trajectory Generation
- Urban Mobility
- Benchmark Frameworks
- Diffusion Models
- Flow Matching
Best for: AI Scientist, Machine Learning Engineer, Research Scientist
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.