Accurate and energy-efficient fall detection on commodity smartphones is crucial for long-term monitoring of older adults and high-risk populations. However, most existing methods emphasize window-level accuracy and F1 score, while largely overlooking probabilistic calibration and event-level behavior that are critical for real-world deployment. We propose LFNET, a lightweight multi-channel convolutional network for fall detection using a single waist-mounted smartphone accelerometer. LFNET integrates multiscale depthwise separable convolutions with efficient channel attention, and jointly leverages tri-axial accelerations and their magnitude to capture both posture dynamics and impact intensity. On the UniMiB-SHAR dataset, LFNET attains competitive or superior F1 and AUPRC on both the binary fall detection task (AF2) and the 17-class activity recognition task (AF17), while using only a fraction of the parameters and FLOPs of larger CNN and Transformer baselines. To better match deployment requirements, we apply temperature scaling and sensitivity-constrained threshold selection to calibrate predicted probabilities and choose operating points tailored to fall-risk preferences. Using semi-synthetic continuous streams constructed from UniMiB-SHAR segments, we further design an event-level post-processing pipeline combining K-consecutive positives, hysteresis thresholds, gap merging, and refractory periods. Event-level experiments show that, under a fixed configuration, LFNET improves event sensitivity from 90.1% to 96.4% over a CNN baseline, while reducing false alarms from 1.4 to 0.2 FA/h and shortening median detection delay from 1.4 s to 0.8 s. These results demonstrate that the proposed framework not only performs well on standard window-level metrics, but also offers favorable false-alarm–delay trade-offs for smartphone-based, long-term fall detection in simulated monitoring scenarios.
Yang et al. (Wed,) studied this question.