(19)
(11)EP 3 405 105 B1

(12)EUROPEAN PATENT SPECIFICATION

(45)Mention of the grant of the patent:
06.05.2020 Bulletin 2020/19

(21)Application number: 17701586.4

(22)Date of filing:  23.01.2017
(51)Int. Cl.: 
A61B 5/024  (2006.01)
G06K 9/00  (2006.01)
(86)International application number:
PCT/GB2017/050162
(87)International publication number:
WO 2017/125763 (27.07.2017 Gazette  2017/30)

(54)

METHOD AND APPARATUS FOR ESTIMATING HEART RATE

VERFAHREN UND VORRICHTUNG ZUR BERECHNUNG DER HERZFREQUENZ

PROCÉDÉ ET APPAREIL D'ESTIMATION DE LA FRÉQUENCE CARDIAQUE


(84)Designated Contracting States:
AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR

(30)Priority: 21.01.2016 GB 201601140

(43)Date of publication of application:
28.11.2018 Bulletin 2018/48

(73)Proprietor: Oxehealth Limited
The Oxford Science Park Oxford OX4 4GA (GB)

(72)Inventors:
  • HUTCHINSON, Nicholas Dunkley
    Oxford Oxfordshire OX4 4GE (GB)
  • JONES, Simon Mark Chave
    Oxford Oxfordshire OX4 4GE (GB)
  • FRAZ, Muhammad
    Oxford Oxfordshire OX4 4GE (GB)

(74)Representative: J A Kemp LLP 
14 South Square Gray's Inn
London WC1R 5JJ
London WC1R 5JJ (GB)


(56)References cited: : 
WO-A1-2010/100593
US-A1- 2014 276 099
US-A1- 2015 221 069
WO-A1-2015/049150
US-A1- 2015 063 708
  
      
    Note: Within nine months from the publication of the mention of the grant of the European patent, any person may give notice to the European Patent Office of opposition to the European patent granted. Notice of opposition shall be filed in a written reasoned statement. It shall not be deemed to have been filed until the opposition fee has been paid. (Art. 99(1) European Patent Convention).


    Description


    [0001] The present invention relates to a method and apparatus for estimating the heart rate of a subject, in particular for analysing video signals from a video camera taking a video image which includes the subject and detecting a photoplethysmographic signal therein.

    [0002] Over recent years techniques have been developed for obtaining an estimate of the heart rate of a human or animal subject which are less intrusive than those using traditional contact sensors, such as electrocardiogram (ECG) sensors or photoplethysmogram (PPG) finger or ear probes. For example Verkruysse et al., "Remote plethysmographic imaging using ambient light", Optics Express, 16 (26), 22 December 2008, PP. 21434-21445 demonstrated that photoplethysmographic signals could be detected in the video signal from a conventional consumer standard video camera where a human subject was illuminated under ambient light.

    [0003] This idea has been developed further in, for example, WO-A2-2013/027027 and WO-A1-2015/049150 which aim to increase the reliability of the detection of the remote PPG signal. The paper "Distance PPG: robust non-contact vital signs monitoring using a camera" by Mayank Kumar et al.; 6 April 2015; Biomedical Optics Express 1565, 1 May 2015, Vol. 6 No. 5, discusses a method of combining skin-colour change signals from different tracked regions of a subject's face using a weighted average, where the weights depend on the blood perfusion and incident light density in the region to improve the signal-to-noise ratio of the camera-based estimate. It discusses the various challenges for camera-based non-contact vital sign monitoring and proposes that improvements in the signal-to-noise ratio of the camera-based estimates reduces the errors in vital sign estimation.

    [0004] Many of the prior art techniques have been based on careful control of the subject being monitored and the lighting conditions in the environment. Thus, although they claim success in detecting the heart rate or vital signs of the subject, in general the subjects were required to remain relatively still, the subjects were not obscured and the lighting conditions were kept relatively constant. It would be useful to be able to detect vital signs, in particular heart rate, of subjects in a wide variety of settings beyond the clinical environment where the subject and environment may be less controlled. For example, being able to monitor a subject in a room, such as a secure room in a detention facility e.g. a prison or police cell, a room in a hospital or care home, or even room in the home, workplace or leisure facility such as a gym, but able to freely move within the room would be useful, but is much more difficult. In real life subjects mix periods of high activity and large movement with periods of relative immobility (seated or lying), will in general be clothed and have bedding to cover themselves. Thus, periods of inactivity while lying down, may coincide with the subject covering themselves partly or completely (known as "tenting") with bedding. Further, within rooms lighting conditions can vary with time, sometimes rapidly. Secure rooms are sometimes lit with visible artificial light and are sometimes completely dark with infrared being the only illumination available. Similar problems of movement and variable illumination occur also in fields such as fitness and health and well-being in the home or elsewhere. Existing systems do not provide vital signs monitoring or heart rate detection which can operate reliably in the face of these difficulties. Being able to detect the heart rate of a subject in these less controlled conditions would significantly improve the ability to monitor the well-being of such a subject and to comply with a duty of care requirement.

    [0005] WO-A1-2010/100593 discloses a method of detecting the presence of a living being by analysing a video image for photoplethysmographic signals. It finds regions of interest in the image by first using a known image segmentation technique such as Viola Jones to detect faces in the image, then selecting measurement zones within the face as regions with minimal intensity variations, and tracking these through the video using a known motion tracking technique. The time-average brightness in the measurement zones can be analysed for a PPG signal.

    [0006] US-A1-2015/0221069 discloses a method of recolouring an image of a face by detecting facial features using the Viola-Jones algorithm and creating polygons from those features. US-A1-2015/063708 discloses analysing an image of people to try to speed up facial detection by having a first step which looks for areas with lots of edges, and then these areas are used as candidates for more detailed facial detection using the Viola-Jones method. US-A1-2014/0276099 relates to detecting PPG signals to derive vital signs, but it aims to assess the reliability of these signals by looking at changes in the spatial and/or temporal properties of the selected regions of interest through the sequence. The aim is to stop trying to derive vital signs when the subject is moving. WO-A1-2015/049150 relates to improving the selection of a PPG signal by looking at signal quality indicators.

