LU93111B1

LU93111B1 - Floor-based person monitoring system

Info

Publication number: LU93111B1
Application number: LU93111A
Authority: LU
Inventors: Renan Serra; Nicolas Vayatis
Original assignee: Tarkett Gdl Sa
Priority date: 2016-06-16
Filing date: 2016-06-16
Publication date: 2018-01-09
Also published as: WO2017216313A1

Abstract

A person monitoring system comprises a resilient floor covering arranged in a monitored area, one or more pressure sensors installed in or under the resilient floor covering for generating an analog measurement signal and a detector for detecting person-related events in the monitored area. The detector includes an analog-to-digital converter for sampling the analog measurement signal and producing a digital measurement signal, a buffer memory for buffering the digital measurement signal and a processor for processing the digital measurement signal. The processor is configured to extract a feature vector from the buffered digital measurement signal and to map the extracted feature vector into one of at least two categories indicating the occurrence or the absence of a person-related event.

Description

DESCRIPTION

FLOOR-BASED PERSON MONITORING SYSTEM

Field of the Invention [0001] The invention generally relates to a floor-based person monitoring system. Such a system could be used for monitoring persons in rooms of a caretaking institution, such as, e.g., a hospital, a nursing home or a retirement home, or a penal institution while preserving, as much as possible, the monitored persons’ autonomy and privacy.

Background of the Invention [0002] Rooms in healthcare facilities are conventionally equipped with nurse call buttons (or other types of switches), enabling the persons in the rooms to request assistance from the nurses or caregivers. Pressing the call button triggers a signal to the caregivers’ room and possibly to caregivers’ pagers or other mobile devices. Such call systems are useful for responding to ad-hoc needs expressed by the room occupants having the button or the switch in reach. However, they are only partly appropriate in case of emergency, especially in the event of a fall.

[0003] There is, therefore, a need for monitoring systems, which allow, inter alia to give the alarm if a fall of a person in the monitored area is detected.

[0004] US 5 877 675 discloses a portable, three-way wireless communication system that provides a patient with a direct link to a caregiver, as well as a central facility such as a nurse's station. The system is comprised of a patient unit, a caregiver unit and a central station. The patient unit is designed to be small and portable, and can be worn on the patient's wrist or the like. The unit permits a patient to send a request for assistance directly to an assigned caregiver, and provides for two-way voice communications between the patient and the caregiver. The unit stores information associated with the patient, such as identification, medications, attending physician, and the like. The caregiver unit is also portable, and provides two-way voice communications with patients and other caregivers. The unit displays information about each patient to whom the caregiver is assigned. The central station functions as a backup, in the event that a caregiver is not able to timely respond to a call from a patient. In addition, it stores more detailed information regarding patients, which can be accessed by the caregiver via their individual units.

[0005] US 2008/0117060 A1 discloses a system for facilitating independent living of individuals. The system, which is adapted to communicate with one or more caregivers, comprises a worn device fitted with a panic button and an activity sensor, sensors placed in the user’s living area and an off-site monitoring center. A first subsystem monitors the activity levels of the individual and determines whether the activity level is indicative of a decline in the individual's health status; a second subsystem can be selectively activated by the individual to alert caregivers that assistance is needed; a third subsystem automatically alerts caregivers that assistance is needed based at least in part on determination of the first subsystem; and a fourth subsystem monitors whether the individual is exhibiting wandering activity.

[0006] Floor-mounted monitoring systems, are, as such, well known in various applications.

[0007] For instance, US 5 592 152 relates to an intruder detecting device, which is to be installed in an integrated raised flooring system. The intruder detecting device includes a floor panel assembly and a plurality of pedestals for supporting the floor panel assembly on a base floor, a housing, a restoring member, a switch unit and a piston. The restoring member biases an upper housing part away from a lower housing part. The switch unit is mounted on the lower housing part and has a resilient switch contact which can be deactivated by the piston.

[0008] EP 2 263 217 discloses an object tracking system, comprising a dense sensor field in the floor. The object tracking system detects sensor activations and links an object to each activation. It further produces event information describing events for immediate or later use. The system detects events according the conditions defined for them, on the basis of sensor observations. The conditions can relate to the essence of the objects, e.g. to the strength of the observations linked to the object, to the size and/or shape of the object, to a temporal change of essence and to movement. The system can be used e.g. for detecting the falling, the getting out of bed, the arrival in a space or the exit from it of a person by tracking an object with the dense sensor field, and for producing event information about the treatment or safety of the person for delivering to the person providing care.

