Automatic Labeling For Personalized IoT Wearable Monitoring

Abstract

Detecting human falls and locating injured workers are promising applications of Internet of things wearable monitoring for improving safety in industrial sectors. Using kinematic sensors and supervised learning, fall detection and localisation models can be built by learning using labeled input data. To reduce false alarms and accuracy degradation in such systems, the models need to be adapted and personalized to meet individual requirements since there are variations in performing the same activity by different people. Automatic labeling, being more efficient and less prone to error, is needed for online model training. In this paper, we propose an automatic labeling algorithm using K-means clustering for personalized falls detection online model training and an automatic labeling algorithm using descendingly ordered set of peak-trough magnitudes for personalized localisation online model training. Our experimental results showed that automatic labeling using maximum sum of cluster means produced a fall detection accuracy of up to 97%, and automatic labeling of peak-trough magnitudes resulted in a step counting accuracy of at least 96%.