    [0007] The present invention provides a method of obtaining an estimate of a periodic vital sign of a subject from a video image sequence of the subject as defined by claim 1.

    [0008] The regions of interest defined to not extend over or include the areas of strong intensity gradients will correspond to visually flatter (i.e. more uniform intensity) regions of the image. This avoids the problem that movement of areas of higher image intensity variation can negatively affect the signal processing and create noise in the intensity variation signals which include the periodic signal of interest.

    [0009] Preferably, the region of interest is defined as a square aligned with orthogonal axes of the frames of the video image sequence. Where the regions of interest are squares aligned with orthogonal axes of the frames of the image, it simplifies and speeds-up image processing, especially if image processing is performed on the integral image.

    [0010] Preferably plural regions of interest are defined in each frame of the video image sequence.

    [0011] The step of detecting an image area with a strong intensity gradient may comprise detecting an image area with an intensity gradient stronger than a predetermined threshold. The threshold may be based on the distribution of magnitudes of the intensity gradients in the image.

    [0012] The step of tracking the region of interest through other frames of the video image sequence may comprise defining the position of the region of interest in other frames of the video image sequence by reference to detected image movement in the video image sequence. Such image movement in the video image sequence may be detected by measuring the optical flow in the video image sequence, e.g. by a dense optical flow algorithm, or by sparse optical flow, e.g. image feature tracking in the video image sequence, e.g. using a standard image feature tracking algorithm such as Lucas-Kanade.

    [0013] By defining regions of interest which do not extend over or include the feature points, the regions of interest correspond to visually flatter (i.e. more uniform intensity) regions of the image. This avoids the problem that feature points - typically areas of higher image intensity variation - can negatively affect the signal processing and create noise in the intensity variation signals which include the periodic signal of interest. The feature points, on the other hand are useful for tracking through the image sequence and allowing the regions of interest that are defined with respect to them to be tracked through the sequence.

    [0014] The method may further comprise the step of defining a grid of image areas whose vertices join the feature points and wherein each region of interest is defined to be entirely within a respective one of said image areas. The image areas may be polygons whose vertices are at the feature points.

    [0015] The step of defining the grid of image areas may comprise defining the grid of image areas on one frame of the sequence and forming grids on the other frames of the sequence by joining the same feature points together. This produces consistent trackable regions of interest.

    [0016] The grid may be triangular, each image area being a triangle, and this can be produced by a triangulation method such as Delaunay triangulation.

    [0017] One way of forming the regions of interest is to define in-circles of said polygonal areas, these naturally avoiding the vertices themselves and thus avoiding the feature points. The regions of interest may be the in-circles or be based on them, e.g. be defined as squares co-centered on the in-circles.

    [0018] The method may further comprise the step of clustering said time series signals to form clusters of time series signals which have greater than a predetermined correlation and are obtained from regions of interest spaced by no more than a predetermined distance in the image, averaging the signals in each cluster, and obtaining the estimate of the periodic vital sign from the averaged signals.

    [0019] The estimate of the periodic vital sign may be obtained by measuring the frequency, or the frequency of the strongest periodic component, of said time series signals or averaged signals.

    [0020] The method may further comprise the step of applying principal component analysis to the time series signals or averaged time series signals, calculating a signal quality index of the principal components and obtaining the estimate by measuring the frequency, or the frequency of the strongest periodic component, of the principal component with the best signal quality index.

    [0021] The intensity variations may include a periodic component corresponding to a photoplethysmogram signal, or may include a periodic movement signal.

    [0022] The periodic vital sign may be the heart rate or breathing rate.

    [0023] Another aspect of the invention provides apparatus for estimating a periodic vital sign of a subject comprising: a video camera for capturing a video image sequence of the subject; an image data processor for executing the method above; and a display for displaying the estimate of the periodic vital sign.

    [0024] The invention also extends to a computer program comprising program code means for executing on a computer system the method above.

    [0025] One embodiment of the invention provides a method and apparatus for estimating heart rate of a subject from a video image sequence of the subject. Regions of interest are generated by detecting and tracking feature points through the video image sequence, triangulating the feature points and generating square regions of interest corresponding to the in-circles of the triangles. The variation in intensity in each region of interest through the video sequence is taken as a time series signal, bandpass filtered in the physiologically expected range and the signals are clustered according to their similarity and spacing in the image. Signals in the same cluster are averaged together and for clusters with at least a preset number of signals, the frequency of the average signal is taken as indicating a heart rate. In an alternative embodiment, the intensity variations from the square regions of interest through the frame sequence are taken as time series signals and those signals which have a strong peak in the power spectrum are selected and subject to principal component analysis. The principal component with a highest signal quality is selected and its frequency is found and used to estimate the heart rate.

    [0026] The invention will be further described by way of non-limitative example with reference to the accompanying drawings, in which:-

    Figure 1 schematically illustrates a subject in a secure room and an apparatus in accordance with one embodiment of the invention;

    Figure 2 schematically illustrates example frames from a video image sequence and corresponding signals in Figure 2B;

    Figure 3 is a flow diagram of signal processing in accordance with an embodiment of the invention;

    Figure 4 is a flow diagram of signal processing in accordance with an embodiment of the invention;

    Figure 5 is a flow diagram of signal processing in accordance with an embodiment of the invention;

    Figure 6 is a flow diagram of signal processing in accordance with an embodiment of the invention;

    Figure 7 is a frame of video data showing the results of signal processing in accordance with the method of Figure 3;

    Figure 8 schematically illustrates a signal and quantities used in calculating a signal quality index;

    Figure 9 schematically illustrates a power spectrum of a signal and quantities used in calculating a second signal quality index;

    Figure 10 is a flow diagram of signal processing in accordance with another embodiment of the invention;

    Figure 11 schematically shows a distribution of square regions of interest over an image frame;

    Figure 12 illustrates an example log normal distribution over width of region of interest.



