Lightweight Change-Detection Algorithm Reduces HAR Energy by 67%
A novel lightweight algorithm for change detection in wearable devices has been developed, achieving a reduction in computational demands by more than 67%. This non-parametric method employs dynamic template matching and operates without the need for offline training, processing at around 16k FLOPs per step. It demonstrates 98% sensitivity on the UCA-EHAR dataset, ensuring that all genuine activity changes are detected, while maintaining 75% specificity to minimize unnecessary HAR activations. The algorithm was tested using data from smart glasses, smartwatches, and smartphones, requiring only a short calibration phase specific to each device. This innovative approach effectively mitigates energy waste from ongoing classification during extended periods of stable activity, making it ideal for ultra-low-power applications.
Key facts
- Algorithm reduces computational load by over 67% in realistic monitoring settings.
- Runs at approximately 16k FLOPs per step.
- Requires no offline training and no prior definition of target activity classes.
- Achieves 98% sensitivity on UCA-EHAR dataset.
- Achieves 75% specificity.
- Evaluated on smart glasses, smartwatch, and smartphone data.
- Requires only a brief device-specific calibration phase.
- Based on dynamic template matching.
Entities
—