[0009] US 8 138 882 discloses an electronic multi-touch floor covering that has numerous sensors arranged in a dense two-dimensional array to identify shapes. The electronic multi-touch floor covering identifies the shape of an object that is in contact with the surface of the electronic multi-touch floor covering. An entity record is then retrieved from a data store, such as a database, with the retrieved entity record corresponding to the identified shape. Actions are then retrieved from a second data store with the actions corresponding to the retrieved entity record. The retrieved actions are then executed by the computer system. For instance, if the system detects that the family dog has entered an area that is “off-limits” for it, a notification to the owner can be dispatched in order to have the dog removed from the off-limits location.

[0010] US 6 515 586 relates to a floor covering integrated with a tactile sensory layer so as to form a tactile sensory surface. The tactile sensory layer has a plurality of sensors arranged in a dense two-dimensional array. A controller is connected to the tactile sensory surface to track a person or object. The tactile sensory surface may be flexible and manufactured in bulk on a roll, so that it is adjustable in both length and width.

[0011] US 2006/0171570 discloses a “smartmat” that monitors and identifies people, animals and other objects. Objects are differentiated based on weight, footprint and floor/wall pressure patterns such as footfall patterns of pedestrians and other patterns.

General Description [0012] A person monitoring system according to the invention comprises a resilient floor covering arranged in a monitored area, one or more pressure sensors installed in or under the resilient floor covering for generating an analog measurement signal and a detector for detecting person-related events in the monitored area. The detector includes an analog-to-digital converter for sampling the analog measurement signal and producing a digital measurement signal, a buffer memory for buffering the digital measurement signal and a processor for processing the digital measurement signal. The processor is configured to extract a feature vector from the buffered digital measurement signal and to map the extracted feature vector into one of at least two categories indicating the occurrence or the absence of a person-related event.

[0013] As used herein, the term “feature” designates a characteristic of a signal, obtained by a transformation of the initial set of measured data (i.e. the samples of the digital measurement signal). The initial data points, i.e. the individual samples, are not considered as “features” in the context of the present document. The “analog measurement signal” designates the analog signal that is applied to the input of the analog-to-digital converter (ADC), possibly after signal conditioning (e.g. filtering or smoothing) by electric or electronic components. The “digital measurement signal” is the digital signal fed into the buffer memory, possibly after some basic signal processing in the digital domain (e.g. filtering, interpolating, or the like).

[0014] The ADC is preferably configured to sample the analog measurement signal at a sampling rate (fs) comprised in the range from 50 Hz to 1 kHz, more preferably in the range from 50 Hz to 500 Hz and yet more preferably in the range from 100 Hz to 200 Hz. The resolution of the ADC is preferably at least 8 bits (28 = 256 quantization levels), more preferably at least 12 bits (212 quantization levels) and yet more preferably at least 16 bits (216 quantization levels). Advantageously, the ADC is connected to the pressure sensor via a charge amplifier and a low-pass filter. The low-pass filter preferably has a cutoff frequency of half the sampling frequency in order to satisfy Shannon’s law.

[0015] The buffer memory preferably implements a FIFO (first in first out) buffer, e.g. a cyclic buffer. When the capacity of the FIFO buffer is reached (i.e. when the buffer is full), insertion of a new sample of the digital measurement signal into the FIFO buffer is accompanied with deletion of the oldest sample of the digital measurement signal from the FIFO buffer. The content of the FIFO buffer thus changes at the sampling rate (i.e. the frequency at which new samples of the measurement signal are provided). Preferably, the processor is configured to extract the feature vector from the entire content of the FIFO buffer and to map the extracted feature vector into one of at least two categories. The processor preferably carries out these steps at the sampling rate (fs = 1/Ts), i.e. each time the FIFO buffer is updated, the new feature vector is calculated and a new mapping operation is carried out. Alternatively, the processor may carry out these steps at a lesser rate. With N being the size of the FIFO buffer (i.e. the maximum number of samples it may contain) and fproc the said rate, one has the relationship:

The processor recalculates the feature vector each time n new samples have entered the FIFO buffer. Between each calculation, N-n samples in the FIFO buffer remain

unchanged, which means that the calculations are performed on a “sliding window”. With n = N, the contents of two subsequent windows are disjoint, which may make it difficult to detect the signature of an event beginning in one window and ending in the next or a further one. By increasing the buffer size, one may reduce the probability of unsuitably positioned windows but one cannot completely eliminate that risk. Accordingly, n is preferably chosen such that there is a non-zero overlap between the contents of two subsequent windows, i.e. n < N. More preferably, in terms of duration, the overlap amounts at least approximately to the average recording time that is necessary for correct classification of an event. In practice, this means that N is preferably chosen such that the corresponding duration N/fs is of the order of the duration of an individual event (e.g. in the range from 0.5 s to 20 s, more preferably in the range from 1 s to 10 s and still more preferably in the range from 1 s to 5 s) and n is preferably chosen such that n < 9N/10, more preferably such that n < 3N/4 and still more preferably such that n < N/2. Particularly preferred are the following choices of n: n = 1,2, 3, 4, 5, 6, 7, 8, 9, 10.

[0016] Preferably, the feature vector comprises (as vector components) one or more features computed as specified below.

[0017] Let x denote a vector with components x1,x2, :.,xK> with K being a positive integer : x = (x1,x2> -,Χκ)· In what follows, x could be : o the vector composed of the samples Si contained in the FIFO buffer: x =

o the vector representing the derivative of the digital measurement signal, having the components si = (sî-sm) or (si-Sj-i)/Ts: x = (s1\s2/, with K = N, o the vector representing the (fast) Fourier transform of the digital measurement signal

even number, or M=(N+1 )/2 otherwise, o the vector representing the (fast) Fourier transform of the derivative of the digital measurement signal

with M=N/2 if N is an even number, or M=(N+1 )/2 otherwise, o etc.

[0018] The features could be obtained by using mathematical formulas, such as :

o maximum, i.e. maxfxji = 1, o minimum, i.e. minfxji = 1, o median (or 2nd quartile or 50th percentile), o mean, i.e. μ = ^Σ£=ι*ϊ> o variance, i.e. var = ^-Σ-iilXi-μ|2, ο n-th order central moment (η > 2), i.e. Mn = -Σ*=ι(χζ - μ)η, particularly preferred are the 2nd order central moment (variance), the 3rd and the 4th order central moments, [0019] If x = s (and K = N), the above-defined features will be termed “time-domain” features. If x = T(s) (and K = N/2 or K = (N+1 )/2), they will be referred to as “frequency-domain" features.

[0020] The processor is preferably configured to use supervised classification algorithms on discrete data that are labeled and to create a training base by computing features on them. Preferably, simulated event data are selected to create a training base. Additionally or alternatively, the training base may containreal data recorded onsite. According to a preferred embodiment, both data types are used by the training base.

[0021] Preferably, data gathering is carried out as a continuous task that feeds the algorithms over time and thereby makes them more robust.

[0022] The processor is preferably configured to use a classifier that splits the feature space in sub-spaces corresponding to event labels such as a K-nearest-neighbour (KNN) classifier for instance-based algorithms, support vector machines (SVM) for linear and non-linear classification algorithms, a random forest classifier for interpretable classifiers, a neural network for black box classifiers or a combination of at least two of these classifiers to map the extracted feature vector into one of the at least two categories.

[0023] According to an embodiment, the processor is configured to implement a neural network classifier based upon time and frequency domain features.

[0024] According to a further embodiment of the invention, the processor is configured to implement an instance-based learning algorithm such as K-nearest neighbor (KNN) classifier. This KNN classifier preferably takes a classification decision based upon the time-domain features and upon the frequency-domain features.

[0025] Preferably, the processor is configured to detect one or more of the following person-related events: fall of a person, a person walking (activity), a person entering the monitored area, a person leaving the monitored area, a person leaving a first sub-area of the monitored area and entering a second sub-area of the monitored area. Preferably, the processor is configured to detect all of the above person-related events. More preferably, the processor is configured to detect all events that can be labeled.