    [0027] Figure 1 schematically illustrates a secure room 1 occupied by a subject 3 who is monitored by a video camera 5 illuminated selectively by visible spectrum artificial lighting 11 or infrared lighting 13. The output from the video camera 5 is processed by a signal processor 7 and the results are displayed on a display 9. The results may be an estimate of the heart rate, or an indication that no heart rate can be detected, together with an indication of the length of time for which no heart rate has been detected.

    [0028] The output from the video camera is a conventional digital video output consisting of a series of image frames, typically at twenty frames per second, with red, green and blue intensities across the image as an array of pixel values forming each frame. The red, green and blue sensors typically also provide a response in the infra-red (IR), allowing an IR signal to be obtained. Alternatively a monochrome digital video camera providing only one channel can be used - but such cameras also provide an IR signal. The video signal is analysed by a signal processor 7 which may be a programmed general purpose computer or a dedicated signal processing device and the display 9 can display the video image as well as other information, such as the estimated heart rate, or other vital signs obtained by analysis of the video image.

    [0029] The processing of the video signals to obtain an estimate of the heart rate in accordance with one embodiment of the invention will now be described. This embodiment is based on detecting PPG signals in various regions of interest defined in the image frames. Thus the first aspect of this embodiment of the invention is the way in which the regions of interest in the video image are defined. Having defined regions of interest, the image intensities (e.g. average or sum) in the regions of interest through the frame sequence forming the video image then form time series signals which are analysed to detect a PPG signal.

    Defining regions of interest (ROIs)



    [0030] Figure 3 illustrates the signal processing for defining regions of interest in accordance with a first embodiment of the invention. Firstly, in step 100, the average frame intensity of each frame is set to a constant value to reduce image flicker, e.g. by multiplying each pixel value by the mean pixel value over the whole sequence and dividing by a constant to scale the values as desired (e.g. 0-255 for 8 bit values).

    [0031] In step 101, feature points in the video sequence are detected. There are many ways of detecting feature points in a video sequence using off-the-shelf video processing algorithms based on sparse optical flow. For example, feature points consisting of recognisable geometrical shapes such as corner or edges can be detected based, for example, on the gradient of intensity variation in one or two dimensions, and any such conventional algorithm which identifies image feature points can be used in this invention. The feature points are tracked through the whole batch of video frames under consideration, e.g. by using a conventional tracking algorithm such as KLT tracking, to form "tracks" consisting of the x and y coordinates of each feature point in each image frame through the sequence. A measure of the strength of each feature point may also be calculated and stored associated with the feature points, for example corresponding to the strength of the image intensity gradient forming the feature.

    [0032] In general feature detecting and tracking algorithms will generate many more candidate feature points than are required. Preferably in this embodiment the strongest feature point (as measured by gradient intensity) is used and then other feature points are taken in turn and either included or ignored based on their feature strength and spacing from already selected feature points, weighting them, e.g. proportionally, to their feature strength and their minimum distance in the sequence from already-selected feature points. This achieves a reasonably even distribution of feature points across the image. It is also preferable that the extent of movement of each feature through the image sequence under consideration is calculated (i.e. the variation in its x and y coordinates through the time window) and features for which the movement satisfies a predetermined definition of moderate, e.g. movement which is of the order of or greater than the typical movement found in a ballistocardiogram (BCG) and less than the gross level of movement which would preclude detection of a PPG signal, are preferred. This avoids selecting features which either do not move or which correspond to gross movement.

    [0033] This process of detecting features and tracks through the time window and selecting them based on strength, movement and spacing continues until a desired number, for example several hundred, tracks have been selected.

    [0034] In step 103 a time window (e.g. 6 seconds = 60 frames at ten frames per second) is taken. Thus the next steps of the process are conducted on a time window of the video sequence, and then the process will be repeated for another time window shifted along by some time increment. The successive windows may overlap, for example if a six second window is stepped forwards by one second each time the overlap will be five seconds. An estimated heart rate is output (if detected) for each time window. Thus if the window is moved along by one second each time, a new heart rate estimate is, potentially, output every second.

    [0035] A set of "persistent tracks" is defined as the set of all tracks that span all frames in the current window. In step 106, the central frame of the time window is taken and Delaunay triangulation is performed on the persistent tracks. Delaunay triangulation is a process which creates triangles favouring large internal angles. Figure 2 illustrates schematically three successive frames at times n - 1, n and n + 1 with the central frame n having five feature points 20 connected to form triangles. As can be seen in Figure 2, the position of the feature points varies from frame-to-frame. Having formed the triangles in the central frame of the sequence, the same triangles are formed in each other frame of the sequence (i.e. the same feature points are connected together) so that each triangle is defined throughout the whole six second time window by three KLT tracks specifying the positions of its vertices. In step 108, the in-circle 22 of each triangle is formed and then a square 24 concentric with the in-circle 22 is formed, aligned with the x and y axes of the image frame and with a side length equal to the diameter of the in-circle. Each of these squares 24 then constitutes a region of interest from which a signal will be obtained for further processing.

    [0036] In a separate step 104, the integral image is calculated for each frame. As is well known in the art of image processing, in the integral image the value at any point (x, y) is the sum of all of the pixels above and to the left of (x, y), inclusive. The reason for using the integral images is that it simplifies and speeds up the image processing steps involving summing intensities and the steps in the method which involve such sums - e.g. step 110 are preferably conducted on the integral image, though they can be conducted on the original image frames with some loss of speed.

