MM-DOS: A Novel Dataset Of Workout Activities

Ahmed Sharshar, Ahmed Fayez, Ahmed Abo Eitta, Walid Gomaa

Abstract: Human Activity Recognition (HAR) is now one of the most widespread fields in research and application, especially within the area of artificial intelligence. A wide variety of applications are based on HAR, including healthcare, entertainment, and sports performance analysis. Such systems can be used as a reliable source of performance feedback for sports. Hence, coaching and the athletes' performance enhancement became the main application for HAR. Considering a data-oriented approach, the need for quality datasets increased to cover these springing needs. In this paper, we present a new multi-modal dataset MMDOS; the dataset contains a variety of data forms from different sensors, including Red-Green-Blue (RGB) videos, inertial motion data, depth, and thermal data, all synchronized with regard to the activities/exercises performed. The dataset consists of four workout exercises: free squats, shoulder presses, push-ups, and lunges. The exercises are performed by 50 participants indoors at a fitness center. Professional trainers label the data with the type of activity and the mistakes that the participant made. In this paper, we offer further details about the nature of the data and how the dataset was collected and post-processed. Finally, the paper shows some analysis performed to evaluate the dataset quality to recognize the different exercises. We reached accuracy up to 96% for RGB videos, 90% for depth videos, and 91% for IMU data.