Title: CityTrajBench: A Unified Benchmark for City-Scale Vehicle Trajectory Generation

Abstract:

The generation of vehicle trajectories at a city-wide scale is a critical component for mobility analytics, urban planning, and transportation simulation. Despite its importance, conducting systematic comparisons among various trajectory generation methods has proven challenging. Existing literature frequently employs disparate datasets, preprocessing workflows, trajectory representations, and evaluation metrics. This lack of standardization obscures whether performance variations stem from the core generation mechanisms or from inconsistencies in experimental protocols.

To resolve these inconsistencies, we introduce CityTrajBench, a comprehensive benchmark framework and standardized protocol designed for city-scale vehicle trajectory generation. CityTrajBench establishes a common setting that harmonizes data ingestion, trajectory normalization, feature engineering, model adaptation, map-aware post-processing, model selection, and multi-level evaluation. The framework accommodates a wide array of heterogeneous generators, ranging from statistical baselines to advanced architectures based on Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), diffusion models, and flow-matching techniques. These models are rigorously evaluated using three real-world urban trajectory datasets.

The benchmark assesses performance across five key dimensions: global spatial realism, trip-level distribution fidelity, trajectory-level geometric similarity, conditional mobility consistency, and computational efficiency. Our experimental results highlight distinct trade-offs among different model families. DiffTraj demonstrates superior performance in trajectory-level geometric fidelity, while DiffRNTraj excels in structure-sensitive global realism. TrajFlow offers a robust balance, performing well across realism, quality, conditional consistency, and efficiency. Notably, a simple Markov baseline remains highly competitive regarding coarse-grained trip statistics and local movement patterns. These insights underscore that the quality of urban trajectory generation is inherently multi-objective, with no single model achieving dominance across all criteria. CityTrajBench provides a reproducible testbed and protocol to facilitate future advancements in urban mobility generation research.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC