Keywords

1 Introduction

The modern office worker spends a considerable amount of time sitting on a chair in front of a screen. The popularity of this style of work has led to the growth of related health issues in recent decades. Most notably lower back pain is a major risk factor that causes considerable economic impact in many western countries [1]. Supporting the office worker by providing more ergonomic seating, as well as encouraging physical exercise in the office can alleviate some of this impact [2]. In this work we present the Capacitive Chair. This office chair is equipped with eight capacitive proximity sensors that are able to detect body posture, respiratory rate and work activities. In order to detect a high number of poses, the electrodes have to be applied on the chair in a suitable layout. To demonstrate potential adaptations to common chair features we use a variety of different electrode materials and shapes, including one conductive thread electrode integrated into the mesh of a back rest. This thread electrode reacts to both presence and geometric deformation, which makes it suitable to detect subtle movements, such as respiration.

We envision three different use cases that can be supported by the Capacitive Chair. The first is the tracking of working situations, based on the current level of movement. We can determine if the chair is occupied and how active a person has been moving on it. The second use case is a posture recognition that recognizes different common poses on a chair. The last use case is breathing rate recognition, based on analyzing the chest movement. We have performed two different studies. The first tracked the working activity of a single user throughout a week. The results are activity graphs that show active phases, inactive phases and periods of not being at the chair. In a second evaluation we were testing the posture recognition of the chair in a short study with 10 participants. This work is extended from internal technical reports and project reports. Excerpts and intermediate results can be found in [3, 4].

2 Related Works

In the previous years there have been various attempts to attach different sensors to pieces of furniture, in order to provide additional functionality or provide contextual information for other systems in the environment.

The Health Chair by Griffiths et al. uses pressure sensors in the backrest to detect the respiratory rate from chest movement and EKG electrodes on the armrests to measure the heart rate [5]. Based on a pre-study with several hundred users focusing on common sitting postures, they have optimized their data processing for detecting physiological signals of users in those postures. They conducted a study with 18 users that were taking different poses on an equipped chair. The Health Chair was able to detect the heart rate in 32 % of the cases and respiratory rate in 52 % of the time.

The Smart Bed is a system created by our group that integrates capacitive proximity sensors into a bed [6]. A first algorithm detects the posture of one or two persons on the bed and the estimated stress distribution on the spine, using a combination of pressure and presence. This data is acquired by attaching flexible electrodes to a slatted frame. In an extended work a movement-based algorithm was added that evaluates the current sleep phase [7]. Based on the variance of sensor signals over time the extent of user movement can be estimated. This is correlated to the current sleep phase. This method was adapted for detecting the work activity levels of the Capacitive Chair. It will be briefly discussed in the following sections.

Another similar prototype is the Smart Couch [8]. This system by Grosse-Puppendahl et al. uses capacitive proximity sensors integrated into a regular couch to recognize different postures of one or two occupants based on a machine learning classifier. A similar approach is used for the posture recognition in this paper. We extend on this work by providing an integrated layout for capacitive proximity sensors in a chair and a classification method that is focused on the requirements of one person on one chair.

3 Layout for Capacitive Sensors in an Office Chair

The layout of capacitive proximity sensor systems should be driven by the requirements of the chosen applications. In our case there are three measured properties - posture, activity and breathing rate. The requirements for the first two are similar. Posture can be determined by the position of body parts relative to the sensors, while activity is calculated based on the change in position of body parts relative to the sensors. Therefore, a high number of sensors should be reserved for placement close to the body parts performing most of the movement - the limbs. As the office chair has no parts around the lower legs or feet the leg movement is instead indicated by a sensor in front of the seat area that detects presence of the upper legs. The chair has two armrests that can be equipped to estimate arm positions and activity. In order to detect postures that are similar but vary on orientation of the user the weight distribution of the body should be identified - whether it is more on the left or right side. For this purpose two electrodes are added in the back area of the seat. This is suitable as most of the body weight is put in this area and variations can be easily identified.

To detect how far an occupant is leaning to the front several electrodes should be attached in the back rest. Here, the third requirement comes into play. The detection of the respiratory rate is based on the movement of the chest. As it is beneficial if electrodes are approximately the size of the object to be detected, the electrodes are not evenly spaced - instead we are using a larger electrode close to the chest. This leads to the final electrode layout shown in Fig. 1. (1) electrode on the upper part of the backrest (covered by faux leather), (2) electrode in the central part of the backrest (using conductive thread), (3) electrode in the lower part of the backrest (covered by faux leather), (4) electrode below the right armrest (5), electrode below the left armrest, (6) electrode for the left hip area below the left part of the seat, (7) electrode for the right hip area below the right part of the seat, and (8) electrode for detecting both legs below the front part of the seat.

Fig. 1.
figure 1

Electrode layout of the capacitive chair

Full size image

4 Data Processing

In this section we describe three different data processing methods that allows us to detect activity levels, recognize the posture and acquire a breathing rate from a single sensor.

