MicroFallNet: A lightweight model for real-time fall detection on smart wristbands

Falls are a major public health concern for the aging population, leading to significant injuries, loss of independence, and increased healthcare costs. While wearable devices present promising solutions, existing algorithms are often hindered by the limitations of microcontroller units (MCU) in terms of computational power, memory, and energy consumption. To overcome these challenges, we introduce MicroFallNet, a lightweight convolutional neural network designed for accurate and efficient fall detection. MicroFallNet features a novel FireModel architecture, incorporating Squeeze and Expand layers to optimize computational efficiency and enhance feature extraction. The proposed algorithm demonstrates superior performance on the UMAFALL and FallAllD datasets, achieving geometric mean accuracies of 97.91 % and 97.86 %, respectively, significantly surpassing traditional methods. Additionally, MicroFallNet showcases excellent deployment efficiency across various microcontrollers, particularly excelling on the ESP32 smart wristband platform, where it achieves an inference time of just 30.3 milliseconds. This capability makes MicroFallNet ideally suited for real-time fall detection applications, advancing the development of wearable devices for the elderly and contributing substantially to the field of smart health monitoring. Our code will be publicly available at https://github.com/qwer12330/MicroFallNet-A-Lightweight-Model-for-Real-Time-Fall-Detection-on-Smart-Wristbands-Using-Sm.