An Approach to Sport Activities Recognition Based on an Inertial Sensor and Deep Learning

In recent years, due to changing the human lifestyle, the number of sport trainers has been increased. The conventional classifiers as Naive Bayes (NB), Decision Trees (DT) and Convolutional Neural Networks (CNNs) can be used in this domain to recognize and count sports activities of subjects and provide them qualified feedback. This paper uses literature studies and selected sport activities, namely squats, pull-ups and dips as the dataset based on three UWB sensors with additional inertial data, which contains the reduced data set consisting of 17 training sets and next for CNN training the 1444 samples describing exercises and 2024 samples with breaks, which were grouped in the ratio 70:15:15. The recognition accuracy of the NB and DT were 89.4 and 92.9 accordingly. Next, the extensive performance analysis of the CNN based on experiments for different kernel sizes, different number of filters for single and dual layer networks was carried out. Moreover, the innovative model for sport activities recognition in the form the combination of several networks forming Ensemble Neural Network (ENN) was created. The accuracy was at the level 94.81 of CNN and exceeded 95% of ENN. The proposed prototype of the measurement system and data acquisition platform for sport activities recognition was highlighted as the great potential in the privacy-training sport system