Motion Prediction of Vulnerable Road Users at Signalized Intersections Using Lightweight Feed-Forward Neural Networks

Jekl, Bence and Fényes, Dániel and Németh, Balázs (2026) Motion Prediction of Vulnerable Road Users at Signalized Intersections Using Lightweight Feed-Forward Neural Networks. In: IEEE 24th International Symposium on Intelligent Systems and Informatics, https://conf.uni-obuda.hu/sisy2026/, Pula, Croatia. (In Press)

Preview

Text SISY_2026_paper_31.pdf Download (1MB) | Preview

Abstract

The accurate prediction of Vulnerable Road Users (VRUs) is essential for the safe and efficient operation of autonomous vehicles in urban environments. This paper proposes a lightweight Feed-Forward Neural Network (FFNN) for multi-step acceleration prediction of bicycles and e-scooters at signalized intersections using the IMPTC dataset. A frequency-domain analysis of VRU velocity profiles is performed to identify the dominant motion time scale, which guides the selection of the history window and prediction horizon. The model uses motion history and traffic light state information without requiring complex scene representations. Quantitative evaluation demonstrates that the proposed FFNN outperforms zero-acceleration and constant-acceleration baselines in both road and sidewalk scenarios. Cross-scenario evaluation confirms that scenario-specific training improves performance. The model contains only 1803 trainable parameters and achieves an inference time of approximately 0.03 ms per sample on a standard CPU, demonstrating very high computational efficiency and suitability for real-time automotive applications.

Actions (login required)

View Item