Link to paper

The full paper is available here.

You can also find the paper on PapersWithCode here.

Abstract

  • Designed a personalized smart chair system to recognize sitting behaviors
  • Used a liquid state machine and logistic regression classifier to construct a spiking neural network
  • Designed an algorithm to encode map-like data into cosine-rank sparsity data
  • Experimental results show prediction precision of 88.52%

Paper Content

I. introduction

  • Smart cities are using digital twins to improve quality of life
  • Global Burden of Disease study shows people are suffering from lower back pain due to sitting behaviors
  • Sensing systems are needed to provide real-time feedback on sitting posture
  • This paper presents a deep learning-powered chair to teach best sitting posture and provide feedback
  • Chair uses pressure sensors to obtain pressure data and a spiking neural network to recognize sitting posture

A. pressure sensing

  • Pressure-mapping sensors are used to analyze weight-supporting surfaces in shoes, beds, and chairs.
  • Examples of shoe studies include assessing pressure discomfort thresholds, pain thresholds, and tissue stiffness.
  • Examples of bed studies include helping caregivers adjust bedridden patients and developing a monitoring framework.
  • Examples of chair studies include monitoring sitting behavior, developing a driving assistant, and using pressure sensors to give feedback.

B. smart chair

  • Smart chair designed to process real-time information
  • Includes chairs for different purposes, sensors, microcontrollers, and classification ability
  • Can recognize 8-15 postures
  • Includes feedback mechanism to alert user to risky posture

C. spiking neural network

  • Artificial Neural Networks (ANNs) are simulations of biological neural networks.
  • Timing phenomena, such as time differences between spikes, are part of information-processing mechanisms in biological nervous systems.
  • Spiking Neural Networks (SNNs) are the closest models to biological neural networks.
  • Supervised learning algorithms for SNNs can be divided into gradient descent learning algorithms, supervised STDP learning algorithms, and learning algorithms based on spike sequence convolution.

Iii. system description and pressure sensor

A. system description

  • Designed hardware instrument and software interface for data acquisition
  • 15 common sitting postures designed to be recognized
  • Trained network to identify real-time sitting postures

B. hardware description

  • Chair composed of transducer, data collection, and calibration modules
  • Transducer module includes interconnected sensors and actuator units
  • Data collection module collects measured data from transducer module and passes to interface module
  • Calibration module calibrates transducer units before each test

C. software description

  • User interface provides communication between chair and external environment
  • Three communication media: USB, Bluetooth, ZigBee
  • System supports Bluetooth and USB
  • Data collection module exchanges data and control signals with posture classifier
  • Posture classifier module receives sensor data, stores in database, analyzes, and informs user of posture
  • Machine learning mechanism used to investigate poses and activate embedded actuators if posture is incorrect

Iv. posture recognition approach a. spiking neuron and synapse

  • Neurons in SNN determined by membrane potential and activation threshold
  • Membrane potential determined by excitatory and inhibitory postsynapse potentials
  • Neuron fires when membrane potential reaches activation threshold
  • Neuron model used is Leaky Integrate-and-Fire (LIF)
  • Weight assigned to each synapse affects membrane potential of target neuron
  • Liquid State Machine (LSM) is a type of recurrent neural network

B. liquid state machine

  • The LSM is composed of an input layer, liquid reservoir and readout layer.
  • The liquid reservoir maps the signal from the input layer onto a liquid state.
  • The readout layer is a detector that can be trained to recognize patterns from the liquid filter.

C. encoding method

  • Algorithm 1 is designed to encode pressure matrix data into a cosine-ranked 0/1 matrix with sparsity.
  • Algorithm 1 maps the pressure matrix data to a range of [0, π] and adjusts the sparsity of the 0/1 matrix with scalar ratios A and n.

V. experiment and results

A. data acquisition

  • 19 participants (14 male, 5 female)
  • Age range 22-58, median 25, mean 28.53
  • Pressure range 691-1023, median 1007, average 988.74
  • Signed consents obtained before experiment
  • Data divided into training and validation sets

B. experiment setup

  • Set up 2 groups of models for testing SNN performance
  • First model was LR classifier, second model was LSM with LR as readout layer
  • Two trials with encoded/nonencoded data in first group, three trials with nonencoded/two sets of encoded data with different sparsities in second group
  • Set value of C in equation 4 to 0.3 (EE), 0.2 (EI), 0.4 (IE), 0.1 (II) and λ to 1.6667
  • Set number of E neurons to 1600, I neurons to 400, input neurons depending on coding number n
  • Activation threshold of each neuron was 15 mV, resting potential 13.5 mV, refractory period of E neurons 3 ms, I neurons 2 ms
  • Set amplitude of A to 30, duration for each simulation to 30 ms, weights initialized with random numbers following gamma distribution with shape α of 30 (EE), 60 (EI), 19 (IE), 19 (II), 18 (input to E), 9 (input to I)

C. experiment results

  • Performance comparison of two models shown in Table I
  • Logistic regression achieved satisfactory classification accuracy
  • Encoded data still can be classified by logistic regression
  • Structure used in paper verified to be effective
  • Coding method can modify number of coded neurons
  • Runtime of proposed system can reach 10 fps
  • Data-set split by subjects for training and testing
  • Proposed approach superior to other machine learning algorithms
  • SNN and ANN have similar performance
  • Proposed method has better learning ability than linear classifiers
  • Proposed system can recognize largest sitting posture classes

Vi. conclusion

  • System presented for sitting posture recognition
  • System is cost-effective and has high accuracy
  • Hardware used is an FSR matrix
  • Low-density pressure matrix is sufficient to distinguish between different postures

Algorithm

  • Sensor Pos.Num ANN used an 8x8x1 FSR grid with 8 sensors
  • SVM used a 16x16x2 FSR grid with 8 sensors
  • Determine used 8 FSR sensors
  • PCA used a 42x48 Tekscan BPMS with 14 sensors
  • ANN used a 42x48 Tekscan BPMS with 9 sensors
  • SIR used a 42x48 Tekscan BPMS with 10 sensors
  • SNN+LR used 9x9 and 9x10 FSR grids with 15 sitting postures
  • SNN was able to accurately classify 15 common sitting postures