[0026] False positive detections by a realistic classifier cannot completely be ruled out, especially if high sensitivity is desired. As known to those versed in the field of the invention, sensitivity (or true positive rate, TPR) is defined as the proportion of positives that are correctly identified as such: TPR = TP/(TP + FN), where TP designates the number of true positive event detections and FN designates the number of false negative event detections (i.e. misses). Specificity (SPC, or true negative rate), on the other hand, is defined as the the proportion of negatives that are correctly identified as such: SPC = TN/(TN + FP), where TN designates the number of true negative event detections and FP designates the number of false positive event detections (i.e. an event was detected but there was, in fact, no event). According to a preferred embodiment of the invention, the processor is configured to carry out a validation of the classification decision (i.e. the mapping of the extracted feature vector into one of at least two categories), which takes into account previous feature vectors and/or previous classification decisions. As will be appreciated, the validation helps to reduce the proportion of false positive detections, which is especially useful if the rate at which the feature vectors are computed and classification decisions are taken (fproc) is significantly smaller than the inverse of the typical duration of the person-related events to be detected. The validation step may be implemented as a hysteresis function: only if a classification decision output by the classifier remains the same during a minimum amount of time it is validated. Using the hysteresis approach, “freak” classifications (outliers), which have a tendency to disappear after a very short time may be efficiently reduced. It is also possible to implement the validation step as a majority decision based on a certain number (L) of previous classifications. For instance, a counter may be used to keep track of the number of positive event detections among the L previous classifications. The processor may be configured to validate a positive event classification only if the count of positive event detections lies above a certain threshold. It will be appreciated that there may be several counters if several types of events have to be detected.

[0027] The validation step may furthermore take into account the relationships between different kinds of events. Certain events may rule out each other. For instance, if it is detected that there is one person in the monitored area and if it is detected that the person is walking, the event “fall of a person” would require that the walking stops. Accordingly, if the walking goes on immediately after the presumed detection of a fall, then there was obviously no fall (or at least no serious fail) of the person. Thanks to the validation step, such an event may then be disregarded or reclassified into a different category (e.g. one collecting “unknown events”). The detector preferably has storage space, wherein one or more files are stored, which indicate the compatible and/or incompatible types of events and/or the circumstances in which the different types of events are compatible or incompatible. These files may further contain rules defining how the processor should handle conflicts (incompatible events).

[0028] Additionally or alternatively, the validation step may comprise computing another feature vector and/or applying another (potentially more computationally costly) classifier in case of a positive event detection.

[0029] As a further option, if the monitored area comprises plural individually monitored sub-areas, the validation step may take into account events detected previously or subsequently in the other sub-area(s). Such a configuration may be especially useful for detection of events occurring in a zone overlapping with more than one sub-area (e.g. a person falling on the boundary of two sub-areas) or of events that occur typically in pairs (end of walking in one sub-area, beginning of walking in an adjacent sub-area).

[0030] The processor may be configured to record (log) the raw data (i.e. the digital measurement signal) all the time. Preferably, however, the processor is configured to record only time intervals during which an event (of any type or of one or more specific types) was detected.

[0031] According to a preferred embodiment of the invention, the one or more pressure sensors are configured to convert variations in pressure exerted thereon into an electric signal that serves as the analog measurement signal. The one or more pressure sensors may e.g. comprise one or more sheet-type pressure sensors, each comprising a ferroelectret polymer film sandwiched between a first electrode layer and a second electrode layer. As used herein, the term “ferroelectret polymer film” designates a cellular polymer film structure that exhibits piezoelectric properties and, more specifically, that generates an electric potential difference and/or a current between the first and second electrode layers applied on its surfaces, in response to pressure being applied on the polymer film structure.

[0032] Preferably, the pressure sensors comprise plural sheet-type pressure sensors arranged in substantially non-overlapping manner in different sub-areas of the monitored area. The different sub-areas may have any suitable size and form. AS each sub-area has a dedicated pressure sensor, it may be individually monitored. The size preferably amounts to any value in the range comprised between 1 cm2 and 100 m2. Particularly preferred are sub-area sizes comprised in the range from 1 dm2 to 20 m2.

