A parallel CNN architecture for sport activity recognition based on minimal movement data

Abstract

Novel Human Activity Recognition (HAR) methodologies, which are built upon learning algorithms and employ ubiquitous sensors, have achieved remarkable precision in the identification of sports activities. Such progress benefits all age groups of humanity, and in the future, AI will be used to address difficult problems in scientific research. A novel approach is introduced in this article to utilize motion sensor data in order to categorize and distinguish various categories of sports activities. This is achieved through the parallel implementation of Convolutional Neural Networks (CNN) and machine learning methods. The methodology being proposed consists of four fundamental phases. The preliminary stage consists of sensor data preprocessing and normalization. In the subsequent phase, the signal characteristics are characterized using Discrete Wavelet Transform (DWT) and Short-Time Fourier Transform (STFT). Both are utilized in order to lay the foundation for the two CNN models that follow. Every signal representation is utilized as an input for a Separated convolutional model, which constructs the motion features using the sports motion information. When the two sets of motion pointsets from each CNN are merged, the situation becomes more balanced, and the Random Forest classification model is able to identify the type of sports activity by detecting and classifying the features. Using the DSADS dataset, the effectiveness of the proposed method in classifying a variety of sports activities was evaluated. A mean precision of 99.61% was achieved in this particular domain.