MiLift: Efficient Smartwatch-Based Workout Tracking Using Automatic Segmentation

The use of smartphones and wearables as sensing devices has created innumerable context inference apps including a class of workout tracking apps. Workout data generated by mobile tracking apps can assist both users and physicians in achieving better health care, rehabilitation, and self-motivation. Previous approaches impose extra burdens on users by requiring users to select types of exercises or to start/stop sessions. In this paper, we propose MiLift, a practical end-to-end workout tracking system that performs automatic segmentation to remove user burdens. MiLift uses commercial off-the-shelf smartwatches to accurately and efficiently track both cardio and weightlifting workouts without manual inputs from users. For weightlifting tracking, MiLift supports both machine-based and free weight exercises, and proposes a lightweight repetition detection algorithm to ensure efficiency. A research study of 22 users shows that MiLift can achieve above 90 percent average precision and recall for cardio workout classification, weightlifting session detection, and weightlifting type classification. MiLift can also count repetitions of weightlifting exercises with an average error of 1.12 reps (out of an average of 9.65). Our empirical app study on a Moto 360 watch suggests that MiLift can extend watch battery lives by up to 8.25χ (19.13h) compared with previous approaches.