Citation

Ahmed, Farhana and Khaliluzzaman, Md. and Ghazali, Anith Khairunnisa and Besar, Rosli and Abdul Aziz, Nor Hidayati Optimized Fall Detection Using CNN and Quantized Xception Model for Real-Time Applications. HCIS.

Text HCIS_Paper_R2_.pdf - Published Version Restricted to Repository staff only Download (1MB)

Abstract

According to the World Health Organization (WHO), falls are the leading cause of injury-related deaths worldwide, making efficient fall detection systems a priority. Current fall detection research relies on computationally expensive structures like 3D CNNs (e.g., ResNet3D) and CNN-LSTMs that are extremely accurate but suffer from high latency (>200ms per inference) and high memory costs (>150MB), limiting realworld deployment. To bridge this gap, we introduce a novel lightweight fall detection system that outperforms deep learning techniques using the UR-Fall Dataset by balancing the accuracy, efficiency, and deployment feasibility. Our approach integrates three key steps: (1) A CNN with a self-attention mechanism that extracts critical spatial features with 99.73% accuracy but remains computationally heavy. (2) The Xception Model, a lightweight customized transfer learning model with depth-wise separable convolutions, which is fast and efficient with 100% accuracy, but still too large for practical deployment. Finally, this paper sheds light on a novel contribution, (3) the Quantized Xception Model, which not only addresses the gaps of existing fall detection models but also reduces the model size from 80.89 to 20.78 MB (a 74% reduction over standard Xception), improving the inference speed from 108 to 70 ms per image, while maintaining 99.74% accuracy. Compared to prior quantized models (e.g., TensorRT optimized EfficientNet) that trade interpretability, our model balances efficiency with interpretability via Grad-CAM visualizations and achieves 100% confidence for fall/non-fall detection. This real-time, lightweight model is scalable for research and potential industrial applications in hospitals, home automation, and security surveillance.

Item Type:	Article
Uncontrolled Keywords:	Deep Learning
Subjects:	Q Science > Q Science (General) > Q300-390 Cybernetics
Divisions:	Faculty of Engineering and Technology (FET)
Depositing User:	Ms Rosnani Abd Wahab
Date Deposited:	02 Mar 2026 00:39
Last Modified:	02 Mar 2026 00:39
URII:	http://shdl.mmu.edu.my/id/eprint/15382

Downloads

Edit (login required)