Edinburgh Napier University

  Information about the motor performance, i.e. how well an activity is performed, is valuable information for a variety of novel applications in Activity Recognition (AR). Its assessment represents a significant challenge, as requirements depend on the specific application. We develop an approach to quantify one aspect that many domains share – the efficiency of motion – that has implications for signals from body-worn or pervasive sensors, as it influences the inherent complexity of the recorded multi-variate time-series. Based on the energy distribution in PCA we infer a single, normalised metric that is intimately linked to signal complexity and allows comparison of (subject-specific) time-series. We evaluate the approach on artificially distorted signals and apply it to a simple kitchen task to show its applicability to real-life data streams.