Activity Recognition Based on Accelerometer Data with Enhanced ROCKET Algorithm

Automated recognition of activities based on ac-celerometer data is an important task due to its applications in various domains, such as sport, medicine, healthcare and mental health. State-of-the-art solutions for this task are based on machine learning. In this paper, we consider activity recognition as a multivariate time series classification task. Our solution is based on Random Convolutional Kernel Transform (ROCKET), which is a recent time series classifier. We enhance ROCKET by using dynamic convolution instead of standard convolution. Dynamic convolution was originally defined for univariate time series, therefore, we extend this notion to multivariate time series so that it can be used to classify accelerometer data. We perform experiments on publicly available real-world accelerometer data related to the detection of various activities, including epileptic seizures. The experimental results show that the proposed enhancements of the ROCKET algorithm reduce the average classification error by almost an order of magnitude, in particular from 1.74% to as low as 0.18%. This may be considered important in certain applications in medicine or healthcare. Moreover, the observed improvement is statistically significant according to t-test at a significance level of 0.01. Our implementation is publicly available at https://github.com/kr7IROCKET-DTW.