    [0037] As illustrated in step 110 the intensity in each region of interest in each frame is calculated (the sum of all the pixel intensity values) and the intensity for each square region of interest (ROI) through the time window corresponds to a signal (il to im) to be processed. In visible light, for a camera outputting three R, G, B colour channels, only the green channel is used. However if the room is illuminated by infra-red light, the mean of the three colour channels is used. The image intensity of each ROI through the frame sequence will typically vary as schematically illustrated in Figure 2B. The mean location (xm, ym) of each square (for example the centre) is also calculated and associated with the corresponding signal im over the time window. The intensity signals and associated locations are then output as time signals as illustrated in step 112.

    [0038] Figure 7 illustrates triangulation, and the corresponding squares on an actual video image frame.

    [0039] The embodiment above is based on sparse optical flow. A different, unclaimed, method of defining ROIs will now be described based on using dense optical flow, this processing being illustrated in the flowchart of Figure 10.

    [0040] After the same initial steps 90 and 100 of acquiring a video image sequence and reducing flicker, in step 700 a density matrix is initialised, as a matrix of zeros of dimension equal to the video resolution. The density matrix will (after updating) quantify the amount of signal believed to have recently come from each image region (being a pixel and a small area around it) and it is used to influence the distribution of regions of interest used for each time window. It is only initialized at the start of the video sequence and is then used for all frames and time windows in that video sequence, being updated for each time window.

    [0041] In this method, for each time window (e.g. each set of 60 frames), the regions of interest are defined as image-axis-aligned squares of side length w and centred at position (x, y) in the image frame. A set number of regions of interest will be defined, typically from 100 to 500, e.g. 200. The regions of interest are defined by drawing their location and size randomly from probability distributions over (x, y) and w. Figure 11 schematically shows a distribution of square regions of interest (seven for the sake of illustration) over an image frame centred at different positions (x1, y1) to (x7, y7) and with different sizes w1 to w7. The distribution of locations is a function of both the density matrix and image intensity gradients as explained below. The distribution of sizes (over w) is chosen to give a variety of different size squares appropriate to the video sequence to be processed (i.e. based on the scene and the set-up of the camera). For example, a log normal distribution may be used such as: LN(2,0.5) as shown in Figure 12. Such a distribution can be used an initial distribution which is then updated to shift towards values of w (side length) that provided better SQIs in previous windows. For example, suppose that during the previous window there were 200 squares with SQI values and w values (si,wi), (s2,w2), ..., (s200, w200). Then the new distribution over w could be set to LN(k,0.5), where k=exp(∑i(wilog(si))/200).

    [0042] In step 701 the video sequence is divided into time windows as in the first embodiment, e.g. of 60 frames.

    [0043] In step 702 a standard dense optical flow algorithm (such as Horn & Schunk or Farneback) is applied to each pair of consecutive video frames. For each pair of frames this generates a vector field corresponding to two matrices, each of dimension equal to the video resolution, representing the x and y movement associated with each location in the image, respectively.

    [0044] In 703 the set of image axis-aligned square ROIs are defined as triples, (x,y,w) according to the distributions for location and size.

    [0045] The distribution over locations is a function of both the density matrix and image intensity gradients as mentioned above. For the gradient contribution the absolute values of the intensity gradients are calculated in the central frame of the time window. These values are then smoothed spatially using a 2-D box filter to form a smoothed matrix of intensity values.

    [0046] The distribution over (x,y) is then given by the density matrix divided by the smoothed matrix of intensity values (if all values in the density matrix are currently zero, e.g. as initialized, then a uniform distribution is used instead). The distribution thus favours image regions with a high density (density represents the quality of signals previously obtained from each image area) but with low image intensity gradients (i.e. favouring visually flatter image areas).

    [0047] The (x,y) coordinates for the required number of square regions of interest are then randomly drawn from that distribution and they define the ROI locations in the final overlapping frame of the time window. The locations in other frames of the window are then obtained in step 704 by updating them by the frame-to-frame vector fields obtained in step 702.

    [0048] In step 705 the density matrix is updated. First it undergoes a decay step in which each element in the density matrix is multiplied by some value, c, where 0<c<1 (larger values of c represent a less substantial decay. The value of c may be a constant. Alternatively c may depend on the extent of movement either globally (whole image) or locally (within the pixels near the element of the density matrix under consideration). Next the density matrix undergoes a growth step in which the elements of the density matrix near to signals with strong SQIs (see below) have their values increased. For each signal a Gaussian centred on the centre of the square to which that signal corresponds is added to the density matrix, with a weight that is proportional to the SQI corresponding to the signal and is on average about one tenth of the size of the density values.

    [0049] Each of the squares is a region of interest and, as with the first embodiment, as illustrated in step 110 the intensity in each region of interest in each frame is calculated (the sum of all the pixel intensity values) and the intensity for each square region of interest through the time window corresponds to a signal (il to im) to be processed. In visible light, for a camera outputting three R, G, B colour channels, only the green channel is used. However if the room is illuminated by infra-red light, the mean of the three colour channels is used. The image intensity of each ROI through the frame sequence will typically vary as schematically illustrated in Figure 2B. The mean location (xm, ym) of each square (for example the centre) is associated with the corresponding signal im over the time window. The intensity signals and associated locations are then output as time signals as illustrated in step 112.

    [0050] In a variation of this embodiment, the array of square ROIs are generated (in one frame) according to the distributions over location and size as above, but then for the movement of the ROIs through the frames of the time window (i.e. their locations in other frames in the time window) feature tracking (e.g. KLT tracking) is used and the movement of the square ROIs is set to match the mean movement undergone by the three tracks that were closest to the given square during the central frame of the time window. The location distribution is updated for each time window in the same way as above using the density matrix which has a time decay and a signal strength growth, and the image gradients in the central frame of the time window.

    Estimating physiological signals



    [0051] The intensity signals output from step 112 of Figure 3 or Figure 10 will be further analysed to attempt to detect PPG, and in particular heart rate, signals. Two ways of achieving this will be described below.

