Researchers unveil Gated-CNN for smartwatch fall detection, beating Transformers with 97% F1
Sigmoid gating outperforms self-attention on wearables, achieving zero missed falls on Pixel Watch 3.
A new paper from researchers including Sana Alamgeer and Anne H. H. Ngu challenges the dominance of Transformer-based attention in wearable fall detection. Their proposed model, Gated-CNN, uses independent one-dimensional convolutions on accelerometer and gyroscope streams, followed by a sigmoid gating module that selectively suppresses background noise while amplifying fall-specific features. This avoids the quadratic overhead of self-attention, which distributes weights across all time steps and misses brief impact signatures.
Evaluated offline across five wrist-mounted IMU datasets (SmartFallMM, WEDA-Fall, FallAllD, UMAFall, UP-Fall), Gated-CNN achieved average F1-scores of 90-93%, outperforming Transformer baselines. For real-world testing, the team deployed the model on a Google Pixel Watch 3 with 12 participants, achieving an average F1 of 97% and accuracy of 98% with zero missed falls. The results demonstrate that sigmoid gating offers a structurally aligned, computationally efficient alternative to attention for commodity smartwatches.
- Gated-CNN replaces self-attention with sigmoid gating, reducing computational overhead from quadratic to linear.
- Achieved 97% F1 and 98% accuracy on a Google Pixel Watch 3 across 12 participants with zero missed falls.
- Outperformed Transformer baselines on five public datasets (90-93% F1) while being lightweight enough for real-time wearable deployment.
Why It Matters
A more efficient fall detection model could save lives by enabling accurate, always-on monitoring on existing smartwatches.