Dynamic Workout Detection in Smart Watch Using Artificial Neural Network

Recent advancements in smart technology like smart watches make it easy for the user to keep track of their daily activities. Nowadays, smart watches are capable of efficiently monitoring routine exercises and sport activities. But the current solution of Human Activity Recognition (HAR) and Sports Activity Recognition (SAR) uses multiple sensors like accelerometer, gyroscope, and magnetometer, which is battery intensive. Further most of the HAR and SAR solutions have not been subjected to real time testing. To overcome this problem, we propose a method for SAR by making use of data from a single sensor namely accelerometer. In this work SAR covers 4 activities (badminton, table tennis, squash and football), and together we denote them as Dynamic Workout (DW). The proposed method makes use of a light weight Artificial Neural Network (ANN) model with only 2301 trainable parameters. The model is deployed on Samsung Galaxy Watch3 for real time testing. The proposed method achieved True Positive rate (TPR) of 89.5% and True Negative Rate (TNR) of 92.3%. The model achieved 100% accuracy in real time testing.