Idea
On-device smartwatch system for real-time, privacy-preserving human activity recognition benefiting fitness and health monitoring users
Research Paper
Core Innovation
This paper introduces WatchHAR, a unified end-to-end trainable model that integrates sensor preprocessing with inference directly on smartwatches. It achieves significantly faster processing speeds and high accuracy across many activity classes compared to prior models. This enables real-time, privacy-preserving activity recognition without relying on external devices or cloud services.
Market Size (TAM)
$10–20B TAM, $2–5B SAM; assumption: growing wearable device market and increasing demand for on-device AI in health and fitness.
Potential Customers & Pain Points
- Wearable Device Manufacturers Needing Efficient On-device Activity Recognition
- Fitness App Developers Seeking Accurate Real-time Data
- Healthcare Providers Requiring Continuous Patient Monitoring
- Privacy-conscious Users Avoiding Cloud Data Processing
Business Model
Licensing the WatchHAR technology to wearable manufacturers and fitness app developers; offering SDKs and APIs for integration.
Competitive Landscape
- Google Fit
- Apple Activity Recognition
- Fitbit SDK
Implementation Challenges
- Limited smartwatch hardware resources
- User adoption of new activity tracking apps
- Integration with diverse smartwatch platforms
Validation Strategy
- Develop prototype smartwatch app demonstrating real-time activity recognition
- Conduct user studies measuring accuracy and latency in real-world scenarios
- Partner with device makers for pilot integration and feedback
Research Paper Overview
WatchHAR: Real-time On-device Human Activity Recognition System for Smartwatches
Summary
WatchHAR is an audio and inertial sensor-based human activity recognition system that runs entirely on smartwatches, enabling privacy-preserving and low-latency activity tracking. It features a novel end-to-end trainable architecture that unifies sensor preprocessing and inference, achieving 5x faster processing and over 90% accuracy across 25+ activity classes. WatchHAR outperforms state-of-the-art models with processing times under 12 ms, making smartwatches standalone, minimally invasive continuous activity trackers.