WO2015127491A1

WO2015127491A1 - Monitoring system

Info

Publication number: WO2015127491A1
Application number: PCT/AU2014/000171
Authority: WO
Inventors: Ingrid Zukerman; Masud MOSHTAGHI; Robin Andrew RUSSELL
Original assignee: Monash University
Priority date: 2014-02-25
Filing date: 2014-02-25
Publication date: 2015-09-03

Abstract

A monitoring system for monitoring activity of a user in a physical environment having one or more regions, the monitoring system including one or more sensors for detecting movement of the user in each region; a data logger for logging each sensor event, the time of the event and the region in which movement was detected; and a user inactivity detector configured to determine, for each region, whether the time elapsed since a last detected event is an outlier in a distribution of transition times between consecutive detected sensor events for that region, the outlier being indicative of abnormal user inactivity.

Description

MONITORING SYSTEM

Technical of the Field

[0001] The present invention relates generally to a monitoring system and method of monitoring activity of a user in a physical environment. The invention has particular application in the automatic monitoring of the elderly and infirm in their living environment and the invention will be described in relation to that exemplary application. It is to be appreciated however, that the invention can be used in other applications, for example the monitoring of workers in a dangerous industrial environment or the monitoring of prisoners or detainees in a highly secure environment. Other possible applications of the present invention relate to the monitoring of children, domestic animals, livestock, or the like.

Background of the Invention

[0002] The aging of the world's population, coupled with the desire of elderly people to maintain their independence, underscores the need for in-home monitoring systems that help elderly people remain safely in their houses.

[0003] Many existing in-house monitoring systems for seniors rely upon transmitters which can be attached to a wheelchair or can be worn around the neck, wrist or on a belt of a senior person. If the senior person falls, the transmitter is configured to transmit an alert signal to a carer or other emergency contact. Such systems have only found minimal acceptance in the community, principally due to their intrusiveness.

[0004] Other in-home monitoring systems have been proposed that rely upon a network of non-intrusive sensors to detect activity of a monitored person. Such systems generate an alert which is sent to carers whenever unusual activity patterns are detected. Whilst such systems are less intrusive than systems requiring transmitters to be worn, to date systems that rely upon detecting deviation of a user's behaviour from standard activity patterns are rule-based, rigid in their behaviour, error-prone and generate an unacceptable number of false alerts. [0005] It would be desirable to provide a monitoring system and method for monitoring activity of a user in a physical environment which alleviates or overcomes one or more inconveniences or disadvantages of known monitoring systems and methods. It would also be desirable to provide a monitoring system and method for monitoring activity of a user in a physical environment that is simple and easy to use, and is less prone to errors than known monitoring systems and methods.

[0006] The discussion of documents, acts, materials, devices, articles and the like is included in this specification solely for the purpose of providing a context for the present invention. It is not suggested or represented that any or all of these matters formed part of the prior art base or were common general knowledge in the field relevant to the present invention as it existed before the priority date of each claim of this application.

Summary of the Invention

[0007] According to a first aspect of the invention there is provided a monitoring system for monitoring activity of a user in a physical environment having one or more regions, the monitoring system including

one or more sensors for detecting movement of the user in each region; a data logger for logging each sensor event, the time of the event and the region in which movement was detected; and

a user inactivity detector configured to determine, for each region, whether time elapsed since a last detected event is an outlier in a distribution of transition times between consecutive detected sensor events for that region, the outlier being indicative of abnormal user inactivity.

[0008] In one or more embodiments of the monitoring system, the user inactivity detector is configured to

model at least part of the distribution of transition times between consecutive detected sensor events for each region,

for each model, determine an upper threshold for transition times indicative of unusual amounts of user inactivity, and

determine when time elapsed since a last detected event exceeds the upper threshold for each model.

[0009] In one or more embodiments of the monitoring system, the user inactivity detector is further configured to separately model different temporal intervals during a recurring time period, wherein user activity typically exhibits consistency during the same temporal interval of different time periods.

[0010] In one or more embodiments of the monitoring system, the user inactivity detector is further configured to

determine whether parameters of the model can be estimated from the detected events, and if there is no data to estimate the upper threshold for a particular temporal interval in a region,

set the upper threshold for that temporal interval and region by calculating a function of thresholds from temporally proximate intervals or similar regions.

[0011] In one or more embodiments of the monitoring system, the user inactivity detector is further configured to

determine whether parameters of the model can be estimated from the detected events, and if there is not enough data points to estimate the upper threshold for a particular temporal interval in a region,

set the upper threshold for that temporal interval and region to a maximum of the observed data points.

[0012] In one or more embodiments, the user engages in specific behaviour in each region.

[0013] In one or more embodiments of the monitoring system, the distribution of transition times between consecutive detected sensor events is modelled as a long-tailed distribution.

[0014] In one example, the user inactivity detector is configured to model the distribution as a Pareto distribution.

[0015] In another example, the user inactivity detector is configured to model the distribution as a hyperexponential distribution.

[0016] The user inactivity detector may be configured to model the tail of the hyperexponential distribution by selecting a partition of data by accumulating transition times from the largest sample to the smallest sample until an estimated Coefficient of Variation is greater than or equal to 1 ; and fitting an exponential distribution to that partition of data.

[0017] The user inactivity detector may alternatively be configured to model the tail of the hyperexponential distribution by selecting a partition of data by accumulating bins from a histogram of the transition times starting from the bin with largest value to that with the smallest value until the estimated Coefficient of Variation from these bins becomes greater than or equal to 1 ; and fitting an exponential distribution to that partition of the data.

[0018] The user inactivity detector may alternatively be configured to model the tail of the hyperexponential distribution by direct estimation of the exponential distribution in the tail from the histogram of the data by identifying the bins that are not affected by the exponential distribution(s) at the head of the hyperexponential distribution, and directly estimating an exponential distribution using the density values estimated from these unaffected bins.

[0019] In the example described in the preceding paragraph, the bins may be identified by pinpointing the bin in the histogram that produces the largest change in the mean of the exponential distribution estimated from that bin onwards compared to the mean of the exponential distribution estimated from the previous bin.

[0020] In one or more embodiments, the user inactivity detector is further configured to set the upper threshold for each model to no lower than a minimum preset threshold value.

