Sitting Posture Recognition Using a Spiking Neural Network

Jianquan Wang, Basim Hafidh, Haiwei Dong, Abdulmotaleb El-SaddikarXiv 2022 · DOI · OpenReview

Abstract

To increase the quality of citizens' lives, we designed a personalized smart chair system to recognize sitting behaviors. The system can receive surface pressure data from the designed sensor and provide feedback for guiding the user towards proper sitting postures. We used a liquid state machine and a logistic regression classifier to construct a spiking neural network for classifying 15 sitting postures. To allow this system to read our pressure data into the spiking neurons, we designed an algorithm to encode map-like data into cosine-rank sparsity data. The experimental results consisting of 15 sitting postures from 19 participants show that the prediction precision of our SNN is 88.52%.

Introduction

In our increasingly sedentary world, poor sitting habits are a leading cause of lower back pain and long-term musculoskeletal issues. While "smart chairs" exist, they often face a trade-off: they are either prohibitively expensive or computationally limited, recognizing only a handful of basic positions. This research introduces a system that is both affordable—using standard Force-Sensitive Resistors (FSRs)—and sophisticated enough to distinguish between 15 nuanced postures. By utilizing Spiking Neural Networks (SNNs), which are more biologically realistic and energy-efficient than traditional networks, the system provides a path toward real-time, energy-efficient health monitoring.

System Architecture and Methods

The proposed system architecture transitions from raw data acquisition to complex posture classification through a multi-stage pipeline.

The overall architecture of the Liquid State Machine system for posture recognition, showing the flow from encoding to the liquid reservoir and final classification.

The Encoding Algorithm

A critical innovation of this work is the cosine-ranked sparsity algorithm. SNNs require time-varying "spikes" as input, whereas pressure sensors provide static "maps." This algorithm transforms static pressure values into a temporal sequence of 0/1 spikes that the neural network can process. This transformation not only satisfies the SNN requirements but also improves classification performance by tuning the data's sparsity.

The Liquid State Machine (LSM)

The core model is a Liquid State Machine, a type of recurrent SNN. It consists of a "reservoir" of Leaky Integrate-and-Fire (LIF) neurons (75% excitatory, 25% inhibitory). The "liquid" state of this reservoir captures the spatiotemporal patterns of the pressure data. Because the internal recurrent structure naturally creates a high-dimensional map of the input, a simple Logistic Regression (LR) layer can act as the "readout" to classify the 15 postures with high accuracy.

Hardware Implementation

The physical system consists of a custom-built chair equipped with two FSR grids: a 9x9 array on the seat and a 10x9 array on the backrest, totaling 190 sensing points. These are managed by an Arduino Mega 2560.

Hardware setup showing (a) the pressure sensing grid, (b) front view, and (c) back view of the smart chair.

Experimental Results and Insights

The study involved 19 participants and a total of 5,944 pressure maps. To test generalization, the authors used a "subject-split" approach, training on 15 subjects and testing on 4 entirely new subjects.

  • High Precision: The SNN+LR model achieved a precision of 88.52% and an F1 score of 0.86, outperforming traditional baselines like SVM (82%), KNN (72%), and standard ANNs (85%).
  • Efficiency: The system achieved a processing speed of 10 frames per second (fps) on a standard desktop CPU, making it viable for real-time haptic feedback via vibration motors.
  • Bio-inspired Advantage: By using LIF neurons, the system carries information in the timing between spikes, offering significant energy efficiency for future embedded or wearable applications.

Conclusion and Future Work

While the results are promising, future improvements could address parameter sensitivity in the encoding algorithm and expand the dataset to include a more diverse range of body types. Additionally, exploring more robust sensing materials could mitigate the potential for FSR sensor drift over time. This work demonstrates that bio-inspired computing can effectively bridge the gap between low-cost hardware and high-fidelity behavioral recognition.

Summary updated

Abstract

To increase the quality of citizens' lives, we designed a personalized smart chair system to recognize sitting behaviors. The system can receive surface pressure data from the designed sensor and provide feedback for guiding the user towards proper sitting postures. We used a liquid state machine and a logistic regression classifier to construct a spiking neural network for classifying 15 sitting postures. To allow this system to read our pressure data into the spiking neurons, we designed an algorithm to encode map-like data into cosine-rank sparsity data. The experimental results consisting of 15 sitting postures from 19 participants show that the prediction precision of our SNN is 88.52%.
BibTeX
@misc{wang2022sitting,
  author = {Jianquan Wang and Basim Hafidh and Haiwei Dong and Abdulmotaleb El-Saddik},
  title = {{Sitting Posture Recognition Using a Spiking Neural Network}},
  year = {2022},
  eprint = {2212.12908},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2212.12908},
}
Figure