A Location-Velocity-Temporal Attention LSTM Model for Pedestrian Trajectory Prediction

Xue, Hao; Huynh, Du Q.; Reynolds, Mark · 2020 · DOAJ

DOI: 10.1109/ACCESS.2020.2977747

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of pedestrian trajectory prediction, a critical task for applications such as autonomous vehicles and location-based services. While many existing methods rely on rich contextual information—such as neighboring pedestrians, static obstacles, or background scene features—these approaches often suffer from high computational costs and poor generalizability to unseen scenes. The authors propose a method that predicts future trajectories using only the observed trajectory of the person of interest, without requiring external context. This approach is motivated by the need for simpler, more efficient models suitable for resource-constrained environments like Advanced Driver Assistance Systems. The proposed model, named LVTA (Location-Velocity-Temporal Attention LSTM), utilizes a dual-LSTM architecture to process location and velocity embeddings in parallel. The core innovation involves two specific mechanisms. First, temporal attention mechanisms are applied to the hidden state vectors of both LSTM layers to weight the relationships between the observed trajectory history and the prediction phase, inspired by machine translation techniques. Second, a "tweak module" containing a location-velocity attention layer refines the predicted coordinates at each time step before they are passed to the next iteration. This module fuses location and velocity information to improve prediction accuracy without relying on pooling layers for social or scene context. Experiments were conducted on three benchmark datasets: the Central Station dataset, the ETH/UCY dataset, and the Edinburgh dataset. The LVTA model was compared against eleven existing trajectory prediction methods. On the Central Station dataset, LVTA achieved an Average Displacement Error (ADE) of 9.19 pixels and a Final Displacement Error (FDE) of 17.28 pixels. On the ETH/UCY datasets, it attained an ADE of 0.46 meters and an FDE of 0.92 meters. The study also included ablation studies to verify the contribution of the temporal attention and tweak modules, as well as hyperparameter tuning for dropout rates and hidden dimensions. The results demonstrate that LVTA achieves competitive, state-of-the-art performance despite its lack of contextual inputs. Crucially, the model exhibits strong generalizability, performing well on new, unseen scenes without the need for retraining. The authors conclude that incorporating temporal attention and the location-velocity tweak module significantly improves prediction accuracy and robustness compared to their previous work and other context-free methods. This confirms that high-performance trajectory prediction is achievable using only individual trajectory data, offering a viable alternative to complex context-dependent models.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• anticipation