[0021] In one or more embodiments of the monitoring system, the user inactivity detector is further configured to derive an alert line by connecting the thresholds for a particular region throughout the recurring time period; and adjust inconsistent thresholds by applying a smoothing function to the alert line.

[0022] In one or more embodiments of the monitoring system, the user inactivity detector is further configured to adjust the upper threshold for each model by decreasing unreachable thresholds to their maximum useful value.

[0023] In one or more embodiments of the monitoring system, the user inactivity detector is further configured to adapt each model by applying a weighting to recorded transition times, wherein a greater weighting is applied to more recent transition times than to older transition times.

[0024] In one or more embodiments, the monitoring system further includes an alert generator configured to employ one or more decision procedures to decide whether an alert message should be sent in response to detection of an outlier in the logged sensor events, and, in the event that an alert message is to be sent, to determine an intended recipient of the alert message and the urgency of the message.

[0025] In one or more embodiments of the monitoring system, the decision procedures include a decision theoretic algorithm which selectively generates the alert message based on preferences of the user, uncertainty associated with the information regarding the user, and cost of different actions.

[0026] In one or more embodiments of the monitoring system, the decision theoretic algorithm includes one or more Bayesian Decision Networks (BDNs).

[0027] In one or more embodiments of the monitoring system, the parameters of the BDNs are elicited from stakeholders or experts or based on the preferences of similar user groups or a combination of these techniques.

[0028] In one or more embodiments of the monitoring system, the BDNs are configured to take into account the effect of previous decisions on subsequent decisions.

[0029] Another aspect of the invention provides a method of monitoring activity of a user in a physical environment having one or more regions, the method including the steps of

detecting movement of the user in each region by one or more sensors; logging each sensor event, the time of the event and the region in which movement was detected in a data logger; and

determining, for each region, whether time elapsed since a last detected event is an outlier in a distribution of transition times between consecutive detected sensor events for that region, the outlier being indicative of abnormal user inactivity.

[0030] In one or more embodiments of the method, the outlier detection includes the steps of: modelling at least part of the distribution of transition times between consecutive detected sensor events for each region; for each model, determining an upper threshold for transition times indicative of unusual amounts of user inactivity; and determining when time elapsed since a last detected event exceeds the upper threshold for each model.

[0031] In one or more embodiments of the method, the method further includes the step of: separately modelling different temporal intervals during a recurring time period, wherein user activity typically exhibits consistency during the same temporal interval of different time periods.

[0032] In one or more embodiments of the monitoring system, the method further includes the steps of:

determining whether parameters of the model can be estimated from the detected events, and if there is no data to estimate the upper threshold for a particular temporal interval in a region,

setting the upper threshold for that temporal interval and region by calculating a function of thresholds from temporally proximate intervals or similar regions.

[0033] In one or more embodiments of the monitoring system, the method further includes the steps of:

determining whether parameters of the model can be estimated from the detected events, and if there is not enough data points to estimate the upper threshold for a particular temporal interval in a region,

setting the upper threshold for that temporal interval and region to a maximum of the observed data points.

[0034] In one or more embodiments of the method, the distribution of transition times between consecutive detected sensor events is modelled as a long-tailed distribution.

[0035] In one example of the method, the distribution is modelled as a Pareto distribution.

[0036] In another example of the method, the distribution is modelled as a hyperexponential distribution.

[0037] The tail of the hyperexponential distribution may be modelled in the method by: selecting a partition of data by accumulating transition times from the largest sample to the smallest sample until an estimated Coefficient of Variation is greater than or equal to 1 ; and fitting an exponential distribution to that partition of data.

[0038] The tail of the hyperexponential distribution may alternatively be modelled in the method by selecting a partition of data by accumulating bins from a histogram of the transition times starting from the bin with largest value to that with the smallest value until

the estimated Coefficient of Variation from these bins becomes greater than or equal to 1 ; and fitting an exponential distribution to that partition of the data.

[0039] The tail of the hyperexponential distribution may alternatively be modelled in the method by direct estimation of the exponential distribution in the tail from the histogram of the data by identifying the bins that are not affected by the exponential distribution(s) at the head of the hyperexponential distribution, and directly estimating an exponential distribution using the density values estimated from these unaffected bins.

[0040] In this case, the bins may be identified by pinpointing the bin in the histogram that produces the largest change in the mean of the exponential distribution estimated from that bin onwards compared to the mean of the exponential distribution estimated from the previous bin.

[0041] In one or more embodiments, the method further includes the step of setting the upper threshold for each model to no lower than a minimum preset threshold value.

[0042] In one or more embodiments, the method further includes the step of: deriving an alert line by connecting the thresholds for a particular region throughout the recurring time period; and adjusting inconsistent thresholds by applying a smoothing function to the alert line.

[0043] In one or more embodiments, the method further includes the step of: adjusting the upper threshold for each model by decreasing unreachable thresholds to their maximum useful value.

[0044] In one or more embodiments, the method further includes the step of: adapting each model by applying a weighting to recorded transition times, wherein a greater weighting is applied to more recent transition times than to older transition times.

[0045] In one or more embodiments, the method further includes the step of: employing one or more decision procedures to decide whether an alert message should be sent in response to detection of an outlier in the logged sensor events, and, in the

event that an alert message is to be sent, to determine an intended recipient of the alert message and the urgency of the message.

[0046] In one or more embodiments of the method, the decision procedures include a decision theoretic algorithm which selectively generates the alert message based on preferences of the user, uncertainty associated with the information regarding the user, and cost of different actions.

[0047] In one or more embodiments of the method, the decision theoretic algorithm includes one or more Bayesian Decision Networks (BDNs).

[0048] In one or more embodiments of the method, the parameters of the BDNs are elicited from stakeholders or experts or based on the preferences of similar user groups or a combination of these techniques.

[0049] In one or more embodiments of the method, the BDNs are configured to take into account the effect of previous decisions on subsequent decisions.

[0050] Another aspect of the invention provides a user inactivity detector for use in a monitoring system for monitoring activity of a user in a physical environment having one or more regions, the monitoring system including one or more sensors for detecting movement of the user in each region; and a data logger for logging each sensor event, the time of the event and the region in which movement was detected, the method of monitoring activity of a user in a physical environment having one or more regions, the user inactivity detector including a processor and a memory storing a series of instructions to cause the processor to:

determine, for each region, whether time elapsed since a last detected event is an outlier in a distribution of transition times between consecutive detected sensor events for that region, the outlier being indicative of abnormal user inactivity.

