Deep Recurrent Neural Networks for Human Activity Recognition During Skiing

Abstract

In recent years, deep learning has been successfully applied to an increasing number of research areas. One of those areas is human activity recognition. Most published papers focus on a comparison of different deep learning models, using publicly available benchmark datasets. This article focuses on identifying specific activity—skiing activity. For this purpose, a database containing information from the three inertial body sensors, placed on skier’s chest and on both skis, was created. This database contains synchronized data, from an accelerometer, gyroscope or barometer. Then, two deep models based on the Long Short-Term Memory units, were created and compared. First is a unidirectional neural network which can remember information from the past, second is a bidirectional neural network, which can memorize information from both the past and the future. Both models were tested for different window sizes and the number of hidden layers and the number of units on the layer. These models can be used in the alpine skiing and biathlon training support system.