Abstract
eSports is a rapidly developing multidisciplinary domain, yet research into player behavior remains in its early stages. This work proposes a "smart chair" platform—an unobtrusive approach to collecting data on eSports athletes for processing with machine learning. By analyzing accelerometer and gyroscope data from players of varying skill levels in Counter-Strike: Global Offensive (CS:GO), we demonstrate that professional athletes can be identified by their unique physical behavior during gameplay.
Introduction: The Physical Gap in eSports Coaching
Despite the multi-million dollar scale of the eSports industry, coaching tools often remain limited to basic in-game statistics like Kills/Deaths/Assists (KDA) or subjective "gut feelings." There is a significant lack of understanding regarding the physical state and body language of players. While previous studies have examined peripheral inputs like keyboard and mouse logs, few have explored how a player’s overall posture and movement correlate with skill. This research aims to fill that gap using an unobtrusive sensing platform to optimize training and provide deeper behavioral feedback.
Methods: The Smart Chair Platform
The researchers developed a custom hardware-software system integrated into a standard gaming chair.
- Hardware: A sensing unit containing a Raspberry Pi 3 and an MPU 9250 Motion Processing Unit was mounted to the bottom of the chair.
- Sensors: The system captures data from an accelerometer, gyroscope, and magnetometer at a frequency of 100Hz.
- Feature Engineering: Raw data was transformed into 13 features, including "Active Movements" (large position shifts), "Subtle Oscillations" (micro-movements like swaying), and "Posture" (time spent leaning against the backrest).
- Machine Learning: Four classification algorithms were evaluated: Logistic Regression, Support Vector Machines (SVM), K-Nearest Neighbors (KNN), and Random Forests.
Dataset and Experimental Setup
The study involved 19 participants playing the "Retake" modification of CS:GO. The cohort consisted of 9 professional athletes (primarily from the "Monolith" team) and 10 amateurs (newbies and semi-pros). Each 35-minute session was sliced into 3-minute segments to create a robust dataset of 154 sessions for model training and testing.
Key Insights: The "Pro" Signature
The analysis revealed distinct physical signatures associated with high-level play:
- Controlled Stability: Professional athletes perform significantly fewer "active movements" than amateurs, suggesting higher concentration and physical discipline.
- Engagement Posture: Amateurs were found to lean back more frequently, indicating a less engaged or more relaxed posture compared to the forward-leaning "pro" stance.
- Subtle Dynamics: While pros move less overall, they exhibit specific subtle rotational swaying (y-axis) characteristic of their focused playstyle.
Key Results
The study successfully demonstrated that physical behavior is a strong predictor of gaming skill. The Support Vector Machine (SVM) model emerged as the top performer.
The SVM achieved an Area Under the Curve (AUC) of 0.86 for binary classification (Pro vs. Amateur). Logistic Regression followed closely with an 0.85 AUC, suggesting a stable, linear relationship between physical features and skill. Notably, rectilinear motion along the x-axis (side-to-side movement) was identified as the most characteristic feature of professional athletes.
Conclusion and Future Directions
This research proves that eSports expertise manifests in the behavior of the entire body, not just the hands. While the current sample size is small, the results provide a foundation for more advanced systems. Future work will focus on expanding the sensor array (e.g., pressure and temperature sensors), using Hidden Markov Models for better game-round segmentation, and testing across different genres like MOBAs to see if these physical signatures are universal across eSports.
Summary updated