4.1 Activity Level Detection

Detecting a general activity level during work is a useful measure for long term data analysis and in an aggregated form allows an organizational oversight of how workers behave during a regular work day. Thus it may inform changes in organization of work if undesired activity levels are detected. The individual user is getting a feedback on his activity. This may prompt a more active style of work with less sitting and increased activity on and off the chair. It is foreseeable that this could be combined with different gamification approaches, e.g. to facilitate activity competition between workers of a group.

The chosen method is similar to an approach we previously used for sleep detection. The variations in subsequent sensor readings are used to determine an activity level. We will at this point just give a shortened description of this process with a thorough description being found in Djakow et al. [7]. The following three steps are performed:

  1. 1.

    Acquire a time-series of sensor readings and their variations

  2. 2.

    Detect movement if aggregate or individual variation exceeds a certain threshold

  3. 3.

    Analyze the frequency of movements over time and associate them to a certain activity level

For the Capacitive Chair we are considering three activity levels. The first is “not at chair”, whereas no occupancy is recorded on the chair. This usually occurs during lunch break or certain meetings. The second level is “active work”, such as writing and typing. The third level is “inactive work” that does not involve movement, such as reading screen content. The tasks of writing and typing are indicated by considerable movement of the hands and arms. Thus, we can use the electrode layout to our advantage and prefer readings from the sensors attached to the arm rests. We tested two approaches - the first explicitly favored variation from armrest sensors, while the second implicitly uses the sensors with most significant variation. The first method is preferable if we only want to detect writing and typing, as it will ignore movements of other body parts. The second method considers activity whenever any sensor has high variations. We observed that typing and writing often lead to selection of armrest sensors. However, since we wanted to capture more styles of active movement of the chair the second method is used in this work.

4.2 Posture Recognition

Support vector machines (SVM) are a supervised learning method that is primarily used for linear classification of n-dimensional features [10]. They are clustering data by calculating a hyperplane from training data that maximizes the distance from the closest features. A fast learning method is using sequential minimal optimization and was proposed by Platt [11]. The algorithm requires normalized values. A dynamic normalization algorithm constantly analyzes the sensor data for minimum and maximum values and accordingly calculates the normalized value. There is a variety of different software frameworks for machine learning that support training and recall of SVMs, thus there is no need for reimplementing these methods. As there is an implicit weighting of features according to significance, there is no need to preprocess or weigh the sensor data. The training data is collected from a set of persons that have a significant variance in body shapes in both height and girth. SVMs support an arbitrary number of different groups for classification. However, the number of significant poses on a chair is limited. The Global Posture Study by office furniture manufacturer Steelcase Inc. analyzes the most common poses with a focus on information consumption from modern technical devices, such as smart phones or tablets [9]. The different postures are shown in Fig. 2. A capacitive office chair should be able to distinguish most of these poses if training data has been collected from a sufficiently large number of suitable candidates. Additionally, using sensors clearly positioned on a certain side of the chair, e.g. the armrest, it is possible to associate directional varieties of the asymmetric postures.

Fig. 2.
figure 2

Selected set of postures from global posture study and own gestures. From top left to bottom right: The strunch, the draw, the smart lean, the take it in, upright, no person (first four taken from [9])

Full size image

The processed values of all sensors are compared to previously trained sitting positions of a user. The position with the lowest deviation is considered the current posture.

4.3 Detection of Breathing Rate

The volume changes of the chest while breathing have been a topic of research for a long time [12]. If the body of a person is not moving and can be considered at a static distance from a capacitive proximity sensor, the chest movement should translate into a periodically changing sensor value. The breathing rate detection is operating on a single electrode that is placed close to the chest. The basic concept is shown in Fig. 2. The surface of the electrode is large and close to the surface. Therefore, it is able to pick up the chest movement. Two different methods of data processing are used and fused to get the final breathing rate. Using a fast Fourier transformation the signal is transformed into the frequency space. We are looking for significant signal portions in frequency areas that can be associated to breathing, between 0.1 Hz and 3 Hz. The above Fig. 3 shows an example of the sensor data curve generated by the conductive thread sensor behind the back of a person. The chest movement is clearly visible as sinusoidal oscillation of the sensor value. If there is a sufficiently stable baseline, the zero-crossings can be calculated. However, as this can’t be guaranteed in all cases an adaptive baseline should be used that is reconfigured according to changing states of the sitting person.

Fig. 3.
figure 3

Chest movement when breathing in and out

Full size image

5 Capacitive Chair Prototype

As our basis we are using a regular office chair that was acquired from a local furniture store for approximately €40. Its backrest has a mesh texture in the middle and two non-movable armrests on each side. The electrode locations are chosen according to the layout present in Sect. 3. It is outlined in Fig. 4. Electrodes 1 and 3 are created from flexible copper foil suspended and held in place by duct tape that is attached to the aluminum frame of the backrest. Electrode 2 is created by weaving conductive thread through the mesh and fixing it on two sides using conductive copper tape. This structure is shown in Fig. 5. Electrodes 4 and 5 are also based on copper foil, yet fixed below the hard plastic of the armrests. In order to avoid signal attenuation from the aluminum pipe in the armrest, they are cut to be wider and cover the whole lower structure. Electrodes 6, 7 and 8 are made from single layer PCB boards that are glued to the bottom of the seat area.