    [0052] In a first embodiment as illustrated in Figure 4, the intensity signals such as those schematically illustrated in Figure 2B are input in step 200 and the signals are scaled in the range of+1 to -1 and bandpass filtered in the expected physiological range for the heart rate of 0.8 to 2 Hz as illustrated in step 202. A standard bandpass filter such as a Butterworth filter may be used. The aim of steps 204 to 212 is then to combine together signals that are similar to form a smaller number of averaged signals. This is done by clustering, and in particular in step 204 a random pair of signals is selected and, if their Pearson correlation is greater than a threshold (for example 0.5) and if they are from locations within sixty pixels distance of each other in the image (Euclidean distance), the signals are averaged together and noted as being in the same cluster. The averaged signal then replaces the signals that formed it in the set of signals being considered for clustering.

    [0053] The process of selecting random pairs of signals, or averaged signals, continues until no more combinations can be formed. It should be noted that as indicated in step 206, when already averaged signals are averaged together, they are weighted (e.g. in proportion) according to the number of original signals that formed them. In step 208, clusters with fewer than an empirically-set number, typically about ten, signals contributing are discarded and then in step 210 signal quality indices of the surviving average signals are calculated.

    [0054] In this embodiment the signal quality index indicates how consistent the waveform is. For example such an SQI can be obtained by calculating the standard deviation of the peak-to-peak distance (SDpp) the standard deviation of the trough-to-trough distance (SDtt) and the standard deviation of the signal amplitude (SDaa). A single signal quality index SQI1 may be formed from these, e.g. as log SDamplitude + max of (log SDpp or log SDtt).

    [0055] If a cluster exists which has a large number of signals in it, but a poor SQI1, then in step 212 the averaged signal of that cluster is then subtracted from the original filtered signals output from step 202 and the clustering is re-performed. The subtraction is performed by linearly regressing each original signal against the average signal from the cluster, and each signal is replaced by its residuals from this regression, such that the correlation between the averaged signal and the result is zero. This step is effective to remove artefacts such as camera wobble or periodic lighting variations which affect the whole image.

    [0056] In step 214, the averaged signal with the best SQI1 is selected and, so as to remove spurious signals unlikely to relate to the cardiac cycle, accepted only if its SQI1 is greater than a predetermined threshold. If accepted the signal and its frequency can be measured to output a heart rate estimate (heart rate in beats per minute = 60 x frequency in Hertz). As illustrated in step 216 one way of obtaining the frequency is to perform a Fast Fourier Transform and look for the highest peak in the power spectrum. Alternatively, the average peak-to-peak and trough-to-trough distances can be used, optionally discarding the first and last peaks and troughs in the time window.

    [0057] If no cluster of signals survives the processing of Figure 4, or if none of the surviving clusters provides an SQI1 greater than the threshold, then no heart rate estimate will be made and instead an output indicating that no heart rate estimate is provided will be given. The processing then returns to step 200 for the next time window from the video image sequence (e.g. the next 9 second window moved along by an increment of 1 second).

    [0058] Figure 5 illustrates an alternative way of obtaining a heart rate estimate.

    [0059] Then in step 302 intensity signals from each of the square regions of interest are obtained as shown in Figure 3, and then in step 304 the signals are bandpass filtered in the passband corresponding to the expected physiological range of 0.8 to 2 Hz.

    [0060] The process then calculates two different signal quality indexes for each of the signals. The first signal quality index SQI1 is calculated in step 306 and is the same as the signal quality index calculated in step 210 of Figure 4 and explained above. It is therefore based on the peak-to-peak and trough-to-trough and amplitude variability in the signal. A second signal quality index SQI2 is obtained in step 308 to 314, this signal quality index relating to the strength of the main peak in the power spectrum of the signal. There are various ways of obtaining a measure of the strength of the highest peak in the power spectrum, but one way is as illustrated in steps 308 to 314. In step 308 the Fast Fourier Transforms of each of the signals is taken and in step 310, separately for each signal, the median power is calculated and added to the entire power spectrum. The natural log of the sum of the power and median power is calculated for each signal. Figure 9 illustrates an exemplary plot for one signal. It is the peaks visible in this plot which will be analysed. In step 314, the peak to trough distance each side of the highest peak PI are summed together (g1 + g2) and the peak to trough distance either side of the second highest peak P2 are summed together (g3 + g4), these sums forming peak quality measures. In step 314, the second signal quality index SQI2 is calculated as the ratio of, or difference between, the peak quality measures of the highest and second highest peaks

    [0061] The power spectrum, as illustrated in Figure 9, is also used to provide a measure of the dominant frequency in the signal, this being the frequency corresponding to the highest peak PI in the plot.

    [0062] Steps 308 to 314 therefore provide a second signal quality index SQI2 and a second estimate of frequency. These can be combined with the first signal quality index SQI1 and corresponding estimate of frequency obtained from the peak to peak, trough to trough and amplitude variability measures of step 306. As illustrated in step 316, the frequency of the signal is taken as a function of the two frequency estimates and the signal quality index of the signal, for example as:-

    where k is a constant, which is high for a signal which has two good individual SQIs and for which the frequency estimates are close to each other. The constant k may be, for example, 5 for frequencies F1, F2 measured in beats per minute.

    [0063] Alternatively, as illustrated in Figure 6, in step 324 each signal is multiplied by its signal quality index SQI, and then in step 326 principal component analysis is applied to the results. In step 328, for the first five principal components PC1 -PC5, a new quality index QI is calculated based on a combination of: from step 600 the principal component's SQI, from step 602 the distribution of the principal component's weights; and, from step 604, 604 the distance between the ROI locations associated with the signals that contribute most towards the principal component.

    [0064] The combination may be, for example, the sum A+B+C found by:

    Let "A" be the SQI for the principal component.

