1. Introduction
Scientific and technological advancements that have been made during the last decades have radically changed the lifestyle of most of the people living in post-industrial countries. Thanks to the modernization of industries and cities, the need for human physical activity in daily tasks has been reduced significantly. Nowadays, machines have taken charge of almost all heavy occupations. As a result, jobs have mainly to do with office tasks in which personal soft skills are mostly required as opposed to hard ones and sitting still in front of computer screens for extended periods of time is the more frequent habit. Commuting is generally performed using personal vehicles or public transportation systems. Houses are structured to provide maximum comfort, obtained by means of modern facilities that require moderate or no physical activities. After many sedentary working hours, television and social media occupy large numbers of people, who once again remain seated in front of their screens. On the other hand, inactivity aiming at preserving the vital energy needed for survival and for food search is the preferred genetic choice [
1]. Indeed, our hunter-gatherer ancestors were considerably more active than we are as individuals in modern societies [
2]. Nevertheless, they had long sedentary hours during the day as well, but opposite to what happens in modern times on modern chair designs, their sitting postures were engaging more muscular activities than today [
3]. Consequences of such habits are that the current sedentary lifestyle is known to increase the risk of several chronic health conditions [
4], such as cardiovascular diseases [
5], obesity [
6], type 2 diabetes [
7,
8], colon cancer [
9,
10], osteoporosis [
11], and depressive illnesses [
12]. Furthermore, thanks to high levels of automation and the integration of Artificial Intelligence (AI) in many aspects of our daily life, future lifestyles with increased sedentary attitudes are easily predictable for the majority of individuals. Therefore, it is very important to investigate and implement methods able to reduce the negative effects of this unavoidable sedentary lifestyle.
In response to the side effects of a sedentary life style, recently, sitting posture monitoring and corrective applications have received much attention among researchers. Several kinds of smart chairs have been developed and experimented using various methods, including image processing [
13,
14] and chairs’ sensor data classification, using various sensor types such as single pressure sensors [
15,
16,
17,
18], pressure distribution sensors [
16,
19,
20,
21], textile sensors [
22,
23], accelerometers [
24], and combined sensors [
25,
26], just to name a few. For example, in [
20] the authors adopted an 8 × 8 pressure sensor array positioned on the sitting cushion of a chair. They employed an Artificial Neural Network (ANN) model to classify eight sitting postures. The model achieved a classification accuracy of
in the subject-independent case by using the data collected from eight subjects as the training set and the data collected from another eight subjects as the test set. In [
19], pressure mapping sensors were mounted on the surface of the sitting cushion and backrest of a chair, and a Principal Components Analysis (PCA) model was used to identify pattern similarity and to classify 10 sitting postures. This study achieved a sitting postures classification accuracy of 96% and 79% in the subject-dependent and subject-independent cases, respectively. In [
16], the authors presented an optimization model for sensor positioning to reduce the number of pressure sensors to be deployed. Using a logistic regression classifier, they could achieve a classification accuracy of 87% for 10 postures with 31 pressure sensors, and of 78% with 19 pressure sensors in subject-independent cases. In [
23], the design of a textile sensor, named eCushion, was presented. The system included a 16 × 16 sensor textile array, a signal acquisition unit based on an Arduino board equipped with Bluetooth for data transmission, and a smartphone featuring a custom application for computation and data monitoring. The experimental results for seven postures classification showed a recognition rate of 92% in the subject-dependent case, and 79% in the subject-independent case. In [
25], the authors designed a mixed sensors system combining six pressure sensors and six infrared sensors. This combination allows the user to perform pressure measurements on the sitting cushion by means of the pressure sensors and the backrest to measure back distance by means of the reflective ones. The authors employed a k-Nearest Neighbor (KNN) algorithm to classify 11 sitting postures, achieving 92% of classification accuracy in the subject-independent case. Finally in [
15], a smart Internet of Things (IoT) system for sitting posture detection was proposed. It was based on the installation of six force sensors, communicating with a mobile application. In this design, several chairs were equipped with the six force sensors and a NodeMCU board. Resistance measurements read by the sensors were sent to a cloud-based platform, where a real-time algorithm was running, providing postures classification in three categories, green (correct), orange (wrong) or red (unhealthy), defined by means of three experimentally determined separation thresholds. A smartphone application was monitoring the postures classification results for the user in real-time.