[0051] In one or more embodiments of the user inactivity detector, the series of instructions further cause the processor to carry out a method as described hereabove in paragraphs [0029] to [0044].

[0052] Another aspect of the invention provides an alert generator for use in a monitoring system for monitoring activity of a user in a physical environment having one or more regions, the alert generator including a processor and a memory storing a series of instructions to cause the processor to carry out a method as described hereabove in paragraphs [0045] to [0049].

Brief Description of the Drawings

[0053] The invention will now be described in further detail by reference to the accompanying drawings. It is to be understood that the particularity of the drawings does not supersede the generality of the preceding description of the invention.

[0054] Figures 1 and 2 are schematic diagrams depicting a monitoring system for monitoring activity of a user in a physical environment according to an embodiment of the present invention;

[0055] Figure 3 is a flowchart depicting operations performed by the monitoring system shown in Figures 1 and 2;

[0056] Figures 4 to 7 are depictions of different Bayesian Decision Networks implemented in an alert generator forming part of the monitoring system shown in Figures 1 and 2; and

[0057] Figure 8 is a schematic diagram showing various functional elements of part of the monitoring system depicting in Figures 1 and 2.

Detailed Description

[0058] Referring now to Figure 1 , there is shown generally a monitoring system 10 for monitoring activity of a user 12 in a physical environment, such as a home 14, having one or more regions, in this case the rooms or other regions n to r_B. The monitoring system also includes a series of sensors for detecting movement of the user 12 in each of the regions n to r₈.

[0059] Various non-intrusive sensors are depicted in this exemplary embodiment. Passive infrared (PIR) or microwave motion sensors - depicted by a triangle - detect movement in their coverage area. When movement is detected, the sensor sends a violated signal, and when movement is no longer observed, it sends a normal signal. Reed sensors - depicted by a rectangle - are attached to doors (e.g. fridge door, entry door) and detect changes in status. They send a violated signal when a door is opened and a normal signal when it is closed. Unlike motion sensors, the time between violated and normal signals is not necessarily a sign of activity, since doors can be left open. Beam breakers - depicted by a line terminated at both ends by a diamond - send a violated signal when the infrared beam between the emitter and the receiver of the device is interrupted and a normal signal when the beam becomes unbroken again. These sensors do not show the direction of the movement. Finally, pressure mats - depicted by a solid rectangle - consist of a set of pressure switches and send a violated signal when at least one switch makes contact due to pressure on the mat. A normal signal is sent when all the switches disconnect. The sensor is currently position under the mattress.

[0060] It will be appreciated that the above-described non-intrusive sensors depicted in Figure 1 are illustrative only, and that the type, number and location of sensors selected will depend upon the particular monitoring application. In order to make an appropriate initial selection of sensors it can be helpful to ascertain the types of activities performed in different regions within the home 14. For example, if the user 12 often watches television in the lounge room, a TV remote control sensor (not shown) may be used to log actions performed with the TV remote control, whilst if the user normally reads in that room, several motion sensors may be required to detect small movements. Additional sensors that may be used include, but are not limited to, a smoke detector with a wireless transmitter, power usage meters, which log the wattage used by specific appliances, e.g. computer monitor, TV and electric kettle, and microwave motion sensors

[0061] Each of these sensors transmits a signal indicative of a sensor event to a data logger 16. In this case signal transmission occurs wirelessly by means of a radio frequency signal, but in other applications one of more wired connections may be used. The data logger 16 acts to log each sensor event, the time of that event and the region of the home 14 in which the sensor event (e.g. detection of movement) occurred. The data logger 16 transmits this information to an Abnormal Inactivity Detector (AID) 18 which models the user's normal length of inactivity in each region of the home 14. The AID 18 transmits information about a threshold being exceeded to an alert generator 20 which employs decision making procedures to determine whether or not to transmit an alert message to one or more potential message recipients, such as the user 12, carers and/or medical professionals, so that an appropriate response can be taken.

[0062] The home 14 depicted in Figure 1 is a specific example of a more general case in which a set of JV sensors is provided in a physical environment, the sensors being distributed in regions R = {r_t \i = \, ... , L] where L < N. A sensor event is a signal that indicates events such as a door being opened or closed (reed sensor), or movement followed by its termination (PIR motion sensor). A transition time from region r_; is the time elapsed between two consecutive events such that the first event is logged in a region r,.

[0063] As seen in step 40 of Figure 3, a data stream from the sensors is logged by the data logger 16 (Figure 1 ) as a time-stamped sequence of sensor events. For reed sensors, the events of interests are violated and normal signals indicating opening and closing of doors respectively. For beam breakers, the events of interests are pairs of violated-normal signals. In the case of PIR motion sensors, movement is considered to have ended when a normal signal is received or when a violated signal from another region is received (this can happen due to transmission delays from the sensor in the originating region). Finally, pressure mats are treated in the same way as reed sensors, because a person might become incapacitated while in bed. That is, the time between violated and normal signals from pressure mats does not indicate an activity. Power sensors send analog signals (wattages), with sensors for different appliances being treated individually, e.g., high wattage in an electric kettle or the computer monitor indicate activity, but since the kettle and the monitor turn themselves off, low wattage is not indicative of anything. In contrast, for the TV, changes between high and low wattage

may indicate activity. The TV remote control sensor is treated like a reed sensor, as any change in status indicates activity.

[0064] The AID 18 in Figure 1 includes a user inactivity modeller 22 and decision making logic 24 to determine whether or not abnormal user inactivity has occurred. As seen in step 42 of Figure 3, the user inactivity modeller 22 is configured to model at least part of a distribution of transition times between consecutive detected sensor events for normal inactivity occurring in each region n to rs in the home 14, and as seen in step 44 of Figure 3, for each such model, the user inactivity modeller 22 determines an upper threshold for transition times indicative of abnormally long periods of user inactivity. The model and its relevant thresholds need not be updated continuously, but following an initialisation phase can be updated periodically as the system 10 adapts to changing conditions.