Fig. 4.
figure 4

Capacitive chair electrode positions on actual prototype

Full size image
Fig. 5.
figure 5

Capacitive electrode based on conductive thread woven through the mesh of the chair

Full size image

The different electrodes of the Capacitive Chair are connected to a single OpenCapSense board that supports eight channels [13]. This board additionally supports various pre-processing steps. In this case we use a floating average filter to attenuate high-frequency noise. OpenCapSense provides non-normalized integer values that are sent to a connected PC through a USB cable. Any further processing is done on this PC. Normalization is used before all further steps. The Activity Level detection and Posture Recognition use these normalized value for their processing, while the Breathing Rate recognition performs a FFT with a sliding window of 20 s and a 200 ms overlap. This provides sufficient information in the required range between 0.1 Hz and 3 Hz. The most significant amplitude in this interval is considered the user’s breathing rate.

6 User Studies

We have performed two different studies. The first tracked the working activity of a single user throughout a week. The results are activity graphs that show active phases, inactive phases and periods of not being at the chair. An example day activity graph is shown in Fig. 6. The test user was asked to annotate specific activities throughout the day. The graph shows the hourly distribution of the three phase. Notable are a very inactive phase at the beginning of the workday (green), in this case caused by reading emails. At the middle of the work day the lunch break is visible as “not at chair” period (grey). However, the distinction between active and inactive work often is not closely correlated to noted activities. In general even fairly active work periods only resulted in a low percentage of active work. The method has to be either tuned to individually set different activity levels on a per-person basis or consider other data sources, such as sensors attached in different places of the work environment or monitoring tools installed on the work computer.

Fig. 6.
figure 6

Work activity throughout a day

Full size image

The aggregated values enable self-monitoring of levels of activity throughout the week. This can be used as feedback to employ a more active lifestyle or in an aggregated fashion by management to facilitate a more active style of work, e.g. by using micro break programs.

In a second evaluation we were testing the posture recognition of the chair in a short study with 10 participants. Our system was tuned to distinguish three poses from the global posture study and a non-pose: sitting upright, the strunch, take it in, and close to chair. The latter is defined by providing some signal due to presence without actually sitting on the chair.

Before the study a set of training samples was collected by two of the developers and used to train the system. The data collected from the study participants was not used in the training process, in order to test a pre-trained system.

The persons were given a short introduction, in which the different postures were displayed. After that the participants were asked to perform the postures in order. When testing “close-to-chair” the subjects were asked to rattle at the chair, stand close, move it around and thus disturb the potential sensor readings. Each class was tested for 10 s, collecting 200 samples. Overall the results were very convincing. Of the 40 different measurements series only two were not achieving 100 % accuracy. The upright and “close to chair” positions were classified correctly for all candidates. A single candidate had an 86 % rating on the strunch posture. A different candidate had a 55 % rating on the “take it in” position. The average of correctly classified postures is 98.5 %.

The following Table 1 shows the percentage of correct samples for each posture of all participants.

Table 1. Results of posture recognition test in percent
Full size table

7 Conclusion and Future Work

On the previous pages we have presented the Capacitive Chair - a regular office chair equipped with eight capacitive proximity sensors that is able to detect presence, posture, activity level and breathing rate of its occupants. The sensor system is invisibly integrated into the chair and uses variations of electric fields to determine the physiological data over a distance. We have shown that a combination of different electrode materials and shapes that are attached to a single evaluation board for capacitive proximity sensors provide sufficient information about presence and proximity of occupants to gather physiological data. The system was installed in a prototype and evaluated during a long term study for activity level monitoring at work, as well as a test on the precision of the posture recognition method with ten subjects. The activity levels correlate with the noted activities, however active and inactive work are difficult to distinguish using the chair alone. The posture recognition achieved a very high overall classification rate of 98.6 %. However, the set of postures is limited to only four. A test with more postures might require modifications to our method or adding a higher number of sensors.

As future work we are planning to improve the range of recognized postures by collecting a more comprehensive set of training data from various subjects. Another approach is to use the proximity data, in order to allow a more fine-grained posture recognition, e.g. to enable tracking of office chair exercises. This can create various applications in the domain of applied ergonomics. We are working on expanding our breathing rate recognition to also detect the heart rate of occupants. However, as the movement of the heart muscle within the chest is weak compared to the chest movement while breathing this may require higher precision sensors and higher data acquisition frequencies. Notably, the breathing rate recognition wasn’t evaluated in this work, which is planned in the immediate future.

There are a few factors related to the hardware that can be improved. A major point to improve is the restriction to a single chair. It took approximately 12 h to equip a single chair with our sensing system. Ideally, the electrodes and sensors should be able to be installed as a kit that can be easily attached to various different types of chairs. Also, the system has a comparatively high energy consumption of about 200 mA. Therefore, it is currently wired and should be modified for reasonable battery powered application.