Movement recognition from wearable sensors data: power-aware evolutionary training for template matching and data annotation recovery methods

Human activities recognition finds numerous applications for example in sport training, patient rehabilitation, gait analysis and surgical skills evaluation. Wearable sensing and Template Matching Methods (TMMs) offer significant advantages compared to manual assessment methods as well as to more cumbersome camera-based setups and other machine learning (ML) algorithms. TMMs require less data for training than other ML methods, they are low-power and therefore suitable for integration on wearable sensor. They compute a sample-by-sample distance between two time series. When applied to gestures sensors data, this even enables a richer and more movement-specific assessment and feedback. However, TMMs lack of a standard training procedure. In this thesis, we introduce an innovative evolutionary training algorithm for TMMthat not only can maximize recognition performance, but it can also prefer power-minimisation by reducing the TMM’s computational cost with a configurable trade-off. We exhibit that a reduction is even possible without sacrificing recognition performance by exploiting the long-established concept of “time warping”. We demonstrate that our method is suitable for a wide variety of raw data as well as processed, fused and encoded sensor data. We present a new original multi-modal, multi-user dataset of beach volleyball movements that allowed to evaluate our training methods on a real-case of sport training actions. Moreover, the collection of this dataset helped to generate a set of guidelines for the collection of movement data in the wild, using wearable sensors. We introduce a 3D human model that can be animated through inertial wearable sensors data for troubleshooting, movement analysis and privacy-safe annotation of human activities. Finally, through a case study on a dataset of drinking actions, we demonstrate how TMM can improve the quality of a badly annotated but highly valuable dataset.