[0065] The AID 18 is configured to carry out this modelling and detection for different temporal intervals during a recurring time period, where the user activity typically exhibits consistency during the same temporal interval of different time periods. Typically, a recurring time period may be a day, and temporal intervals may be hours during the day. However in other examples the recurring time period may be a month or year, and temporal intervals may be days or weeks during that extended time period.

[0066] In the particular embodiment depicted in Figure 1 , the AID 18 generates a threshold for the duration of a plausible period of inactivity in each region of the house for each hour of the day. The alert generator 20 selectively issues one or more alerts when the time elapsed since the last event in the most recently visited region exceeds the threshold for this region at the current point in time. To generate these thresholds, the AID 18 first models the distribution of the transition times for each region and temporal interval, and then sets an upper limit for each distribution.

Models for inactivity detection

[0067] The inactivity detection mechanism is implemented by the AID 18 in two main stages: (1) building models of normal inactivity periods from sensor observations; and (2) detecting abnormally long periods of inactivity on the basis of these models. The models of normal inactivity are in turn constructed in two stages: initialization, where the user inactivity modeller 22 learns a start-up model for the detection of normal inactivity periods from sensor observations; and, optionally, adaptation, where the model gradually adapts to changes in behaviour.

Model initialization

[0068] User activities depend on the time of the day and the particular region of the home 14, and the AID 18 provides a trigger signal to the alert generator 20 when an abnormally long inactivity period is detected.

An exponential distribution is widely used to model the time between two events. However, the distribution of transition times has a longer tail than the exponential distribution. Therefore, the user inactivity modeller 22 is configured to use models suitable for long tail or heavy tail distributions. In that regard, the user inactivity modeller 22 can be configured, in alternative embodiments, to use different distributions used in modelling such long-tailed distributions, such as a Pareto distribution and a hyperexponential distribution.

Pareto distribution

[0069] The probability density function of a Pareto distribution is shown in Eqn. 1. The parameter is usually called the shape parameter and x_m the scale parameter or minimum value.

[0070] Since the logged data contains zeroes, and for the Pareto distribution, the minimum value x_m should be greater than 0, the user inactivity modeller 22 truncates the distribution of the data at 1 before estimating a Pareto distribution. It is worth noting that the Pareto distribution is closed under the operation of truncation from below. For data in each region-hour pair, the user inactivity modeller 22 calculates the sample mean of the data and uses the techniques described in paragraphs [0087] to [0088] to calculate an alert threshold for the distribution.

Hyperexponential distribution

[0071] Long-tailed distributions can be approximated arbitrarily closely by a mixture of exponential distributions or hyperexponential distribution. The use of a hyperexponential distribution allows the user inactivity modeller 22 to approximate long- tail probability distributions by convenient exponential probability distributions. The user inactivity modeller 22 models only the exponential distribution in the tail of the hyperexponential distribution to calculate an upper bound for the distribution of inactivity periods. That is, the user inactivity modeller 22 identifies an exponential distribution that characterizes long inactivity periods.

[0072] The user inactivity modeller 22 can be configured, in alternative embodiments, to perform this task in two different ways. A first configuration is based on a Coefficient of Variation (CV), which is defined as σ/μ, the ratio of the standard deviation to the mean, and shows the extent of variability in relation to the mean of the population (CV = 1 for an exponential distribution).

[0073] In a second configuration, the user inactivity modeller 22 carries out direct estimation of the exponential distribution in the tail by using special locations in the histogram of the data samples.

Coefficient of Variation

[0074] The user inactivity modeller 22 uses a heuristic divide-and-conquer method that splits a long-tailed distribution into a sequence of exponentially distributed partitions so that CV≥ 1 for each partition. The user inactivity modeller 22 uses this approach to identify the last exponential distribution of the tail by first building a continuous histogram from the data. The user inactivity modeller 22 then uses h - 2^■ IQR^■ n,^{~1 3} as the bin width in the histogram, where IQR is the interquartile range of the data and n, is the

number of inactivity periods in a particular region and hour. This simple and robust way of selecting the width of a bin often gives quite reasonable results in real applications.

[0075] The user inactivity modeller 22 accumulates bins from the largest to the smallest until the CV estimated from the sample mean fi and standard deviation 8 of the data points in these bins exceeds or equals 1.

[0076] If the selected bins contain enough data points to reliably estimate the parameter of the last exponential distribution, i.e., the sample mean (as described in paragraph [0077]) the user inactivity modeller 22 applies the estimator in Equation 2 to estimate the sample mean: μ =— med I

^ ln2 ' (2)

This is a robust estimator that handles outliers, and uses the median of the selected data points {¾, - , xi_C[ to estimate the sample mean, where / and / are the indices of the first and last selected bins respectively, and c_f is tne number of data points in bin .

[0077] To check whether there are enough data points to estimate the sample mean with a certain degree of confidence, the user inactivity modeller 22 uses ((ζ_γ/₂) σ/Ε^Λ\ , which calculates the minimum number of data points required for an estimation within a margin of error E, where Ζ_γ/2 is the value that yields 1 - y/2 for the cumulative Normal distribution. In one embodiment, setting E to μ/2 and choosing a 95% confidence interval (for which Z_Y/2 « 2) yields roughly fe— 16 data points. It will be appreciated that in other embodiments there may be a different number of data points. Using E - μ/2 yields a confidence interval of size μ, therefore with less data points the estimation will have a confidence interval that is larger than the parameter that the user inactivity modeller 22 is trying to estimate, i.e., μ. Due to this fact, if the number of selected data points is less than k = 16, the user inactivity modeller 22 does not estimate the parameter of the distribution from the current region-hour pair.

Direct estimation

[0078] The user inactivity modeller 22 requires at least two bins of the histogram of the data in the tail of a hyperexponential distribution to estimate the parameter of an exponential distribution in the tail. In direct estimation, unlike CV, the user inactivity modeller 22 identifies the tail part of the histogram which is not influenced by earlier components of the distribution.

