Artificial Intelligence Design for Real-time Fall Detection

Falls represent a serious senior care issue worldwide, especially among individuals aged 65 and above. The World Health Organization (WHO) reports approximately 646,000 related fatalities annually due to falls. This paper addresses the limitations of existing fall detection systems, characterized by high false alarm rates and constraints in deployments. We introducing an edge artificial intelligent (AI) design for real-time fall detection. The compact $25 \times 32 \times 4 \mathrm{~mm}$ device, leveraging machine learning and Field Programmable Gate Arrays (FPGA). It combines a threshold-based method with a neural network model, dynamically activated through a threshold comparator, thus enhancing adaptability while minimizing redundant power consumption. Our research adopts a dual evaluation strategy. Initially, we benchmark the proposed model against existing frameworks using publicly available datasets focused on tri-axial sensor readings. Subsequently, we validate the device using simulated fall data from 10 healthy subjects, achieving performance metrics with an average accuracy, sensitivity, and specificity surpassing 98.3% and notably reducing the false positive rate to $\mathbf{1. 3 \%}$. Furthermore, the device incorporates a hardware AI accelerator, evidenced by experimental results demonstrating a throughput rate of 16,995 operations per second, coupled with an overall system power consumption of 109.1 mW. Integration of an RF energy harvesting module ensures uninterrupted functionality without sole reliance on batteries. To safeguard user privacy during data transmission and storage, a customized homomorphic encryption scheme is employed.