    Let "weights" be the vector of weights associated with the principal component.

    Then let B=1/(1-abs(sum(weights^3))/sum(abs(weights^3))), where abs(k) is the absolute value of k, for arbitrary k.

    Let "C" be the mean Euclidean distance between the ROI locations associated with the four signals for with the greatest values of abs(weights).



    [0065] In step 330 whichever of the first five principal components PC1 -PC5 has the best quality index QI is selected and a frequency estimate and a quality estimate are output. The frequency estimate can be simply the highest peak in the power spectrum of the selected principal component, or can be obtained by measuring average peak-to-peak and trough-to-trough distances, or by taking the average of the two frequencies obtained from these methods and the principal component quality index can be obtained by taking the square of the principal component quality calculated in step 328 and dividing it by the principal component quality of the second best principal component.

    [0066] The frequency estimate will be used to output a heart rate estimate (the heart rate in beats per minute equals 60 times the frequency in Hertz), and the quality index is output as a measure of the confidence of the measurement. The processing will then be repeated for the next time window.

    [0067] The invention may be embodied in a signal processing method, or in a signal processing apparatus which may be constructed as dedicated hardware or by means of a programmed general purpose computer or programmable digital signal processor. The invention also extends to a computer program for executing the method.


    Claims

    1. A computer-implemented method of obtaining an estimate of a periodic vital sign of a subject (3) from a video image sequence of the subject (3), comprising the steps of:

    detecting an image area with a strong intensity gradient in a frame of the video image sequence;

    defining a plurality of regions of interest (24) in the frame of the video sequence, the regions of interest (24) being defined not to include said image area;

    tracking the regions of interest (24) through other frames of the video image sequence forming a time window consisting of a predetermined number of frames;

    detecting intensity variations in said region of interest (24) through the image sequence to form respective time series signals (il-im) and obtaining an estimate of said periodic vital sign from said time series signals(il-im);

    characterized in that:
    the steps of detecting an image area with a strong intensity gradient and tracking the regions of interest through other frames of the video image sequence comprise detecting and tracking image feature points (20) through the sequence; defining a set of persistent tracks as the set of all image feature point tracks that span all frames in the time window; and the step of defining regions of interest (24) comprises defining regions of interest (24) each of which is entirely within an area of the image between the tracked image feature points (20) forming the persistent tracks and which does not overlap the feature points (20).
     
    2. A method according to claim 1 wherein the regions of interest are defined as squares aligned with orthogonal axes of the frames of the video image sequence.
     
    3. A method according to claim 1 or 2 wherein the step of detecting an image area with a strong intensity gradient comprises detecting an image area with an intensity gradient stronger than a predetermined threshold.
     
    4. A method according to any one of the preceding claims wherein the step of tracking the regions of interest through other frames of the video image sequence comprises defining the position of the regions of interest in other frames of the video image sequence by reference to detected image movement in the video image sequence.
     
    5. A method according to claim 4 wherein image movement in the video image sequence is detected by measuring the optical flow in the video image sequence.
     
    6. A method according to claim 1 further comprising the step of defining a grid of image areas whose sides join the feature points and wherein each region of interest is defined to be entirely within a respective one of said image areas.
     
    7. A method according to claim 6 wherein the image areas are polygons whose vertices are at the feature points.
     
    8. A method according to claim 6 or 7 wherein the step of defining the grid of image areas comprises defining the grid of image areas on one frame of the sequence and forming grids on the other frames of the sequence by joining the same feature points together.
     
    9. A method according to claim 6, 7 or 8, wherein the grid is triangular, each polygonal image area being a triangle, optionally wherein the regions of interest are defined by forming in-circles of said image areas and optionally wherein the regions of interest are defined as squares co-centered on the in-circles.
     
    10. A method according to any one of the preceding claims further comprising the step of calculating a signal quality index representing the strength in said time series signals of said periodic vital sign and combining estimates from the regions of interest in dependence upon the signal quality index.
     
    11. A method according to any one of the preceding claims further comprising the step of clustering said time series signals to form clusters of time series signals which have greater than a predetermined correlation and are obtained from regions of interest spaced by no more than a predetermined distance in the image, averaging the signals in each cluster, and obtaining the estimate of the periodic vital sign from the averaged signals.
     
    12. A method according to any one of the preceding claims wherein the estimate of the periodic vital sign is obtained by measuring the frequency, or the frequency of the strongest periodic component, of said time series signals or averaged signals, optionally further comprising the step of applying principal component analysis to the time series signals or averaged time series signals, calculating a signal quality index of the principal components and obtaining the estimate by measuring the frequency, or the frequency of the strongest periodic component, of the principal component with the best signal quality index.
     
    13. A method according to any one of the preceding claims wherein the intensity variations include a periodic component corresponding to a photoplethysmogram signal and wherein the periodic vital sign is the heart rate or breathing rate.
     
    14. Apparatus for estimating a periodic vital sign of a subject comprising:

    a video camera (5) for capturing a video image sequence of the subject;

    an image data processor (7) configured for executing the method of any one of the preceding claims; and

    a display (9) for displaying the estimate of the periodic vital sign.