[0079] After identifying the tail part of the distribution, the user inactivity modeller 22 uses the following procedure to estimate the parameter of the exponential distribution responsible for the tail part of the data. Since the density function of an exponential distribution s p(x) = ae ^{" x} , log(p(;t)) = log(a) - ax is linear in x. Therefore, the user inactivity modeller 22 needs at least two distinct values for a pair (x, ]og(p(x))) to estimate a using the least squares (LS) method. To obtain probability estimates from our data, the user inactivity modeller 22 divides n, data samples into I bins (Z^,^■·^■ , i>J with i^ci- "· , ¾] samples respectively, where the width of a bin is determined by h =

2^■ IQR^■ Then, it assigns a density to the center of each bin using the number of points in the bin. The density in the center Q_b of bin b_u is approximated by the fraction of points in the data that fail into b_u:

^PW ^ - ⁽³⁾

The user inactivity modeller 22 then identifies the first bin bf of the tail section of the distribution using the procedure described in paragraph [0080]. This yields the pairs

from which the user inactivity modeller 22 estimates the parameter a of the last exponential distribution in the hyperexponential distribution.

[0080] The mean of the component that is responsible for the tail is larger than the mean of the whole distribution. Hence, the user inactivity modeller 22 starts from the first bin in the histogram b_s whose center 9_b is larger than the sample mean of the data, and estimates an exponential distribution starting at each possible bin from this point onwards using the procedure described in paragraph [0079]. Then it chooses a bin b_f that satisfies the following constraint: fibf - ¾,·_,. = max 0 _s+1 - fib_s-^■■■■ fib[ - i) (4) where fi_bu is the estimated mean of the exponential distribution obtained using the bins from b_u onwards. That is, the algorithm chooses the bin that produces the largest change in the estimated mean compared to the previous bin. This idea is based on the fact that the effect of earlier exponential distributions of a hyperexponential distribution decreases exponentially. Hence, when the bins affected by these earlier components are removed, the estimated mean jumps to the mean of the tail.

[0081] The following rules are applied by the user inactivity modeller 22 when searching for the initial bin of the exponential tail:

1. Any bin for which the mixture coefficient for the estimated exponential

(calculated as described below) is less than a threshold t_e is ignored;

suggested values for t_e are in the range [0.05, 0.15].

2. Any bin yielding less than k (=16 data points in the tail (calculated as

described in paragraph [0077]) is ignored.

To estimate the mixture coefficient for the tail component of the hyperexponential distribution, the user inactivity modeller 22 first estimates the data in the segment prior to bin bf that are accounted for by the exponential distribution in the tail. This is done by using the parameter obtained from the bins in the tail by the LS method to calculate log (p((9ftj) for e {1, ··· , / - 1}, the log of the probability at the center of each bin at the head of the distribution. The number of data points in these bins, c_u, that is accounted for by the exponential distribution in the tail, can then be calculated using Equation 3.

The mixture coefficient for the exponential in the tail is estimated by

Model adaptation

[0082] To accommodate changes in normal user activity over time, the user inactivity modeller 22 implements an adaptation regime based on the idea of gracefully forgetting data from the past. This is done by introducing a forgetting factor 0 < λ < 1, where the inactivity periods from A adaptation steps ago are weighted by λ^Α. This type of exponential forgetting is widely used in the estimation literature, and the suggested value for λ is [0.9, 1 ). Accordingly, a previously learned model is adapted by applying a weighting to recorded transition times, wherein a greater weighting is applied to more recent transition times than to older transition times.

[0083] In the case of the above-described CV approach, the last exponential distribution for a particular region in the house and hour of the day is identified as described in paragraph [0074] to [0076]. However, the adapted CV is calculated from the weighted mean ft_w and standard deviation S_w of the data. Given a sequence of inactivity periods {x_fi, ... , x_lCl} for a region-hour pair, weighted by W = {w_{f x}, . .. , w_iCl) (where w_it e {λ^Α, 1} for i— fl, .. . , lc_u several w_ir may have the same value, λ is the same for all regions, and it is possible to use different values for A), fi_w and 8_W are calculated as follows:

where n_fl = ¾₌₁ c_t (the number of data points being examined). The parameter of the resultant exponential distribution is estimated by means of a weighted median of the data (instead of the regular median in Equation 2), which is a value such that the total weight of the samples above and below this value is approximately equal to half the total weight of all samples.

[0084] In the case of the above-described Pareto approach, the weighted sample mean is calculated directly from the samples in the entire distribution as follows:

[0085] For the direct estimation method, the user inactivity modeller 22 uses the sum of weights of the samples in each bin b_u, rather than the number of samples in each bin, to compute (#fc,J in Equation 3.

Detecting abnormally long inactivity periods

[0086] To detect abnormally long periods of inactivity, the user inactivity modeller 22 (1 ) calculates an alert threshold τ for each region-hour pair from the distributions obtained above, (2) adjusts the thresholds, and (3) derives and adjusts an alert line for each region. This is a line that connects the thresholds estimated for a region throughout the day, indicating how long a resident would have to be inactive at each point in time to warrant an alert.

Calculating an alert threshold

[0087] To set an alert threshold for a region-hour pair, the user inactivity modeller 22 determines how much variation can be expected from the mean of the distribution fitted by the above processes. That is, the threshold must be such that it is highly likely that no inactivity period from the distribution exceeds this threshold. The threshold τ in Equation 6 is used to identify outliers. An alert is triggered if an inactivity period exceeds the threshold τ for a particular region in the house and hour of the day. τ = <" 3(η_ΰ, β) , (6) where ¾ is the number of samples used to estimate the sample mean fl; β marks the β - outlier region of the distribution, such that the probability of seeing a value from the distribution is less than β (β is a configurable parameter); and g(n_q^^~) is a multiplier that defines the width of the no-alert region.

[0088] The user inactivity modeller 22 uses Equation 7 or 8 to estimate values of g for the Pareto distribution and the exponential distribution. P(¾ < τ) = 1 - β for k (7)

P (x_fc < log (l - (1 - ^)^{1 n}")) = 1 - /J for k = 1, ... Λ (8)

Adjusting the thresholds

[0089] There may be situations where thresholds cannot be estimated due to insufficient data points or the absence of data points (missing thresholds). In addition, the above processes may produce low thresholds of only a few minutes in high-traffic regions, such as hallways, which are too short for issuing an alert.

[0090] When there are not enough data points to estimate a threshold for a temporal interval in a region, the user inactivity modeller 22 uses the maximum of the data as the threshold.

