TrainingPal: An Algorithm for Recognition and Counting Popular Exercises Using Smartphone Sensors

Nowadays, most of the smartphones come with 3D-accelerometer, magnetometer, and gyroscope. There is a growing interest in utilizing these sensors for monitoring and tracking exercises. In this paper, we proposed a framework, called TrainingPal, for recognizing five types of cardio exercises (i.e. walking, running, using elliptical machine, rowing, and jumping jack) and five types of resistance exercises (i.e. squat, lunge, situp, push-up, and bench dip). The TrainingPal is also capable of counting number of repetition of each exercise. To train and test the TrainingPal, data was collected by utilizing the built-in accelerometer, magnetometer and gyroscope of Samsung Galaxy S7 edge. Using an armband, the smartphone was attached to the outer side of arm approximately 10 to 12 cm below the shoulder. Leave-one-subject-out cross validation was used to avoid overfitting of the TrainingPal to the study participants. For recognition of different types of exercises, an overall accuracy of 91.71% was achieved. The accuracy was the highest for the recognition of push-ups (100%) and the lowest for the recognition of bench dips (60.33%). For counting the repetitions of the exercises, the TrainingPal achieved an accuracy above 90% for all types of exercises. In conclusion, the proposed framework can be used for tracking and recognizing the popular exercises included in this study. The framework could be potentially extended to other types of exercises and the data collected using other wearable devices.