Overview / Usage

The paper was a byproduct of the course Machine Learning for Mobile and Pervasive Systems conducted by Stephan Sigg at the Aalto University. Together with my friend Ferran Montraveta Roca, we wanted to investigate how we could enhance the workout sessions by using some intelligent solution allowing us to track the exercise we are performing. Originally, we wanted to be even able to detect how accurately specific exercise is performed (if you do a proper push-up or half push-up), but we have quickly realized that even if we would be able, at the end of the course, to distinguish multiple exercises from recorded data, it will be a success. After successfully finishing the course, we were contacted by Stephan Sigg and Rainhard Findling with the proposition of preparing our work/paper for application to IWANN 2019 conference. The application was accepted and we had an opportunity to present our work and publish it as a part of Advances in Computational Intelligence 15th International Work-Conference on Artificial Neural Networks, IWANN 2019, Gran Canaria, Spain, June 12-14, 2019, Proceedings, Part I. There are plans for further development and further study of this work.

References:

1. Skawinski K., Montraveta Roca F., Findling R.D., Sigg S. (2019) Workout Type Recognition and Repetition Counting with CNNs from 3D Acceleration Sensed on the Chest. In: Rojas I., Joya G., Catala A. (eds) Advances in Computational Intelligence. IWANN 2019. Lecture Notes in Computer Science, vol 11506. Springer, Cham

Methodology / Approach

To gather the data that allowed us to recognize the exercises types we used 3D accelerator sensor embedded in the Movesense sensor. After studying the related work and preliminary study we decided to mount the sensor on the chest. After gathering data from 10 subjects performing different exercises (we concentrated on 4 types: pushups, situps, squats, jumping jacks and we added one label for data not covering movement related to any of defined exercises types), we produced the dataset by using a sliding window of 1s length and 1/3s overlap. Considering the 52Hz frequency of the sensor, this gave us 156 features per one window. We labelled the dataset and prepared CNN model. After recognizing the exercise we applied a PCA on the window concerning specific exercise type and perform peak detection allowing for counting the repetitions.

Technologies Used

• TensorFlow
• Python
• Google Colab
• Keras

Other links

• Workout Type Recognition and Repetition Counting with CNNs from 3D Acceleration Sensed on the Chest

Collaborators