[0091] When there is no data to estimate a threshold for a temporal interval in a region, the user inactivity modeller 22 is configured to define the threshold by calculating a function of thresholds from temporally proximate periods or similar regions. For example, missing thresholds can be estimated by the user inactivity modeller 22 by computing a weighted average of the nearest valid thresholds, where the weight for a 'neighbouring' threshold depends on the time difference between the hour with the threshold and the hour without a threshold (longer time differences imply a weaker relation). For example, if the user inactivity modeller 22 cannot calculate a threshold for region at 3 am, and the nearest valid thresholds for region are r_;i at 1 am and r_i7 at 7 am, the 3 am threshold for region r_; will be (1 - ~~) · τ_;ι + (1 - ~~ ) · τ_ί7.

[0092] Low thresholds are replaced by a minimum threshold MT below which an alert cannot be raised. This is a configurable system parameter.

Deriving and adjusting alert lines

[0093] An alert line for a region is derived by the user inactivity modeller 22 by connecting the thresholds for that region throughout a 24 hour period. An alert line may have inconsistent thresholds, which depict sharp changes in behaviour that are unrealistic. It may also have unreachable thresholds, which cannot be physically reached. For example, if the threshold for the bedroom at 3 pm is 4 hours, and at 2 pm it is 1 hour, it is impossible to reach 4 hours of inactivity at 3 pm without first triggering an alert at 2 pm. To address these problems, the user inactivity modeller 22 makes the following adjustments.

[0094] Inconsistent thresholds are corrected by means of a simple moving average low pass filter or other smoothing function to smooth the alert line. This is done continuously, so that the end of a day coincides with beginning of the next day.

[0095] Unreachable thresholds are decreased to their maximum useful value, which in the above example is 2 hours at 3 pm.

Generating alerts

[0096] Once the AID 18 has detected the occurrence of unusual or abnormal user inactivity, the alert generator 20 applies a decision procedure to determine whether a message should be sent, and if so, to whom, e.g. the user or another person, and with what urgency. The alert generator 20 can be configured to transmit an alert signal in a variety of ways. As an example, the alert generator 20 may be configured to transmit an SMS or like message to a smartphone 40 in Figure 2 or another portable communications device associated with a healthcare professional 42 via a mobile communications network 44. In another example, an alert message or signal can be transmitted to a computer terminal 46 associated with the healthcare professional 42 and viewed on a smartphone 40 a graphic user interface 48 of a display 50 connected to the computer terminal 46. Alert messages or signals transmitted via the data network 56 - to which the computer terminal 46 is connected - can be logged at a database server 52 and associated database 54.

[0097] The alert generator 20 may use a single decision procedure or a combination of different decision procedures, or may transmit directly the alert generated by user inactivity modeller 22. Examples of such decision procedures are rule-based procedures, probabilistic procedures, such as decision trees, and decision-theoretic procedures, such as Bayesian decision networks (BDNs), also known as influence diagrams.

[0098] The BDNs implemented in the alert generator 20 provide a user-oriented decision process in response to an alert from the user inactivity modeller 22. The embodiments of the BDNs shown in Figures 4 to 7 take into account stakeholder related parameters such as preferences of the stakeholders, prior knowledge about the monitoring environment and the monitored people, as well as the costs and uncertainties associated with an action.

[0099] The alert generator 20 can use the BDNs to decide on one or more actions. The current decision is related to the decisions made previously by the BDN. The BDNs in the alert generator 20 have a static design where time dependencies are achieved through special nodes that carry the effects of the previous decisions to subsequent ones. For example, if a user alert was previously issued, and there was no response, this affects the next decision.

[0100] The parameters of the BDNs are elicited from stakeholders or experts or are based on the preferences of similar user groups or a combination of these techniques.

[0101] Figure 4 is a depiction of a BDN for deciding whether to interact with an elderly person (user) in his/her home. In this BDN, table 60 is used to record the probable initial level of safety of a user. Table 62 records the time since the last alert from user inactivity modeller 20 regarding the current incident (if any), table 64 records the user's response to the last alert pertaining to the current incident (if any), and table 66 records the probabilities of different levels of risk at the time of the last activation of alert generator 20 for the current incident (if any). Table 68 infers the probability of different levels of risk at present from the information in tables 60-66. Tables 70 and 72 respectively record whether the user is in a location where he or she does not wish to be disturbed, and whether the time of the day is one where the user normally does not wish to be disturbed, and table 74 records the disposition of the user (i.e. to what extent he or

she minds being disturbed in general). Table 76 represents the utilities of different decisions in light of the events in tables 70-74.

[0102] Figure 5 is a depiction of a BDN for deciding on the intensity level of the call to the user. A similar BDN may be designed for carers. Table 78 records the probability of different levels of current risk carried over from table 68 in Figure 4. Table 80 records the probabilities of the different intensity levels at the last activation of the alert generator 20 for the current incident (if any). Table 82 records the time since the last attempt to call the user by the alert generator 20 regarding the current incident (if any). Table 84 records the disposition of the user (i.e. to what extent he or she minds being disturbed in general). Tables 86 and 88 respectively record whether the user is in a location where he or she does not wish to be disturbed, and whether the time of the day is one where the user normally does not wish to be disturbed. Table 90 represents the utilities of different decisions in light of the events in tables 78-88.

[0103] Figure 6 is a depiction of a BDN for deciding whether to interact with caregivers of the monitored user. Table 92 records the probability of different levels of current risk carried over from table 68 in Figure 4. Table 94 records the response of the caregiver to the last activation of the alert generator 20 (if any). Table 96 records the time elapsed since the last attempt to call the caregiver by the alert generator 20 regarding the current incident (if any). Table 98 records the disposition of the care-giver (i.e. to what extent he or she minds being disturbed in general). Table 100 records whether the time of the day is one where the user normally does not wish to be disturbed. Table 102 represents the utilities of different decisions in light of the events in tables 92-100.

