"Physics" vs "Brain": Challenge of labeling wearable inertial data for step detection for Artificial Intelligence

Abstract

Data-driven methods have attracted the research community from all sectors including positioning-based applications. However, the performances of the AI-based methods depend strongly on the quality of the data. With the fast development of powerful hardware, collecting, storing and training huge databases are not problematic anymore. The true bottleneck to AI is rather getting high-quality labeling of the data, especially for supervised learning. This paper aims at discussing the most suitable and efficient way to label the step instants of wearables, between the choices of using physical approaches and the pattern interpretation approach. Physical approaches refer to using highly accurate foot-mounted equipment to get the step instants then project them on the related body parts. While the pattern interpretation approach relies directly on the signal signatures interpreted with the help of human gait knowledge. It is referred to as the "brain" approach. Two machine learning-based step prediction models are trained with respectively the "physic" and "brain" labeling approach. The performance assessment shows that the step prediction model trained with brain labeling has a true positive detection rate around 85.9% - 95.7% with almost no overdetection while the model trained with physical labeling can only reach 54.7% of true positive rate with a high overdetection rate (around 36.7%).