AI-powered Posture Training: Application of Machine Learning in Sitting Posture Recognition Using the LifeChair Smart Cushion

Katia Bourahmoune, Toshiyuki AmagasaIJCAI 2019 · DOI

TL;DR

This paper introduces the LifeChair, a smart cushion that uses pressure sensors and machine learning to recognize sitting postures and guide users through stretches. The system achieves high accuracy in posture detection and demonstrates the importance of user Body Mass Index (BMI) for personalized recognition, offering a promising step towards AI-driven well-being solutions.

What Problem Does This Paper Address?

In today's world, sedentary lifestyles are a growing concern, with many people spending over half their day sitting. This prolonged sitting, often in poor posture, is linked to a host of physical and mental health issues, including back pain, headaches, fatigue, and reduced productivity. While various solutions exist, from ergonomic furniture to active tracking devices, many fall short. Passive solutions don't guarantee good posture adoption, and current active solutions often have limited sensing capabilities or provide inadequate feedback. Furthermore, there's a lack of standardized data for training robust sitting posture recognition systems. This paper tackles the challenge of accurately and in real-time recognizing sitting postures and guiding users through beneficial stretches to promote better health and well-being.

Methods

The core of this research is the LifeChair, a smart cushion equipped with nine pressure sensors and four vibration actuators. This cushion, combined with a smartphone app, forms the basis of their AI-powered posture training system.

  • Pressure Sensing: The LifeChair uses a novel textile sensing technology to capture pressure distribution across the cushion. This data is collected at 5Hz.
  • Calibration: Before use, the system calibrates to the user's upright posture, establishing a reference frame for pressure distribution. This calibration accounts for individual differences.
  • Posture Recognition: Machine learning (ML) algorithms are employed to classify sitting postures based on the pressure sensor data. The authors experimented with several algorithms, including Decision Trees (DT-CART), Random Forest (RF), K-Nearest Neighbours (KNN), Linear Regression (LR), Linear Discriminant Analysis (LDA), and Naïve Bayes (NB).
  • Stretch Pose Recognition: A similar ML approach is used to classify six common chair-bound stretches, with data labeled manually through video cross-referencing.
  • Feature Engineering: The study investigates the impact of including user Body Mass Index (BMI) as an additional feature for posture recognition, hypothesizing that it can account for inter-individual variability.
  • Haptic Feedback: The vibration actuators in the LifeChair are used to provide real-time feedback to the user, guiding them to correct their posture.

Datasets

The paper introduces a newly collected dataset for both sitting posture and seated stretch poses.

  • Sitting Posture Dataset: Collected from 10 healthy subjects (5 male, 5 female) with varying BMIs (Low, Normal, High). This dataset includes raw sensor values, user-input data (height, weight, back-pain history), and posture labels. The authors report detecting over 13 different sitting postures.
  • Seated Stretch Pose Dataset: Also collected from the same 10 subjects, this dataset comprises data for six common chair-bound stretches. The stretch poses were manually labeled using video cross-referencing.

Key Insights and Ideas

  • BMI Matters for Posture Recognition: The study strongly suggests that user BMI is a significant factor influencing the accuracy of sitting posture recognition. Ignoring BMI can lead to lower performance, especially for users with lower BMIs. This highlights the need for personalized approaches in sensor-based health monitoring.
  • Random Forest is a Strong Performer: For both posture and stretch recognition tasks, the Random Forest algorithm consistently delivered the highest accuracy among the tested classifiers. Its robustness and ability to handle complex data make it a suitable choice for this application.
  • The Importance of a Comprehensive Posture Model: The LifeChair's biomechanics-based posture model, combined with a flexible sensor layout covering various body parts, allows for the detection of subtle and common postures (like Forward Head Posture and Rounded Shoulders) that might be missed by simpler systems.
  • Stretching is a Key Component of Well-being: The research extends beyond just posture correction to include stretch pose recognition, acknowledging that regular stretching is crucial for overall health and preventing musculoskeletal issues.

Key Results

  • Sitting Posture Recognition: The system achieved an impressive 98.93% accuracy in detecting over 13 different sitting postures using the Random Forest algorithm with sensor data and BMI as input. When using sensor data alone, the accuracy was 97.76%.
  • Stretch Pose Recognition: The Random Forest model achieved 97.99% accuracy in classifying six common chair-bound stretches using sensor data and BMI.
  • BMI Impact: The study found a statistically significant difference in accuracy based on BMI groups, with lower accuracy for the Low BMI group (0.9704%) compared to Normal (0.9849%) and High BMI (0.9883%) groups when using Random Forest.

Limitations and Future Work

  • Limitations:
    • The study acknowledges that the accuracy of posture recognition is optimized for the specific postures generated by their threshold-based method and the dataset collected.
  • The paper notes that the "stretch pose where subjects had their right leg up and across (RLU) was the most prone to misclassifications," potentially due to user sitting behavior and lumbar region pressure distribution.
  • While the system aims for real-time performance, the computational cost of ML models for mobile deployment is a consideration.
  • Future Work:
    • Personalization: The authors plan to improve accuracy for the Low BMI group by implementing a lower calibration threshold (alpha) for users with a BMI below 18.5 kg/m², enabling user-based personalization.
  • Long-Term Analysis: Extend experiments to include long-term performance analysis and behavior prediction to capture a wider range of posture-related parameters.
  • Real-World Deployment: Validate the AI-based approach in corporate and automotive scenarios by deploying LifeChairs in offices and cars to track worker and driver posture over extended periods and evaluate its impact on well-being.
  • Addressing Dynamicity and Variability: Further address the inherent dynamicity of sitting posture and the vast inter-individual variability in user behaviors and environments.

Summary updated

BibTeX
@inproceedings{bourahmoune2019ai,
  author = {Katia Bourahmoune and Toshiyuki Amagasa},
  title = {{AI-powered Posture Training: Application of Machine Learning in Sitting Posture Recognition Using the LifeChair Smart Cushion}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.24963/ijcai.2019/805},
}
Figure