TrajLearn: Trajectory Prediction Learning using Deep Generative Models

TL;DR

TrajLearn addresses trajectory prediction by representing histories as sequences of hexagonal blocks and learning $P(b_{l+1},\dots,b_{l+k}|B)$ with a decoder-only Transformer. It introduces a constrained beam search to explore multiple plausible futures while enforcing spatial adjacency, and develops mixed-resolution hierarchical maps to balance detail and computational efficiency. Across three real-world datasets, TrajLearn achieves up to ~40% improvements over strong baselines on standard metrics, with extensive analysis of horizon, input length, and map resolution. The approach offers a scalable, reproducible framework for geospatial trajectory forecasting with broad implications for autonomous navigation, urban planning, and mobility analytics, and points to future work in semantic enrichment and transfer learning.

Abstract

Trajectory prediction aims to estimate an entity's future path using its current position and historical movement data, benefiting fields like autonomous navigation, robotics, and human movement analytics. Deep learning approaches have become key in this area, utilizing large-scale trajectory datasets to model movement patterns, but face challenges in managing complex spatial dependencies and adapting to dynamic environments. To address these challenges, we introduce TrajLearn, a novel model for trajectory prediction that leverages generative modeling of higher-order mobility flows based on hexagonal spatial representation. TrajLearn predicts the next $k$ steps by integrating a customized beam search for exploring multiple potential paths while maintaining spatial continuity. We conducted a rigorous evaluation of TrajLearn, benchmarking it against leading state-of-the-art approaches and meaningful baselines. The results indicate that TrajLearn achieves significant performance gains, with improvements of up to ~40% across multiple real-world trajectory datasets. In addition, we evaluated different prediction horizons (i.e., various values of $k$), conducted resolution sensitivity analysis, and performed ablation studies to assess the impact of key model components. Furthermore, we developed a novel algorithm to generate mixed-resolution maps by hierarchically subdividing hexagonal regions into finer segments within a specified observation area. This approach supports selective detailing, applying finer resolution to areas of interest or high activity (e.g., urban centers) while using coarser resolution for less significant regions (e.g., rural areas), effectively reducing data storage requirements and computational overhead. We promote reproducibility and adaptability by offering complete code, data, and detailed documentation with flexible configuration options for various applications.

Figures (19)

• Figure 1: Illustrative example of the trajectory prediction problem using higher-order spatial representations (hexagons); ( a) there are two potential trajectories for the pedestrian, ( b) trajectories are represented on a hexagon-based tessellated map, ( c) given the historical data (red) and the current location (black), two trajectories are predicted (blue and orange), ( d) the actual trajectory followed.
• Figure 2: Construction of higher-order trajectory data.
• Figure 3: TrajLearn high-level architecture.
• Figure 4: Train with (bottom) and w/o (top) teacher forcing.
• Figure 5: Beam search example where $w=2$ and $k=3$.