     
    15. A computer program comprising program code means adapted to cause a computer system to perform the method of any one of claims 1 to 13 when executed on said computer system
     


    Ansprüche

    1. Computerimplementiertes Verfahren zum Erhalten einer Schätzung eines periodischen Vitalzeichens eines Subjekts (3) aus einer Videobildsequenz des Subjekts (3), wobei das Verfahren folgende Schritte umfasst:

    Erfassen eines Bildbereichs mit einem starken Intensitätsgradienten in einem Rahmen der Videobildsequenz;

    Definieren einer Vielzahl von interessierenden Regionen (24) in dem Rahmen der Videosequenz, wobei die interessierenden Regionen (24) so definiert sind, dass sie den Bildbereich nicht enthalten;

    Verfolgen der interessierenden Regionen (24) durch andere Rahmen der Videobildsequenz, die ein Zeitfenster bilden, das aus einer vorbestimmten Anzahl von Rahmen besteht;

    Erfassen von Intensitätsschwankungen in dem interessierenden Region (24) durch die Bildsequenz, um entsprechende Zeitreihensignale (il-im) zu bilden, und Erhalten einer Schätzung des periodischen Vitalzeichens aus den Zeitreihensignalen (il-im);

    dadurch gekennzeichnet, dass:
    die Schritte des Erfassens eines Bildbereichs mit einem starken Intensitätsgradienten und des Verfolgens der interessierenden Regionen durch andere Rahmen der Videobildsequenz Folgendes umfassen: Erfassen und Verfolgen von Bildmerkmalspunkten (20) durch die Sequenz; Definieren eines Satzes persistenter Spuren als Satz aller Bildmerkmalspunktspuren, die alle Rahmen im Zeitfenster umfassen; und der Schritt des Definierens von interessierenden Regionen (24) das Definieren von interessierenden Regionen (24) umfasst, von denen sich jede vollständig innerhalb eines Bildbereichs zwischen den verfolgten Bildmerkmalspunkten (20) befindet, die die dauerhaften Spuren bilden, und die die Merkmalspunkte (20) nicht überlappt.
     
    2. Verfahren nach Anspruch 1, wobei die interessierenden Regionen als Quadrate definiert sind, die mit orthogonalen Achsen der Rahmen der Videobildsequenz ausgerichtet sind.
     
    3. Verfahren nach Anspruch 1 oder 2, wobei der Schritt des Erfassens eines Bildbereichs mit einem starken Intensitätsgradienten das Erfassen eines Bildbereichs mit einem Intensitätsgradienten umfasst, der stärker als ein vorbestimmter Schwellenwert ist.
     
    4. Verfahren nach einem der vorhergehenden Ansprüche, wobei der Schritt des Verfolgens der interessierenden Regionen durch andere Rahmen der Videobildsequenz das Definieren der Position der interessierenden Regionen in anderen Rahmen der Videobildsequenz unter Bezugnahme auf die erfasste Bildbewegung in der Videobildsequenz umfasst.
     
    5. Verfahren nach Anspruch 4, wobei die Bildbewegung in der Videobildsequenz durch Messen des optischen Flusses in der Videobildsequenz erfasst wird.
     
    6. Verfahren nach Anspruch 1, das ferner den Schritt des Definierens eines Gitters von Bildbereichen umfasst, deren Seiten die Merkmalspunkte verbinden, und wobei jede interessierende Region so definiert ist, dass sie sich vollständig in einem der jeweiligen Bildbereiche befindet.
     
    7. Verfahren nach Anspruch 6, wobei die Bildbereiche Polygone sind, deren Eckpunkte an den Merkmalspunkten liegen.
     
    8. Verfahren nach Anspruch 6 oder 7, wobei der Schritt des Definierens des Rasters von Bildbereichen das Definieren des Rasters von Bildbereichen in einem Rahmen der Sequenz und das Bilden von Gittern auf den anderen Rahmen der Sequenz durch Zusammenfügen derselben Merkmalspunkte umfasst.
     
    9. Verfahren nach Anspruch 6, 7 oder 8, wobei das Gitter dreieckig ist, wobei jeder polygonale Bildbereich ein Dreieck ist; optional, wobei die interessierenden Regionen durch Bilden von Innenkreisen der Bildbereiche definiert werden und optional, wobei die interessierenden Regionen als Quadrate definiert sind, die auf den Innenkreisen zentriert sind.
     
    10. Verfahren nach einem der vorhergehenden Ansprüche, das ferner Folgendes umfasst:
    den Schritt des Berechnens eines Signalqualitätsindex, der die Stärke in den Zeitreihensignalen des periodischen Vitalzeichens darstellt, und des Kombinierens von Schätzungen aus den interessierenden Regionen in Abhängigkeit vom Signalqualitätsindex.
     
    11. Verfahren nach einem der vorhergehenden Ansprüche, das ferner Folgendes umfasst: den Schritt des Clusterns der Zeitreihensignale, um Cluster von Zeitreihensignalen zu bilden, die eine größere als eine vorbestimmte Korrelation aufweisen und aus interessierenden Regionen erhalten werden, die um nicht mehr als einen vorbestimmten Abstand im Bild voneinander beabstandet sind, des Mittelns der Signale in jedem Cluster und des Erhaltens der Schätzung des periodischen Vitalzeichens aus den gemittelten Signalen.
     
    12. Verfahren nach einem der vorhergehenden Ansprüche, wobei die Schätzung des periodischen Vitalzeichens durch Folgendes erhalten wird: Messen der Frequenz oder der Frequenz der stärksten periodischen Komponente der Zeitreihensignale oder gemittelten Signale, optional ferner umfassend den Schritt der Anwendung der Hauptkomponentenanalyse auf Zeitreihensignale oder gemittelte Zeitreihensignale, Berechnen eines Signalqualitätsindex der Hauptkomponenten und Erhalten der Schätzung durch Messen der Frequenz oder der Frequenz der stärksten periodischen Komponente der Hauptkomponente mit dem besten Signalqualitätsindex.
     
    13. Verfahren nach einem der vorhergehenden Ansprüche, wobei die Intensitätsschwankungen eine periodische Komponente umfassen, die einem Photoplethysmogrammsignal entspricht, und wobei das periodische Vitalzeichen die Herzfrequenz oder Atemfrequenz ist.
     
    14. Vorrichtung zum Schätzen eines periodischen Vitalzeichens eines Subjekts, die Folgendes umfasst: eine Videokamera (5) zum Erfassen einer Videobildsequenz des Subjekts;
    einen Bilddatenprozessor (7), der zum Ausführen des Verfahrens nach einem der vorhergehenden Ansprüche konfiguriert ist; und
    eine Anzeige (9) zum Anzeigen der Schätzung des periodischen Vitalzeichens.
     