[0033] According to a preferred embodiment of the invention, the processor is configured to detect a heartbeat signal and/or a respiration signal in the digital measurement signal and/or to determine a heartbeat rate and/or a respiration rate. Heart beat and/or a respiration signal detection may include, amongst others, low-pass or band-pass filtering the digital measurement signal (to measure heat beats, one could cut off frequencies lower than about 0.7 Hz (42 beats per minute) and frequencies higher than about 3 Hz (180 beats per minute) and to measure the respiration rate, one could cut off frequencies higher than about 0.3 Hz (18 cycles per minute), signal rectification and/or fast Fourier transform (FFT) of the signals thereby obtained.

Brief Description of the Drawings [0034] The accompanying drawings illustrate several aspects of the present invention and, together with the detailed description, serve to explain the principles thereof without limiting its scope. In the drawings:

Fig. 1 : is a schematic drawing of a person monitoring system in a caretaking facility; Fig. 2: is a schematic illustration of the construction of a floor covering;

Fig. 3: is a schematic illustration of an event detector connected to a plurality of sheet-type pressure sensors;

Fig. 4: is a flowchart illustrating the processing of the signals provided by a pressure sensor by an event detector;

Fig. 5: is a collection of several graphs illustrating the evolution in time of several features computed by the event detector.

Detailed Description of Preferred Embodiments [0035] Fig. 1 schematically illustrates a person monitoring system 10 in a caretaking facility, such as, in the present case, a retirement home or a hospital. There are shown a room 12 of a person to be monitored, a caregivers’ room 14 and a hallway or corridor 15 linking those rooms. The retirement home or hospital may, of course, comprise further rooms, but these are not shown for sake of clarity of the drawing. The room 12 comprises a main, bedroom, partition 16 and a smaller, bathroom, partition 18. The room 12 is accessible from the hallway or corridor 15 via an entrance/exit zone 20, which is adjacent the door (not shown) of the room 12. The floor of the room 12 corresponds, in the illustrated example, to the monitored area of the person monitoring system, whereas the main partition 16, the smaller partition 18 and the entrance/exit zone correspond to individually monitored sub-areas thereof.

[0036] The person monitoring system 10 comprises a resilient polymer-based floor covering 22 having installed thereunder plural sheet-type pressure sensors 24. The construction of the floor covering is best illustrated in Fig. 2. The sheet-type pressure sensor 24 is affixed to the floor pavement 26 with a first adhesive layer 28. The resilient floor covering is affixed on the top surface of the sheet-type pressure sensor 24 with a second adhesive layer 30. Also shown in Fig. 2 is a skirting 32.

[0037] The sheet-type pressure sensors 24, which may be configured as flexible tiles, planks, stripes or bands, are arranged substantially without overlap with one another. In each sub-area, the sheet-type pressure sensors 24 are connected in parallel to a detector 34, in such a way that the analog measurement signals originating from different sensors within the same sub-area are not readily discernable by the detector 34. The sensors of a given sub-area are hereinafter referred to collectively as “sensor group”. The different sensor groups, each associated to a different sub-area of the room, are, however, connected individually to the detector 34, whereby it is known which sensor group an analog measurement signal originates from. In the embodiment illustrated in Fig. 1, there is one sensor group for each one of the following sub-areas: 1) entrance/exit zone 20, 2) bedroom partition 16 and 3) bathroom partition 18. Each sensor group is bijectively associated with one sensor channel of the detector.

[0038] Fig. 3 schematically illustrates the detector 34 and how it is connected to the sheet-type pressure sensors 24. Each sheet-type pressure sensor 24 comprises a ferroelectret polymer film 36 sandwiched between a first electrode 38 and a second electrode 40. When the ferroelectret polymer film 36 is compressed, a voltage is generated between the first and the second electrodes 38, 40. That voltage is input to the detector 34, which converts it into a digital signal for further treatment. A first electrically insulating film 42 is arranged on the second electrode 40 and a second electrically insulating film 44 is arranged between the first electrode 38 and a shield electrode 46. A third electrically insulating film 48 is applied on the opposite side of the shield electrode 46. The second electrode 40 and the shield electrode 46 are connected to ground, in such a way as to shield the first electrode 38, which is the signal electrode of the sensor, from external electromagnetic interference. In the illustrated embodiment, the electrodes 38, 40 and 46 are aluminium layers with a thickness of 5 to 20 pm (e.g. 9 pm) each. The ferroelectret polymer film 36 has a thickness preferably comprised in the range from 50 to 100 pm (e.g. 65 pm). The electrically insulating films 42, 44, 48 can be made of PET (polyethylene terephthalate) or any other electrically insulating polymer. Their thicknesses preferably amount to 50 to 250 pm (e.g. 75 pm). The total thickness of the sheet-type pressure sensor 24 thus amounts to less than 1 mm. The signal electrode (first electrode 38) may be patterned by insulating regions, which preferably extend along straight axes. Those regions allow the pressure sensor to be cut to a desired shape with a reduced risk that the cutting will cause short-circuits between the signal electrode 38 and one of the grounded electrodes 40, 46.

