Real-Time Fall Detection for Care Recipient in Ambient Assisted Living Using Skeleton-Based LSTM on Edge Devices
Abstract
Falls represent a critical biomedical challenge for aging populations, often leading to severe injury or longlife scenarios if assistance is delayed. While fall prevention strategies focus on gait analysis, effective post-impact fall detection is essential for ensuring timely emergency escalation. Ambient vision-based systems offer a compelling alternative to wearable sensors by enabling reliable, passive monitoring without compliance burdens; however, widespread deployment faces significant hurdles regarding privacy, scalability, and system reliability. This paper proposes SOLA: Skeleton-based Optimized LSTM for Action recognition, a resource-efficient, privacy-centric pipeline designed for local execution in multicamera environments. Rather than relying on cloud processing or expensive dedicated workstations per room, SOLA optimizes the trade-off between computational demand and classification precision to enable robust operation on cost-effective local hardware. By synergizing skeletal pose estimation with an evolutionary-optimized LSTM network, the architecture minimizes false alarms, a crucial requirement for dependable emergency alerting. Furthermore, the lightweight design ensures that the system remains operational during network outages or partial infrastructure failures, prioritizing safety in worst-case scenarios. Experimental results on the GMDCSA-24 dataset demonstrate that SOLA achieves superior inference speed and low false-positive rates, offering a scalable solution for privacypreserving, decentralized home monitoring.