This review of the smart chair studies for sitting postures classification obtained with various methods and sensing devices shows that in most cases a high classification rate was achieved when a large number of sensors were deployed (≥30) [
16,
19,
20,
23], or when sensors fusion was adopted [
25,
26]. In some cases, results were not validated with k-fold cross-validation methods [
18,
20,
21,
23,
26]. In many cases, the proposed systems were evaluated by means of experimental setups involving the participation of few subjects (≤10) [
18,
23,
26], as their results could be affected by high variance, thereby producing poor generalization performances [
27].
This paper presents the achievement of our three main objectives for this project: (1) design a new, low cost and simple smart chair sensors system that has the advantage of being at the same time robust, versatile, accurate, and easy to setup; (2) optimize the number and position of the sensors to reach a high classification accuracy; and (3) identify the best deep learning model for our designed smart chair.
To achieve the mentioned objectives, we performed the following steps. (1) We designed the smart chair sensors system employing eight low-cost resistive pressure sensors which were placed on the sitting cushion and on the backrest of a chair. (2) The sensor signals were acquired by a purposely designed data acquisition unit. In contrast to almost all the aforementioned studies that adopted a commercial product such as Arduino, we designed the acquisition system and realized a Printed Circuit Board (PCB) specifically for the smart chair project. The designed PCB presents several advantages with respect to Arduino-based boards, since we could customize the components’ choice according to our specific needs. A first advantage is that the chosen Digital Signal Processor (DSP) has an on-board 12-bit Analog to Digital Converter (ADC) with 16 analog inputs (DSPIC30F6014A from Microchip), while normally Arduino features a 10-bit ADC. This choice increased the measurement resolution by a factor of four, allowing us to resolve 0.00025 kg force acting on each sensor. A second advantage with respect to Arduino-based boards is that we can integrate the entire electronic circuitry on a unique PCB, thus eliminating any auxiliary boards (e.g., Wi-Fi module) that usually must be externally wired to commercial boards. (3) We obtained detailed characterization of the analog front-end of the setup by performing linearity, repeatability, and bandwidth analysis of the sensors system. (4) The designed board was placed on the back of the chair, hosted in a purposely designed 3D-printed enclosure. It is battery powered and can transmit data to a laptop via Wi-Fi continuously up to 30 h without recharging. Therefore, the system is wireless, thus avoiding the need for cumbersome cabling. (5) We developed a Graphical User Interface (GUI) application on a laptop. It receives data and stores them locally, while plotting them in real time on the screen. (6) An experiment has been set up, and a dataset has been collected from 40 individuals sitting in eight different postures. (7) The acquired dataset was employed to train and test seven different deep learning models. (8) To find the most suitable model for our application, the average classification accuracies of all models, obtained by k-fold cross validation, were compared. As it will be shown, the Echo Memory Network (EMN) model scored the best average classification accuracy of . The choice of the sensors type, their number, and positions in combination with the custom designed acquisition board that transmits the acquired data without wires to the developed GUI on a computer are the main innovative aspects of the designed smart chair, which produced relatively high classification accuracy.
The paper is organized as follows.
Section 2 describes the design of the smart chair sensors system (hardware and software components), the experimental setup used to acquire the sitting postures dataset (with preprocessing steps), and the Deep Learning (DL) models’ architecture and hyperparameters.
Section 3 shows the experimental results obtained from the characterization of the analog front-end of the sensors system and the results achieved during the postures classification experiment, including result comparison among the seven DL models. Finally,
Section 4 presents the discussion on various sitting patterns and the related challenges, the computational costs of each DL model, and a comparison between the results of our study and other similar ones. Finally,
Section 5 draws the conclusions of the paper.
2. Materials and Methods
This section is organized in three parts: the first part describes the design of the developed smart chair sensors system as performed according to the required technical specifications; the second part describes the experimental setup and the procedure followed to acquire data from the 40 participants sitting in eight different sitting postures; the third one describes the architecture of the seven DL models employed for the classification of the eight sitting postures.
The general structure of the proposed smart chair posture recognition system is presented in
Figure 1. The smart chair sensors system board (shown as the green box in the figure) has the duty to acquire signals from the chair’s sensors, convert them into digital data, and transmit them to the computer by means of a TCP link. These operations are described in
Section 2.1.1 and
Section 2.1.2, respectively. The designed GUI (shown as the turquoise box in the figure) has the duty to monitor the received data, collect datasets, label them, and store them as text files. The GUI is explained in detail in
Section 2.1.3. Data preprocessing steps (shown as the pink box in the figure) and classification of the eight sitting postures P1,
…, P8 (shown as the yellow box in the figure) were coded in Python scripts and are described in detail in
Section 2.2.2 and
Section 2.3, respectively. Sitting postures are described in
Section 2.2.1.
