Augmenting the Boxing Game with Smartphone IMU-based Classification System on Waist
1. Augmenting the Boxing Game with Smartphone
IMU-based Classification System on Waist
Keio University
Fei Gu
Keio University
Chengshuo Xia
Keio University
Yuta Sugiura
01 Overview
02 Method
Workflow
• Player put a smartphone with calibrated sensors on waist
• Perform 30-minute gaming boxing
• Export data from the smartphone
• 100-frame timeseries data normalized and converted grayscale images
• CNN structure for training
• Inertial sensors contribute
interactively to sports recognition
• Capability to distinguish specific
boxing actions?
Motivation
• No consensus on measurement/
classification method for boxing
• Professional IMUs not commonly
available
Background
• Augment the video game with
smartphone sensors
• Classify six basic punches with a
12-layer CNN model
Approach
04 Discussion
05 Future Work
Possible Quantitative Evaluation
• Use 3D model built for data simulation
• Data generation based on extracted in-game model and empirical
adjustments referencing the professional boxer’s movements
• Score calculation by the distance between the sensor data points of
the players’ and the simulated ones
2022 International Conference on Cyberworlds (CW) Contact: fegu@keio.jp
03
Participants
• 4 female, 6 male
• 20~25 years old
• Varying experiences
Dataset
• 6 types of punches: jab,
straight, left hook, right
hook, left uppercut, and
right uppercut
• 24 times each, 1440 in
total
Training
• TFLearn package
• Commercial computer for
150 epochs
In-Group Test
• Randomly shuffled
• 264 holdout test samples
• Accuracy: 79.2%
In-Subject Test
• Test on 6 samples for
every individual
• Accuracy: 86.4%
Experiment
Phyphox
Fit Boxing 2
9DOF inertial data, 100Hz
• Possible as personal boxing helper
• Small dataset, might be overfitting
• Challenging to predict the movements of unknown users
(a)