[0039] The pressure sensors 24 are connected to the detector 34 by respective coaxial cables 50 comprising each a core conductor 52 and at least one shield conductor 54 surrounding the core conductor 52. The core conductor 52 is connected to the signal electrode 38, whereas the shield conductor 54 is connected to the grounded electrodes 40, 46. The other end of the core conductor is connected to a charge amplifier 56.

[0040] The analog signals output by the charge amplifiers 56 are input to an ADC 60 with multiple sensor channels. The ADC 60 preferably operates at a sampling rate of 100 Hz to 200 Hz (per channel) and with a resolution of at least 8 bits. The digital measurement signals output by the ADC 60 are fed into different buffers 61 and processed by the microprocessor 62. The microprocessor 62 further comprises or is connected to a communication module 64 for connecting the detector 34 to one or more networks (e.g. Ethernet, WiFi™ (IEEE 802.11™ standard), DECT (Digital Enhanced Cordless Telecommunications) and/or a building automation system interface, etc.). It is worthwhile noting that the separation of the sensor channels 59 downstream of the ADC may be effected in software.

[0041] Fig. 4 is a flowchart of the signal processing. The analog measurement signal generated by a pressure sensor 24 is sampled by the ADC 60 and the digital samples are stored at least temporarily in a memory 63 (cf. Fig. 3). The microprocessor is configured to compute a feature vector 66 every time Ts a new sample arrives, using the N most recent samples. As that set changes every Ts, it is termed the “FIFO buffer”. According to preferred embodiments of the invention, fs is selected in the range from 100 to 1000 Hz and N is selected in the range from 100 to 2500. Preferably N and fs are chosen such that N/fs (i.e. the time interval represented by the FIFO buffer) is comprised in the range from 0.5 to 10 s, more preferably in the range from 1 to 5 s, e.g. 2.5 s. Each one of the features (feature vector components) characterizes a specific aspect of the N most recent samples and is obtained at runtime through application of a corresponding formula. The feature vector is input to a classifier, configured to detect whether the feature vector contains the “signature(s)” of one or more predefined, person-related events. The classifier thus assigns the feature vector (and thus the underlying set of N samples) to one of at least two categories indicating the occurrence or the absence of a person-related event (event detection step 68 in Fig. 4). The resulting event class is subject to validation (step 70). For each sensor channel 59-1, 59-2, ... validation 70 takes into account previous a number L (e.g. selected in the range from 10 to 50) of previous classification decisions and primarily aims at increasing specificity (SPC). In the illustrated case, the validation step preferably uses a majority decision based on the L most recent classes issued by the classifier. For each kind of event to be detected and for each sensor channel, a sliding window integrator may be used to keep track of the number of positive or negative classifications in the time interval [t - LTS, t[ (where t is current time). The processor is configured to validate a positive event classification only if the count of positive event detections lies above a certain threshold or if the count of negative event detections lies below a certain threshold. The processor may use auxiliary data 72, such as e.g. rules regarding mutually exclusive event detections or non-detections. For instance, in case of a dense matrix of sensor groups covering each only an area which is small in comparison with the spatial extension of an event (i.e. the size of the region in which an effect of the event should be registered), an event detected by one sensor group but not detected at the same or nearly the same time by any another sensor group is likely due to a false positive detection. For such a pressure sensor configuration, the auxiliary data preferably comprise validation rules, which take the spatio-temporal signal patterns into account.

[0042] Event detection and validation may be implemented as separate steps (as illustrated), as that may facilitate understanding and fine-tuning the behaviour of the detector during the training procedure. However, event detection and validation could also be fused into a single process carried out by a more sophisticated classifier.