[0104] Figure 7 is a depiction of a BDN for deciding whether to interact with emergency response services, e.g., ambulance. Table 104 records the probability of different levels of current risk carried over from table 68 in Figure 4. Table 106 records the response of the medical personnel to the last activation of the alert generator 20 (if any). Table 108 records the time since the last attempt to call the medical personnel by the alert generator 20 regarding the current incident (if any). Table 110 represents the utilities of different decisions in light of the events in tables 104-108 and the direct cost of calling emergency response services.

[0105] Any one of more of the data logger 16, AID 18 and alert generator 20 may be implemented using hardware, software or a combination thereof and may be implemented in one or more computer systems or processing systems. An exemplary computer system 130 is shown in Figure 8. The computer system 130 includes one or more processors, such as processor 132. The processor 132 is connected to a communication infrastructure 134. The computer system 130 may include a display interface 136 that forwards graphics, texts and other data from the communication infrastructure 134 for supply to the display unit 138. The computer system 130 may also include a main memory 140, preferably random access memory, and may also include a secondary memory 142.

[0106] The secondary memory 142 may include, for example, a hard disk drive, magnetic tape drive, optical disk drive, etc. The removable storage drive 144 reads from and/or writes to a removable storage unit 146 in a well-known manner. The removable storage unit 146 represents a magnetic storage unit, such as magnetic tape, optical disk, flash drive, hard drive, etc.

[0107] As will be appreciated, the removable storage unit 146 includes a non- transitory computer usable storage medium having stored therein computer software in a form of a series of instructions to cause the processor 132 to carry out the desired functionality. In alternative embodiments, the secondary memory 142 may include other similar means for allowing computer programs or instructions to be loaded into the computer system 130. Such means may include, for example, a removable storage unit 148 and interface 150.

[0108] The computer system 130 may also include a communications interface 152. Communications interface 152 allows software and data to be transferred between the computer system 130 and external devices. Examples of communication interface 152 may include a modem, a network interface, a communications port, a PC IA slot and

card etc. Software and data transferred via a communications interface 152 are in the form of signals which may be electromagnetic, electronic, optical or other signals capable of being received by the communications interface 152. The signals are provided to the communications interface 152 via a communications path 54 such as a wire or cable, fibre optics, phone line, cellular phone link, radio frequency or other communications channels.

[0109] Although in the above described embodiments the invention is implemented primarily using computer software, in other embodiments the invention may be implemented primarily in hardware using, for example, hardware components such as an application specific integrated circuit (ASICs). Implementation of a hardware state machine so as to perform the functions described herein will be apparent to persons skilled in the relevant art. In other embodiments, the invention may be implemented using a combination of both hardware and software.

[0110] While the invention has been described in conjunction with a limited number of embodiments, it will be appreciated by those skilled in the art that many alternatives, modifications and variations in light of the foregoing description are possible. Accordingly, the present invention is intended to embrace all such alternatives, modifications and variations as may fall within the spirit and scope of the invention as disclosed.

[0111] For example, the various embodiments of the monitoring system 10 described hereabove include an alert generator 20 for receiving an indication of unusual user inactivity from the AID 18 and then deciding whether an alert message should be sent, and if so, to whom and with what urgency. However, in other more simplified embodiments, the monitoring system 10 may omit such an alert generator and may instead simply notify one or more person, such a carer or medical professional, each time abnormal user inactivity is detected.

[0112] Furthermore, in the above-described embodiments, the user inactivity detector 18 determines outliers in a distribution of transition times after a sensor event in a region by determining whether time elapsed since the last event exceeds the upper threshold of its statistically defined normal range for each temporal interval in the region. It should be understood however that the detection of an outlier in a set of transition times need not use an approach based on a statistically obtained threshold, and that other outlier detection techniques may be used. Such techniques may include fixed thresholds, statistical hypothesis testing, and non-parametric techniques such as clustering, Support Vector Machines (SVM), and rule-based, Cumulative Summation (CUSUM)-based, and density-based methods.

Claims

The claims defining the invention are as follows

1. A monitoring system for monitoring activity of a user in a physical environment having one or more regions, the monitoring system including

one or more sensors for detecting movement of the user in each region;

a data logger for logging each sensor event, the time of the event and the region in which movement was detected; and

a user inactivity detector configured to determine, for each region, whether the time elapsed since a last detected event is an outlier in a distribution of transition times between consecutive detected sensor events for that region, the outlier being indicative of abnormal user inactivity.

2. A monitoring system according to claim 1 , wherein the user inactivity detector is configured to

3. A monitoring system according to claim 2, wherein the user inactivity detector is further configured to

separately model different temporal intervals during a recurring time period, wherein user activity typically exhibits consistency during the same temporal interval of different time periods.

4. A monitoring system according to claim 3, wherein the user inactivity detector is further configured to

5. A monitoring system according to claim 3, wherein the user inactivity detector is further configured to

6. A monitoring system according to any one of claims 2 to 5, wherein the user engages in specific behaviours in each region.

7. A monitoring system according to any one of claims 2 to 6, wherein the distribution of transition times between consecutive detected sensor events is modelled as a iong-tailed distribution.

8. A monitoring system according to claim 7, wherein the user inactivity detector is configured to model the distribution as a Pareto distribution.

9. A monitoring system according to claim 7, wherein the user inactivity detector is configured to model the distribution as a hyperexponential distribution.

10. A monitoring system according to claim 9, wherein the user inactivity detector is configured to model the tail of the hyperexponential distribution by

selecting a partition of data by accumulating transition times from the largest sample to the smallest sample until an estimated Coefficient of Variation is greater than or equal to 1 ; and

fitting an exponential distribution to that partition of data.

11. A monitoring system according to claim 9, wherein the user inactivity detector is configured to model the tail of the hyperexponential distribution by

selecting a partition of data by accumulating bins from a histogram of the transition times starting from the bin with largest value to that with the smallest value until the estimated Coefficient of Variation from these bins becomes greater than or equal to 1 ; and

fitting an exponential distribution to that partition of the data.

12. A monitoring system according to claim 9, wherein the user inactivity detector is configured to model the tail of the hyperexponential distribution by

direct estimation of the exponential distribution in the tail from the histogram of the data by identifying the bins that are not affected by the exponential distribution(s) at the head of the hyperexponential distribution, and directly estimating an exponential distribution using the density values estimated from these unaffected bins.

