A Lightweight Federated Prediction Approach for Urban VRU Movement Understanding in Autonomous Driving

Lakshmi Narayana, I; Vamsi, T M N · 2026 · DOAJ

DOI: 10.5935/jetia.v12i57.3066

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 critical need for accurate short-term trajectory prediction of vulnerable road users (VRUs), such as pedestrians and cyclists, in autonomous driving systems. The authors identify that existing centralized learning models fail to scale effectively due to strict privacy regulations, data ownership issues, and poor cross-domain generalization caused by heterogeneous urban environments. To resolve these challenges, the study proposes a lightweight federated prediction framework that enables collaborative model training without sharing raw sensor data, thereby ensuring privacy compliance and improving robustness across diverse traffic scenarios. The proposed methodology integrates onboard perception, tracking, and federated learning into a unified pipeline. First, the YOLOv11 detection model identifies VRUs in video frames, followed by the SORT algorithm, which uses Kalman filtering and Hungarian assignment to maintain consistent object identities across frames. These detections are converted into spatio-temporal trajectory sequences, incorporating social interaction neighborhoods to capture group dynamics. The core prediction module utilizes a Social-LSTM network, which models temporal motion patterns and interpersonal influences through social pooling. To facilitate distributed training, the system employs the FedProx algorithm for global aggregation. Unlike standard FedAvg, FedProx introduces a proximal regularization term that constrains local model updates, stabilizing convergence under non-independent and identically distributed (non-IID) data conditions typical of varied urban landscapes. Experiments were conducted using four benchmark datasets: ETH, UCY, SDD, and NuScenes, which were preprocessed to ensure uniform sampling rates and coordinate systems. The results demonstrate that the federated approach significantly reduces domain drift and enhances stability compared to centralized models. Specifically, the system achieved improved Average Displacement Error (ADE) and Final Displacement Error (FDE) scores across different scenes. The use of FedProx proved essential in mitigating gradient divergence caused by regional variations in VRU behavior, leading to more consistent prediction accuracy. Additionally, the framework maintained high detection precision and recall while adhering to privacy constraints by transmitting only gradient updates rather than raw perception data. The study concludes that this federated spatio-temporal learning system offers a scalable, privacy-preserving, and deployment-ready solution for autonomous vehicle navigation. By overcoming the limitations of centralized data aggregation and addressing cross-domain variability, the approach provides a practical advancement for safer autonomous driving. The findings suggest that integrating lightweight tracking with federated Social-LSTM models can effectively handle the complexity of mixed-traffic environments, supporting more reliable risk assessment and decision-making in real-world applications.

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