2.1. Smart Chair Sensors System
The developed sensors system (see
Figure 2a for the block diagram) evaluates the user’s body pressure on the sensorized chair by means of eight Force Sensing Resistors (FSR); five of them were placed on the sitting cushion and three on the backrest of the chair, as shown in
Figure 2b. The location of each sensor was adjusted to favor the classification of the eight sitting postures. We verified that the conductance of the FSR sensors changes almost linearly with the applied pressure, i.e., the resistance of the sensors decreases by increasing the applied pressure. A sensors system circuit was designed to: (1) convert the conductance of each sensor into a voltage; (2) condition the signals by filtering and amplification according to the specifications described in
Section 2.1.1; (3) process the signals by a DSP; (4) transmit the data by a Wi-Fi module to a laptop where a custom application GUI, as described in
Section 2.1.3, was developed for data storage and real-time monitoring.
The sensors system is powered by a single 2500 mAh Lithium-Polymer (LiPo) rechargeable cell battery (see
Figure 3c). With a multimeter, we experimentally measured the current consumption of the entire board during continuous transmission, obtaining as a result 85 mA; the main contribution of current consumption is due to the Wi-Fi module whose consumption is 70 mA. Thus, the LiPo cell allows for ≈30 h of continuous transmission without battery recharging. The choice of a battery-operated system allows us to easily use the chair without the need of power line plugs or wires, with an advantage of usability with respect to the existing literature (e.g., [
20,
23]). A buck DC–DC converter (not shown in
Figure 2a) provides a +3.3 V supply voltage for the entire circuit. A reference voltage
= 100 mV, required for the Op-Amps implementing low-pass active filters and the conductance to voltage conversion, is generated using a linear voltage reference and routed to the non-inverting input of the Op-Amps.
For the circuit assembly, a two-layer PCB was printed and populated on both sides. Its top layer, shown in
Figure 3a, holds the analog front-end, while the bottom one, shown in
Figure 3b, holds the power supply, the reference voltage
generation, and the Wi-Fi module. The PCB and the battery were accommodated in a custom-designed 3D-printed enclosure, shown in
Figure 3b,c. The external dimensions of the circuit and of the enclosure are 70 × 53
and 75 × 65 × 34
, respectively.
2.1.1. The Analog Section
Referring to
Figure 2a, the sensors’ readout signals
are connected to a conductance to voltage converter circuit with a first-order low pass behavior. The minimum conductance
of the sensors is nearly zero when no pressure is applied and increases with pressure towards a maximum value
. Indeed, sensors on the sitting cushion are subject to a pressure that is significantly higher than that activating the backrest sensors. We experimentally determined the maximum resistance for the two cases, obtaining
= 12 k
for the sitting cushion case and
= 17 k
for the backrest one. To obtain the full scale of the measurement with the minimum resistances of the sensors, we calculated the feedback resistor of the filters to be
= 390 k
for the sitting cushion circuits and
= 560 k
for backrest circuits. The transfer function of the analog section for each channel is:
where
is the angular frequency. Designing
= 32 ms, the analog section behaves as a low pass conductance to voltage converter, having a cutoff frequency of 5 Hz.
2.1.2. ADC, DSP, and Data Transmission
The signals are routed to the analog inputs of the DSP (DSPIC30F6014A from Microchip), featuring an on-board 12-bit ADC and operating at the peak rate of 8 MIPS. The DSP receives the analog signals and converts them into 12-bit numbers operating with a 10 Hz sampling rate. After conversions, the DSP assembles the data packets and sends them through an on-board Universal Asynchronous Receiver Transmitter (UART) to a low-power Wi-Fi module (USR-C216), which sends them to a computer via TCP protocol.
2.1.3. The Software Description
We developed a GUI using LabVIEW™ from National Instruments. The front panel of the GUI is shown in
Figure 4. The GUI is designed to acquire the transmitted data from the chair’s Wi-Fi module, visualize them in real-time, and store them locally on the computer.
To monitor the real-time data from the chair sensors, two graphs are provided as shown at the right side of
Figure 4: the top chart plots the time behavior of the three signals coming from the backrest sensors, while the bottom chart plots the time behavior of the five sitting cushion sensors. Signal values were scaled to be always in the dimensionless range [0, 100]%, where 100% and 0% indicate the maximum and minimum pressure values, hence minimum and maximum sensor resistance, respectively. On the left side of the GUI, in addition to the chair and sensors sketch, few controls were added, such as the IP address, the port of the TCP communication protocol, and the path and name of the store file. In addition, a selectable list of different postures to test was implemented to tag the present acquisition with the corresponding label. The front panel was furnished with a button to connect/disconnect the data acquisition and monitoring. Finally, possible communication errors were visualized in an appropriate TCP error box.
