ExerSense: Real-Tme Physical Exercise Segmentation, Classification, and Counting Algorithm Using an IMU Sensor

TL;DR

This paper presents a real-time, environment-agnostic algorithm for exercise segmentation and classification using IMU sensors, achieving high accuracy with minimal training data, applicable indoors and outdoors.

Contribution

It introduces a novel correlation-based method for exercise recognition that works in both indoor and outdoor settings with minimal training data.

Findings

Achieves 95% classification accuracy on five exercises

Works well in both indoor and outdoor environments

Requires only one sample per exercise for training

Abstract

Even though it is well known that physical exercises have numerous emotional and physical health benefits, maintaining a regular exercise routine is quite challenging. Fortunately, there exist technologies that promote physical activity. Nonetheless, almost all of these technologies only target a narrow set of physical activities (e.g., either running or walking but not both) and are only applicable either in indoor or in outdoor environments, but do not work well in both environments. This paper introduces a real-time segmentation and classification algorithm that recognizes physical exercises and that works well in both indoor and outdoor environments. The proposed algorithm achieves a 95\% classification accuracy for five indoor and outdoor exercises, including segmentation error. This accuracy is similar or better than previous works that handled only indoor workouts and those use a vision-based approach. Moreover, while comparable machine learning-based approaches need a lot of training data, the proposed correlation-based method needs one sample of motion data of each target exercises.