[0043] A simplified example of an algorithm for fall detection is described with reference to Fig. 5, which shows a short portion of a digital measurement signal measured when a person fell with a monitored area (upper left graph) as well as the evolution of several features. The fall occurred after about 10 s of walking activity (another type of event that may be detected). The sampling rate fs was 100 Hz. The feature vector comprised the following component features (calculated with N = 250): o Time-domain energy

o time-domain variance

o time-domain 10th order central moment

o frequency-domain variance

o median frequency, o frequency-domain skewness, o frequency-domain log energy entropy, o frequency-domain interpercentile range (between the 35th and 65th percentiles).

[0044] The microprocessor used a threshold-based classifier. For each feature vector component i, a threshold thn and a weight factor Wi were determined in a training phase. For each feature vector (i.e. for each time step), the microprocessor computes: — Σfeatures where bi is 1 if feature i > thn and 0 else. If W exceeds a global threshold, the classifier classifies the feature vector as indicative of a fall event (positive detection), otherwise it classifies the feature vector as a non-fall event (negative detection). As indicated above, a subsequent validation step is carried out in order to rule out isolated positive detections.

[0045] It may be noted that the bi are not necessarily binary variables. In a variant of the above algorithm, a value from 0 to 1 is assigned to each feature depending on how well it meets a corresponding criterion.

[0046] Turning back to Fig. 5, it can be seen that most of the above-described features show significant peaks in the interval between t = 10 and t = 14. Assuming that appropriate thresholds and weight factors were determined during the training of the classifier, the latter should have had no difficulty in detecting that particular fall event.

[0047] Other events detected by the microprocessor may be the start of activity (walking) and the end of the same. Walking activity detection may e.g. be effected by the microprocessor comparing the energy parameter with a threshold.

[0048] Preferably, each time the microprocessor 62 detects an event (or a new event), it assembles a datagram or data packet containing at least the ID of the detector 34, an identifier identifying the detected event, an identifier identifying the sub-area in which the event occurred and a time stamp indicating at what time the event was detected. Optionally, further indications may be included into the datagram or data packet, such as e.g. the parameters that led to the detection of the identified event. If several events occurred shortly one after the other, they may either be included into separate datagrams or data packets or grouped into one datagram or data packet.

[0049] The datagrams or data packets may be transmitted, e.g., to a cloud server 78 (Fig. 1), via the communication module. The microprocessor 62 preferably keeps a local copy of any detected event in a buffer memory. The size of the buffer memory may be such that several days of data may be stored therein, in case of an interruption of the communication links between the detector and the cloud server.

[0050] As illustrated in Fig. 1, the detector is connected with the nurse or caregiver call system of the caretaking facility. Each room 12 is equipped with a nurse call button 86, which is typically arranged in such a way that the room occupant can reach it from their bed. In its basic configuration, actuation of the nurse call button closes an electrical circuit, which activates an audible and visual alarm signal in the caregivers’ room 14. In this case, one of the relays 72, 74 of the detector 34 is connected in parallel to the nurse call button 86 in such a way that the microprocessor 62 can control the electrical circuit that gives the alarm. If the retirement home or hospital comprises a more modern nurse or caregiver call system, the sensor control system may be interfaced therewith via the communication module 64. When the microprocessor 62 detects a fall of the room occupant 88 (as illustrated in Fig. 1), it triggers an alarm via the caretaking facility’s nurse or caregiver call system. If the nurse or caregiver call system can deal with it, an emergency code, possibly indicating that the occupant has fallen, is sent as well, in order to communicate the urgency of the need for assistance. The detector 34 may further be interfaced with LEDs integrated in the skirting 32 of the room 12. When a fall is detected, the microprocessor 62 may send a control signal to the LEDs causing them to generate a visual signal (e.g. blinking or flashing) informing the room occupant that the fall has been detected and the alarm has been given. If the retirement home or hospital’s nurse or caregiver call system features bi-directional communication, the microprocessor 62 may also inform the room occupant that the caregivers have acknowledged receipt of the alarm by emitting a second visual signal.

