We introduce BlinkWise,
a minimalist wearable add-on that enables detailed blink dynamics
tracking at millisecond resolution
on everyday eyewear.
BlinkWise predicts eye openness, a full
descriptor of blink dynamics.
* Non-sped-up parts are in 480 FPS to reduce total playback time.
** Eye video appears blurry due to cropping from 720p footage.
BlinkWise unlocks diverse blink-centered applications ...
Drowsiness Monitoring
BlinkWise detects a significant correlation between
prolonged blinks and drowsiness.
Workload Assessment
BlinkWise observes significant
reduced blink rate under higher workload.
Health Management
BlinkWise detects partial blinks,
promising for dry-eye disease management.
BlinkWise offers a novel hardware-software solution.
The radio-frequency (RF) sensor detects eyelid movement safely,
sensitively, and with fine granularity.
All processing completes on the edge MCU, supporting private,
real-time applications via 3 novel techniques:
CNN Recurrentization
Executing CNN models as recurrent networks for reduced memory and latency.
Quantization-Aware Normalization
Adapting to variations among users and maintaining performance.
Efficient Two-Stage Detection
Coarse-to-fine prediction of eye openness, reducing computation and latency.
10.67 ✕ reduction in memory; 76.3% reduction in latency.
Acknowledgments
We are grateful to our shepherd and anonymous reviewers for their comments and feedback.
We also thank the members of the WAVES Lab and the PRECISE Center at Penn for their valuable discussions and feedback.
This work was supported by ASSET Center Seed Grants in Trustworthy AI Research for Medicine.
