NeuroVisionHRT
Home / Projects / Published research

Robust Neuromorphic/Event Vision-Based Fall Detection Across Datasets

A cross-dataset benchmark that converts established RGB fall datasets into events, compares nine temporal models, and tests generalisation on real event-camera data.

Expert Systems With Applications · Volume 325 · Article 132675 · 2026

Muhammad Zeeshan Masood, Muhammad Hamza Zafar, Syed Kumayl Raza Moosavi, Florenc Demrozi, Furqan Shaukat, and Filippo Sanfilippo

Project overview

The project tests a central simulation-to-real question: can models trained on event streams generated from RGB fall videos generalise to data captured by a physical event camera? Four RGB benchmarks are converted using the Metavision Simulator, nine deep temporal models are trained under a common protocol, and the resulting systems are evaluated on the real PAF dataset.

Core contributions

  • A public RGB-to-event fall-detection benchmark suite built from four established RGB datasets.
  • A controlled comparison of nine deep learning models under consistent training and evaluation protocols.
  • Cross-dataset testing on real event-camera data to quantify the simulation-to-real gap.
  • A reproducible framework for studying dataset diversity, robustness, and privacy-aware fall monitoring.

Reported evidence

98.65%best reported accuracy on synthetic event data
81.4%reported accuracy on real PAF event data
17.25%reported synthetic-to-real performance gap

TemporalMotionNet achieves the strongest synthetic-domain result but drops on the real event-camera benchmark. This makes the domain gap visible rather than hiding it behind high within-dataset scores.

Datasets and evaluation

  • CAUCA Fall, GMDCSA24, Fall Vision, and Multiple Cameras Fall Dataset converted from RGB video into event streams.
  • PAF used as the real event-camera target dataset for cross-domain evaluation.
  • Nine evaluated models: I3D, 3DCNN, ConvLSTM, MobileNetV3-LSTM, TCN, SwinTransformer, TimeSFormer, EfficientViT, and TemporalMotionNet.

Relevance to human-in-the-loop robotics

Fall detection is directly relevant to healthcare, assistive robotics, smart homes, and human safety monitoring. The study is especially valuable because it stress-tests real-world generalisation instead of relying only on synthetic or within-dataset performance.

Open official article →Open benchmark repository →