A Cascaded CNN-ResNet Architecture for Computation-Aware Fall Detection on TinyML Wearable Devices

Abstract

Falls among older adults are a major public health concern, yet accurate fall detection on resource-constrained wearable devices remains challenging because high-performance deep models often require continuous computation. This study proposes a computation-aware cascaded CNN-ResNet architecture for TinyML-based wearable fall detection. A lightweight CNN processes four accelerometer-derived channels, namely x, y, z, and root-mean-square (RMS), to perform initial classification and extract 32-dimensional embeddings. These embeddings are stored in a first-in, first-out (FIFO) buffer. When the CNN confidence for Activities of Daily Living (ADL) classification falls below a threshold of 0.995, a compact ResNet is activated to analyze six consecutive embeddings for refined spatiotemporal classification. The system was evaluated on the KFall dataset using leave-one-subject-out (LOSO) cross-validation across 32 subjects. It achieved 98.97% ± 0.39% accuracy, 98.48% ± 0.91% precision, 97.96% ± 0.95% recall, and 98.21% ± 0.62% F1-score. On-device benchmarking on ESP32-S3 and STM32F407 MCUs confirmed a combined quantized model size of 0.46 MB and average latencies of 12.3–17.3 ms and 9.6–14.1 ms, respectively, under a conservative daily-activity scenario with 3–5% ResNet invocation. Even under full ResNet activation, the maximum latency remained below 252 ms, supporting real-time on-device inference within the 500 ms sliding-window interval. These results indicate that the proposed cascade provides an effective trade-off between detection accuracy and average computational cost for MCU-based wearable fall detection.