2.2. Experimental Setup
The designed sensors system described above was mounted on the chair shown in
Figure 2b. Three sensors (tagged Back1, Back2, Back3) were mounted on the backrest, while five sensors (tagged Seat1, Seat2, Seat3, Seat4, Seat5) were mounted on the sitting cushion. The 3D-printed enclosure containing the embedded electronics was mounted on the rear of the chair. Since the board is battery-operated and data transmission is wireless, there are no cables extending out of the chair which consequently can be freely moved around, a relevant feature for practical applications.
2.2.1. Sitting Postures
Sitting and standing can occur in various positions, named postures. A posture is the preferred bio-mechanical position of the body in space. The muscular system works together with the skeletal system to keep the body in balance during either dynamic movements or static conditions [
28].
Eight sitting postures were practiced during the setup experiment in which 40 volunteers participated (28 males, 12 females, age 39 ± 17). Each subject sat still on the smart chair assuming in turn, all eight postures, maintaining each one for 60 s and repeating each posture three times. The first posture, P1 of
Figure 5a, represents the sitting position upright with straight back and symmetric limbs. Generally, this is the recommended posture when sitting on a chair, while the second posture, P2 of
Figure 5b, represents slouching with curved back, inclining the head forward and lowering the shoulders. This posture and its variations are very common among office employees. In contrast to P1, in this posture, the functionality of internal organs, such as heart, lungs and digestive system, may in the long term become negatively impacted, and lead to backache and pain [
29,
30]. The third posture, P3 of
Figure 5c, shows a forward bending, with straight back and symmetric limbs, while P4 of
Figure 5d is similar to P3, but the bending direction is backward. The fifth and sixth postures, P5 and P6 of
Figure 5e,f, respectively, represent the bending of the torso toward the left and right sides, respectively, while the legs remain symmetrical and the feet are flat to the ground. Finally, the last two postures, P7 and P8 of
Figure 5g,h, are the case where the right leg crosses over the left one and vice versa, the torso remaining upright and erect.
As a matter of fact, each posture creates a different weight distribution on the eight sensors of the chair. Indeed, several factors (such as weight, height, body shape, muscle strength, etc.) influence the way someone sits on a chair, but in general, the distribution of the weights obtained for a given posture by different subjects is expected to show a similar pattern.
2.2.2. Data Acquisition and Preprocessing
Before sitting on the smart chair, the volunteers were informed about the purposes of the study, the procedures of the test, and data anonymization performed for privacy enforcement. Each session of the experiment was conducted in three steps: (1) the subject was invited to sit on the chair with empty pockets to keep the symmetry of the weight distribution on the sensors (the presence of wallet or smartphone in the back pockets could alter the pressure distribution); (2) all sitting postures were explained and demonstrated to the subject; and (3) the subject was guided to sit in each posture in a natural and comfortable position and to keep it for the 60 s, of the data acquisition phase. As said, each of the eight postures was repeated three times. Thus, 24 datasets were generated for each subject. Since the sampling frequency was 10 Hz, each dataset contained approximately 600 samples. Given the 40 subjects, a total of 960 datasets were generated, for a grand total of samples. Each dataset contains the acquisition time for each sample, the eight readouts of the eight sensors, and the indication of the posture type annotated by the designed GUI.
To standardize the datasets in view of the DL analysis, a simple Python code was developed to preprocess each file. The first and the last 10 samples (i.e., the first and last seconds of acquisition) were discarded to remove possible transient behaviors due to the sitting and get up actions of the subject, or because the posture changed. Consequently, each dataset was reduced to 580 samples for a new grand total of samples, an appropriate procedure that was performed without losing generality.
To reduce the inherent noise contributions, artefacts, and spurious values possibly due to the subjects’ movements during signal acquisition were smoothed using a Savitzky–Golay filter from the scipy.signal library, whose window length and polynomial order were set to 301 and 3, respectively.
Since the sensor values stored by the GUI were designed to be in the percentage range [0, 100], normalization to unity was performed to bound all values within the [0, 1] real interval.