13. A monitoring system according to claim 12, wherein the bins are identified by pinpointing the bin in the histogram that produces the largest change in the mean of the exponential distribution estimated from that bin onwards compared to the mean of the exponential distribution estimated from the previous bin.

14. A monitoring system according to any one of claims 2 to 13, wherein the user inactivity detector is further configured to

set the upper threshold for each model to no lower than a minimum preset threshold value.

15. A monitoring system according to any one of claims 2 to 14, wherein the user inactivity detector is further configured to

derive an alert line by connecting thresholds for a particular region throughout the recurring time period; and

adjust inconsistent thresholds by applying a smoothing function to the alert line.

16. A monitoring system according to any one of claims 2 to 15, wherein the user inactivity detector is further configured to

adjust the upper threshold for each model by decreasing unreachable thresholds to their maximum useful value.

17. A monitoring system according to any one of claims 2 to 16, wherein the user inactivity detector is further configured to

adapt each model by applying a weighting to recorded transition times, wherein a greater weighting is applied to more recent transition times than to older transition times.

18. A monitoring system according to any one of the preceding claims, and further including

an alert generator configured to employ one or more decision procedures to decide whether an alert message should be sent in response to detection of an outlier in the logged sensor events, and, in the event that an alert message is to be sent, to determine an intended recipient of the alert message and the urgency of the message.

19. A monitoring system according to claim 18, wherein the decision procedures include a decision theoretic algorithm which selectively generates the alert message based on preferences of the user, uncertainty associated with the information regarding the user, and cost of different actions.

20. A monitoring system according to claim 19, wherein the decision theoretic algorithm includes one or more Bayesian Decision Networks (BDNs).

21. A monitoring system according to claim 20, wherein the parameters of the BDNs are elicited from stakeholders or experts or based on the preferences of similar user groups or a combination of these techniques.

22. A monitoring system according to either one of claims 19 or 20, wherein the BDNs are configured to take into account the effect of previous decisions on subsequent decisions.

23. A method of monitoring activity of a user in a physical environment having one or more regions, the method including the steps of

detecting movement of the user in each region by one or more sensors;

logging each sensor event, the time of the event and the region in which movement was detected in a data logger; and

24. A method according to claim 23, wherein outlier detection includes the steps of: modelling at least part of the distribution of transition times between consecutive detected sensor events for each region,

for each model, determining an upper threshold for transition times indicative of unusual amounts of user inactivity, and

determining when time elapsed since a last detected event exceeds the upper threshold for each model.

25. A method according to claim 24, and further including the step of:

separately modelling different temporal intervals during a recurring time period, wherein user activity typically exhibits consistency during the same temporal interval of different time periods.

26. A method according to claim 25, and further including the steps of:

27. A monitoring system according to claim 25, and further including the steps of: determining whether parameters of the model can be estimated from the detected events, and if there is not enough data points to estimate the upper threshold for a particular temporal interval in a region,

28. A method according to any one of claims 24 to 27, wherein the user engages in specific behaviours in each region.

29. A method according to any one of claims 24 to 28, wherein the distribution of transition times between consecutive detected sensor events is modelled as a long-tailed distribution.

30. A method according to claim 29, wherein the distribution is modelled as a Pareto distribution.

31. A method according to claim 30, wherein the distribution is modelled as a hyperexponential distribution.

32. A method according to claim 31 , wherein the tail of the hyperexponential distribution is modelled by:

fitting an exponential distribution to that partition of data.

33. A method according to claim 31 , wherein the tail of the hyperexponential distribution is modelled by

fitting an exponential distribution to that partition of the data.

34. A method according to claim 31 , wherein the tail of the hyperexponential distribution is modelled by

35. A method according to claim 34, wherein the bins are identified by pinpointing the bin in the histogram that produces the largest change in the mean of the exponential distribution estimated from that bin onwards compared to the mean of the exponential distribution estimated from the previous bin.

36. A method according to any one of claims 24 to 35, and further including a step of setting the upper threshold for each model to no lower than a minimum preset threshold value.

37. A method according to any one of claims 24 to 36, and further including the step of

deriving an alert line by connecting the thresholds for a particular region throughout the recurring time period; and

adjusting inconsistent thresholds by applying a smoothing function to the alert line.

38. A method according to any one of claims 24 to 37, and further including the step of

adjusting the upper threshold for each model by decreasing unreachable thresholds to their maximum useful value.

39. A method according to any one of claims 24 to 38, and further including the step of

adapting each model by applying a weighting to recorded transition times, wherein a greater weighting is applied to more recent transition times than to older transition times.

40. A method according to any one of the claims 24 to 39, and further including the step of

employing one or more decision procedures to decide whether an alert message should be sent in response to detection of an outlier in the logged sensor events, and, in the event that an alert message is to be sent, to determine an intended recipient of the alert message and the urgency of the message.

41. A method according to claim 40, wherein the decision procedures include a decision theoretic algorithm which selectively generates the alert message based on preferences of the user, uncertainty associated with the information regarding the user, and cost of different actions.

42. A method according to claim 41 , wherein the decision theoretic algorithm includes one or more Bayesian Decision Networks (BDNs).

43. A method according to claim 42, wherein the parameters of the BDNs are elicited from stakeholders or experts or based on the preferences of similar user groups or a combination of these techniques.

44. A method according to either one of claims 42 or 43, wherein the BDNs are configured to take into account the effect of previous decisions on subsequent decisions.

45. A user inactivity detector for use in a monitoring system for monitoring activity of a user in a physical environment having one or more regions, the monitoring system including one or more sensors for detecting movement of the user in each region; and a data logger for logging each sensor event, the time of the event and the region in which movement was detected, the method of monitoring activity of a user in a physical environment having one or more regions, the user inactivity modeller including a processor and a memory storing a series of instructions to cause the processor to: determine, for each region, whether time elapsed since a last detected event is an outlier in a distribution of transition times between consecutive detected sensor events for that region, the outlier being indicative of abnormal user inactivity.

46. A user inactivity detector according claim 45, wherein the series of instructions further cause the processor to carry out a method according to any one of claims 24 to 39.

47. An alert generator for use in a monitoring system for monitoring activity of a user in a physical environment having one or more regions, the alert generator including a processor and a memory storing a series of instructions to cause the processor to carry out a method according to any one of claims 40 to 44.