Evaluation of Deep Learning Architectures on Synthetic LiDAR Sequences for Fall Detection

Abstract

Falls represent a major health risk for the elderly population, often leading to serious injuries and loss of independence. Despite advances in sensing and monitoring technologies, existing fall-detection systems continue to face challenges related to accuracy, privacy, and user compliance. To address these limitations, this study presents a LiDAR-based deep learning framework specifically designed for privacy-preserving, fall detection with stable temporal modeling under simulated LiDAR conditions in ambient-assisted living (AAL) environments. As publicly available LiDAR datasets for human fall activities are scarce and ethically difficult to collect, a synthetically dataset was generated using Blender and Mixamo. This dataset consists of 1,000 multi-frame LiDAR sequences with 60 frames per activity, and captures diverse fall and non-fall motions. The study evaluates four temporal deep learning models including LSTM, GRU, CNN1D and Transformer Lite to analyze spatio temporal motion patterns from 3D LiDAR point sequences and identify the most suitable approach for accurate fall detection. Among these, the LSTM model achieved the best performance, reaching 93.3% accuracy and 0.979 AUC, confirming its ability to capture long term temporal dependencies. These findings highlight the potential and feasibility of LiDAR based temporal modeling as a reproducible, non intrusive and ethically sound solution for continuous elderly monitoring. Future work will focus on validating the framework with real LiDAR sensors and implementing adaptive edge deployment for real-time fall detection in smart home environments.