2.3. Sitting Postures Classifications
To classify the eight sitting postures with the highest accuracy, we tried seven different DL models taken from three main categories: (1) feedforward artificial neural networks, and we adopted Multilayer Perceptron (MLP); (2) Convolutional Neural Networks (CNN); and (3) Recurrent Neural Networks (RNN), where we adopted a Long Short-Term Memory (LSTM), a bidirectional LSTM (BDLSTM), a CNN-LSTM (CNLSTM), a convolutional LSTM (CVLSTM), and an Echo Memory Network (EMN). All models were trained and tested with the very same datasets collected during the experiment. As mentioned in
Section 2.2.2, datasets have two dimensions: the first dimension is the number of samples (580), the second dimension is the number of sensor signals (8). Since CNN and RNN models require three-dimensional input data, we converted the datasets to frames with 10 samples of 58 time steps of 8 sensors signals. Consequently, the input dimensions of the models were set to [samples, time steps, and features] whose values were specified as being [10, 58, 8]. The results of the comparison among the different models are shown in
Section 3.2.
In feedforward ANNs, or equivalently MLPs, information flows acyclically from the input neurons’ layer to the output one, traversing an arbitrary number of so-called hidden layers. The idea behind MLPs is that the complexity of the nonlinear approximation required to classify arbitrary functional behaviors can be tackled by a supervised learning procedure, whose core is based on the use of the notorious back-propagation algorithm which, formalized in a divide and conquer structure, enables the deployment of very deep networks, as opposed to a wide single hidden layer one that can be solved in a reasonable amount of time. In this way, both classification and regression problems can be successfully tackled with reasonable computing resources [
31]. Besides being acyclic, the MLP model is in general fully connected, which means that each neuron of a given layer receives inputs from all the neurons of the preceding one and provides output to all the neurons of the subsequent one.
Nevertheless, we adopted an MLP model with a single hidden layer of 30 neurons, as shown in
Figure 6. We used the Python scikit-learn library to implement the learn and test procedures. A rectified linear unit (ReLU
[
32]) was used as activation function, while ADAM [
33] was used as stochastic gradient-based optimizer, where the regularization parameter (L2 penalty) introduced to limit the data overfitting was set to 0.7, and the maximum number of iterations (training epochs) was set to 100. The input and output layers consisted of eight neurons each, corresponding to the eight input sensors and to the eight postures, respectively, and we used the softmax activation function for the classification. The Cross-Entropy (log loss) was adopted as the loss function for all the models.
CNNs are specially designed to process data with grid-like topology. The kind of data that can be fed to CNNs is either 1D grids, as in the case of time series where signal sampling produces a data stream where neighboring points (with a proper definition of the neighborhood) are correlated in time, or 2D grids, as in the case of digital images, where instead, spatial correlation occurs among neighboring pixels. CNNs apply successive convolution filters to input, performed using proper numerical kernels (filters). Assuming a 2D problem, hence matrix kernels, these filters are usually significantly smaller than input matrices. Consequently, the CNN elaboration aims at extracting the most meaningful features of the input matrix, thus preventing the overfitting problem with the advantage of lower memory requirements for parameter storage [
31]. We implemented a CNN model using the Keras library and designed it with a 1D convolutional layer of 16 filter kernels with size of 3, followed by a dropout layer for overfitting mitigation and a MaxPooling1D layer for the downsampling. Then, a flattening layer was added and a further fully connected layer of 30 neurons with L2 = 0.7 and ReLU activation for information restoration was stacked, followed by an additional fully connected output layer comprising eight neurons with softmax activation functions.
Figure 7 shows the architecture of the CNN model as described above. The model was fitted and validated in 100 epochs with batch size of 64 input samples.
RNNs are neural networks specialized to process sequential data [
31]. They are networks with memory that allows the network to remember the time sequence of the inputs. The memory is fed by recurrent connections, allowing the hidden cells to have access to their previous outputs [
34].
As mentioned before, we trained and tested five different RNN models: LSTM, BDLSTM, CNLSTM, CVLSTM, and EMN.
LSTM networks overcome the vanishing gradient problem typical of RNNs by avoiding the long-term dependency problem. LSTMs, instead of using neurons, possess memory blocks that are connected to layers by gates (input, output, and forget) that are included to manage the block state and output [
35,
36,
37]. We implemented an LSTM model using the Keras library, featuring an LSTM layer of 200 units followed by a dropout layer and a fully connected layer of 200 neurons using L2 = 0.7 and ReLU activation functions. The output was obtained with a single, fully connected layer containing eight neurons with softmax activation functions.
