Vision-Based Fall Prediction With Spatial–Temporal Graph Convolutional Networks
- 29 December 2025
- journal article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (6) , 11171-11183
Abstract
Falls pose a significant risk to the lives and independence of elderly individuals, raising widespread concern in public healthcare. Timely and accurate prediction of fall events prior to impact is essential for enabling effective interventions and minimizing potential injuries. In this paper, we propose a novel Unified Graph Convolutional Network (Uni-GCN) that simultaneously performs skeleton-based pose prediction and fall detection within a multi-task learning framework. The proposed model captures both spatial and temporal dependencies in human motion through a shared graph-based encoder, followed by two task-specific decoders for future pose prediction and fall classification. A unified loss function is introduced to jointly train both tasks, combining prediction loss and classification loss in a mutually beneficial manner. Additionally, we employ a two-stage transfer learning approach to address the challenge of limited data in fall datasets, by pretraining the model on large-scale datasets and fine-tuning it on fall-specific data. Experimental results on public fall datasets Le2i, URFD, and UP-Fall demonstrate the effectiveness of our method, achieving state-of-the-art accuracies of 99.15%, 99.42%, and 99.96%, respectively. Notably, our system is capable of predicting fall events approximately 480 milliseconds before impact, enabling proactive response. In addition, we deploy our model on an NVIDIA Jetson Orin Nano Super edge device, validating its real-time performance and feasibility for deployment in privacy-sensitive and bandwidth-constrained environments.Keywords
Funding Information
- National Natural Science Foundation of China (62372163)
- science and technology innovation Program of Hunan Province (2024RC1033)
This publication has 50 references indexed in Scilit:
- A spatio-temporal graph neural network for fall prediction with inertial sensorsKnowledge-Based Systems, 2024
- Transformer-based fall detection in videosEngineering Applications of Artificial Intelligence, 2024
- OneFall-GAN: A one-class GAN framework applied to fall detectionEngineering Science and Technology, an International Journal, 2022
- Vision-based human fall detection systems using deep learning: A reviewComputers in Biology and Medicine, 2022
- Fall Detection Using Multimodal Data2022
- Elderly fall detection based on multi-stream deep convolutional networksMultimedia Tools and Applications, 2020
- Spatio-temporal fall event detection in complex scenes using attention guided LSTMPattern Recognition Letters, 2020
- DeepFall: Non-Invasive Fall Detection with Deep Spatio-Temporal Convolutional AutoencodersJournal of Healthcare Informatics Research, 2019
- Human fall detection using machine vision techniques on RGB–D imagesBiomedical Signal Processing and Control, 2018
- Human fall detection on embedded platform using depth maps and wireless accelerometerComputer Methods and Programs in Biomedicine, 2014