[0051] Giving feedback to the room occupant that their fall has been detected and that help is underway has the potential to greatly reduce psychological stress in case of a fail. It may furthermore somewhat reduce the room occupant’s fear from getting up at nighttime. It is worthwhile noting that a knocking code (a predefined sequence of knocks and shorter or longer pauses) may be communicated to the room occupant in case they fall and the fall is not detected by the monitoring system. In this case, the microprocessor is configured to identify the knocking code (as a particular kind of events) in the analog measurement signals and to trigger an alert in case of a positive detection. If the fallen person receives no visual feedback from the skirting that the fall has been detected, they may manually trigger the alarm by knocking the predefined knocking code into the floor.

[0052] The microprocessor may optionally be configured to trigger an alarm only if the fall is not followed, with a short period (e.g. in the range from 30 s to 2 minutes), by regular walking activity indicating that the person is able to move. In this case, the fall may be registered without leading to (immediate) intervention by the retirement home or hospital personnel.

[0053] The cloud server receiving the transmitted datagrams or data packets preferably stores the data contained therein (event information and any accessory data) in the database. The data are preferably stored as received. The cloud server may further convert the received data into discrete-time data reflecting the state of the monitored room at discrete times so as to generate a timeline. Each timeline is preferably tied to a detector ID in the database, which permits to look up the status of each room at any past time. The cloud server may further analyze the distribution of the detected events in time, carry out frequency analysis thereon, compute crosscorrelations between different events, etc.

[0054] The following table summarizes results obtained using different types of classifiers:

[0055] While specific embodiments have been described herein in detail, those skilled in the art will appreciate that various modifications and alternatives to those details could be developed in light of the overall teachings of the disclosure. Accordingly, the particular arrangements disclosed are meant to be illustrative only and not limiting as to the scope of the invention, which is to be given the full breadth of the appended claims and any and all equivalents thereof.

Claims

1. A person monitoring system, including a resilient floor covering arranged in a supervised area; one or more pressure sensors installed in or under said resilient flooring to generate an analog measurement signal; a detector for detecting events related to a person in the monitored area, the detector including an analog-to-digital converter for sampling the analog measurement signal and producing a digital measurement signal, a buffer for buffering the digital measurement signal and a processor for processing the digital measurement signal; wherein the processor is configured to extract a feature vector of the buffered digital measurement signal and to map to the extracted feature vector one of at least two categories indicating the occurrence or absence of event related to a nobody.

The personal surveillance system as claimed in claim 1, wherein the buffer implements a FIFO buffer, wherein when the FIFO buffer capacity is reached, inserting a new sample of the digital measurement signal. in the FIFO buffer is accompanied by the deletion of the oldest sample of the digital measurement signal of the FIFO buffer.

The human surveillance system as claimed in claim 2, wherein the processor is configured to extract the feature vector from the entire contents of the FIFO buffer and to map to the feature vector retrieves one of the at least two categories at the sampling frequency.

The personal surveillance system of any one of claims 1 to 3, wherein the processor is configured to use a classifier that divides the feature space into subspaces.

The person monitoring system according to any one of claims 1 to 3, wherein the processor is configured to use a "nearest neighbor K" classifier to map to the extracted feature vector one of the at least two categories.

The human surveillance system according to any of claims 1 to 3, wherein the processor is configured to use a random forest classifier to match the extracted feature vector of at least two categories.

The personal surveillance system according to any one of claims 1 to 3, wherein the processor is configured to combine at least two of the following classifiers: "nearest neighbor K" classifier, "threshold" classifier and classifier " random forest "to match the extracted feature vector one of the at least two categories.

The personal surveillance system according to any one of claims 1 to 7, wherein the processor is configured to detect one or more of the following person-related events: fall of a person, a person walking, a person entering the guarded area, a person leaving the guarded area, a person leaving a first sub-area of the guarded area and entering a second sub-area of the guarded area.

The person monitoring system according to any one of claims 1 to 8, wherein the one or more pressure sensors are configured to convert the pressure variations exerted thereon into an electrical signal that serves as a measurement signal. analog.

The human surveillance system as claimed in claim 9, wherein the one or more pressure sensors comprise one or more sheet-shaped pressure sensors, each comprising a ferroelectric polymer film between a first electrode layer. and a second electrode layer.

The person monitoring system as claimed in claim 10, wherein the one or more pressure sensors comprise a plurality of sheet-shaped pressure sensors arranged in different sub-areas of the monitored area so as not to substantially to overlap.