Understanding activity segmentation for multi-sport competitions

Despite the advances in activity detection, their applications in the sports domain are limited. Athletic environments are fast and challenging. Athletes often perform more than one activity in a single workout, especially if they are training for a multi-sport competition, such as a triathlon. These competitions require an athlete to transition from one activity to another quickly. Current logging applications require the user to select the activity they are about to perform, and start and stop the timer for each activity. This could increase the athlete's transition time, and it would not give the athlete an estimate of how much time they spent in transition between the activities. This paper explores activity segmentation for multi-sport scenarios. Our goal is to identify the activities and segment a user's workout trace into constituent activities, including the transition periods. We use an Apple Watch to gather inertial sensor data and validate our system in the context of a triathlon. The system was trained and tested on 3 activities (running, biking, and swimming), as well as simple actions performed in transition, from 5 different participants. Our system achieves 91% accuracy in detecting the activity, and can accurately identify the start and stop times for each. We also validate our results with data collected from a volunteer at a triathlon.