Sports activity recognition with UWB and inertial sensors using deep learning approach

Nowadays the sensor-based human activity recognition is a key issue in the field of the physical exercises’ recognition. The purpose of this paper is to recognise exercise sports, such as squats, pull-ups and dips, using a dataset based on three ultra-wideband (UWB) sensors with additional inertial data and a deep learning approach in order to achieve an appropriate rate of the recognition of sports exercises with reduced computational effort. Firstly, a dataset, containing 273 samples of dips, 215 samples of pull-ups and 956 samples of squats, plus an additional 2024 samples of the input signals acquired during breaks, was created. Next, an optimal set of hyperparameters of Convolutional Neural Network (CNN) architecture in order to balance the accuracy rate and computational effort was achieved. Finally, it was discovered that acceleration signals have the highest energy and therefore, a comparison of the accuracy rate between CNN, using all signals and CNN, using acceleration only, was carried out (97.5% vs. 97.7%). Moreover, much less computational effort, expressed in number of multiplications, (1.8e4 vs. 1.2e5) was achieved. The practical usefulness is presented, by facilitating the implementation of the presented UWB sensor-based system for recognising such physical exercises as squats, pull-ups and dips on smartphones and commonly available wearable sensor devices.