Figure 8 shows the architecture of the LSTM model. Moreover, this model was fitted and validated in 100 epochs with input batch size of 64 samples.
The BDLSTM [
38] combines an LSTM processing data forward in time starting from the beginning of the sequence with one additional LSTM which instead processes backwards in time starting from the end of the sequence [
31]. We employed a BDLSTM model present in Keras, configured with a bidirectional wrapper comprising an LSTM layer of 200 units, followed by a dropout layer and a fully connected layer of 200 neurons with L2 = 0.7 and ReLU activation functions. The fully connected output layer contained eight neurons with softmax activation functions. This model was also fitted and validated using 100 epochs with batch sizes of 64 samples.
The CNLSTM model [
39] utilizes a CNN to extract important features from the input data, which is combined with an LSTM for sequence memorization. Again, we employed the CNLSTM model present in the Keras library with two time distributed wrapper Layers for the time optimization, including a convolution layer of 64 filters with kernel size
and ReLU activation function, followed by three additional time distributed wrapper layers comprising a dropout layer, a pooling layer for down sampling, and a flattening layer, respectively. An LSTM layer of 200 units was then added, followed by a dropout layer, and by a fully connected layer for reconstruction made of 200 neurons with L2 = 0.7 and ReLU activation function. The output was then obtained by means of a fully connected layer featuring eight neurons and softmax activation functions. As the previous ones, the model was fitted and validated in 100 epochs with batch size of 64.
Ths CVLSTM [
40] uses convolutions to provide inputs to an LSTM. We adopted the CVLSTM model included in the Keras library, in particular its ConvLSTM2D module which is similar to an LSTM but with a convolutional input and feedback transitions. This layer was composed of 16 filters with kernel size 1 × 3, followed by a dropout layer, a flattening layer, and a fully connected layer of 200 neurons with L2 = 0.7 and ReLU activation functions. The output layer was obtained with a fully connected layer of eight neurons with softmax activation functions. Once more, the model was fitted and validated in 100 epochs with batch sizes of 64 samples.
Finally, the EMN [
41] model, designed specifically for time series classification, combines Echo State Network (ESN) and CNN models, exploiting both their peculiar characteristics. EMN is composed by encoding and decoding steps. At the encoding step, memory matrices are created by projecting each input time series frame onto a high-dimensional reservoir state space. At the decoding step, instead, the memory matrices are decoded by CNN convolutional and pooling layers. The full description of the model can be found in [
41]. We employed the EMN model with a reservoir layer as encoding step, which produced the echo memory matrices from the input time series. As per the decoding step, we created eight inputs using the memory matrices of the eight sensors signals. Each input was generated stacking 2D convolutional layer featuring 120 filters and kernel size of 5 × 32, followed by a 2D pooling layer and a dropout one. Then, we implemented data fusion of all the pooled features using a concatenation layer, followed by a fully connected output layer made of eight neurons with L2 = 0.7 and softmax activation functions. In this case, the model was fitted and validated in 100 epochs with batch size of 25.
Figure 9 shows the architecture of the EMN model.
5. Conclusions
In this research study, a smart chair sensors system was designed, realized, and tested with an experiment involving 40 subjects. A large dataset was created with the acquired data. The performance of the designed sensors system was evaluated with seven deep learning models for eight sitting postures classification and secured by k-fold cross validation. Results of all DL models were compared, and the best average accuracy of 91.68% was achieved by an EMN model, obtained in 5-fold, each fold lasting 27 min for computations and to train the 162,248 trainable parameters. The MLP model, instead, achieved the average accuracy of 90.83%, obtained in 5-fold, each fold lasted 3 min for computations and to train 480 trainable parameters. This second one can be considered an appropriate trade-off model for our application in terms of computational cost vs. accuracy.
The sensors system that was designed is innovative, at the same time, simple and versatile. This was obtained thanks to rather low number of deployed pressure sensors, to the optimization of their position, and to the purposely developed PCB. Furthermore, the adopted design is easily applicable to many types of chair and armchair. The signal acquisition board transmits the data via Wi-Fi to the computer for data acquisition and storage without the need of initialization or calibration processes, and since the board is battery powered, no cables are required, thus allowing for the easy repositioning of the chair, while 30 h of continuous operation is at the same time guaranteed with no need of battery recharging. The system can be employed for various applications, such as emotional, behavior and activity identification. Its design is easily extendable to the case of multiple smart chairs to be operated simultaneously and synchronously for larger postures classification experiments.