    15. Computerprogramm, das Programmcodemittel umfasst, die dafür ausgelegt sind, ein Computersystem zu veranlassen, das Verfahren nach einem der Ansprüche 1 bis 13 auszuführen, wenn es auf dem Computersystem ausgeführt wird.
     


    Revendications

    1. Procédé implémenté par ordinateur de l'obtention d'une estimation d'un signe vital périodique d'un sujet (3) à partir d'une séquence d'images vidéo du sujet (3), comprenant les étapes de :

    la détection d'une zone d'image avec un fort gradient d'intensité dans une vue de la séquence d'images vidéo ;

    la définition d'une pluralité de régions d'intérêt (24) dans la vue de la séquence vidéo, les régions d'intérêt (24) étant définies pour ne pas inclure ladite zone d'image ;

    le dépistage des régions d'intérêt (24) dans d'autres vues de la séquence d'images vidéo formant une fenêtre temporelle constituée d'un nombre prédéterminé de vues ;

    la détection de variations d'intensité dans ladite région d'intérêt (24) dans la séquence d'images pour former des signaux de série temporelle respectifs (il-im) et l'obtention d'une estimation dudit signe vital périodique à partir desdits signaux de série temporelle (il-im) ;

    caractérisé en ce que :

    les étapes de la détection d'une zone d'image avec un fort gradient d'intensité et le dépistage des régions d'intérêt dans d'autres vues de la séquence d'images vidéo comprennent la détection et le dépistage de points caractéristiques d'image (20) dans la séquence ; la définition d'un ensemble de pistes persistantes en tant qu'ensemble de toutes les pistes de points caractéristiques d'image qui couvrent toutes les vues dans la fenêtre temporelle ; et l'étape de la définition de régions d'intérêt (24) comprend la définition de régions d'intérêt (24) dont chacune est entièrement à l'intérieur d'une zone de l'image entre les points caractéristiques d'image dépistés (20) formant les pistes persistantes et ne chevauche pas les points caractéristiques (20).


     
    2. Procédé selon la revendication 1, dans lequel les régions d'intérêt sont définies sous forme de carrés alignés avec des axes orthogonaux des vues de la séquence d'images vidéo.
     
    3. Procédé selon la revendication 1 ou 2, dans lequel l'étape de la détection d'une zone d'image avec un fort gradient d'intensité comprend la détection d'une zone d'image avec un gradient d'intensité plus fort qu'un seuil prédéterminé.
     
    4. Procédé selon l'une quelconque des revendications précédentes, dans lequel l'étape du dépistage des régions d'intérêt dans d'autres vues de la séquence d'images vidéo comprend la définition de la position des régions d'intérêt dans d'autres vues de la séquence d'images vidéo par référence à un mouvement d'image détecté dans la séquence d'images vidéo.
     
    5. Procédé selon la revendication 4, dans lequel le mouvement d'image dans la séquence d'images vidéo est détecté en mesurant le flux optique dans la séquence d'images vidéo.
     
    6. Procédé selon la revendication 1, comprenant en outre l'étape de la définition d'une grille de zones d'image dont les côtés joignent les points caractéristiques et dans lequel chaque région d'intérêt est définie pour être entièrement à l'intérieur d'une respective de ladite zones d'image.
     
    7. Procédé selon la revendication 6, dans lequel les zones d'image sont des polygones dont les sommets sont aux points caractéristiques.
     
    8. Procédé selon la revendication 6 ou 7, dans lequel l'étape de la définition de la grille de zones d'image comprend la définition de la grille de zones d'image sur une vue de la séquence et la formation de grilles sur les autres vues de la séquence en joignant les mêmes points caractéristiques ensemble.
     
    9. Procédé selon la revendication 6, 7 ou 8, dans lequel la grille est triangulaire, chaque zone d'image polygonale étant un triangle, optionnellement dans lequel les régions d'intérêt sont définies en formant des cercles internes desdites zones d'image et optionnellement dans lequel les régions d'intérêt sont définies sous forme de carrés co-centrés sur les cercles internes.
     
    10. Procédé selon l'une quelconque des revendications précédentes, comprenant en outre l'étape du calcul d'un indice de qualité de signal représentant l'intensité dans lesdits signaux de série temporelle dudit signe vital périodique et la combinaison d'estimations à partir des régions d'intérêt en fonction de l'indice de qualité de signal.
     
    11. Procédé selon l'une quelconque des revendications précédentes, comprenant en outre l'étape de la mise en grappe desdits signaux de série temporelle pour former des grappes de signaux de série temporelle qui ont plus qu'une corrélation prédéterminée et sont obtenues à partir de régions d'intérêt espacées par non plus d'une distance prédéterminée dans l'image, le moyennage des signaux dans chaque grappe, et l'obtention de l'estimation du signe vital périodique à partir des signaux moyennés.
     
    12. Procédé selon l'une quelconque des revendications précédentes, dans lequel l'estimation du signe vital périodique est obtenue en mesurant la fréquence, ou la fréquence de la composante périodique la plus forte, desdits signaux ou signaux moyennés de série temporelle, optionnellement comprenant en outre l'étape de l'application d'analyse en composantes principales sur les signaux de série temporelle ou signaux moyennés de série temporelle, le calcul d'un indice de qualité de signal des composantes principales et l'obtention de l'estimation en mesurant la fréquence, ou la fréquence de la composante périodique la plus forte, de la composante principal avec le meilleur indice de qualité de signal.
     
    13. Procédé selon l'une quelconque des revendications précédentes, dans lequel les variations d'intensité incluent une composante périodique correspondant à un signal de photopléthysmogramme et dans lequel le signe vital périodique est la fréquence cardiaque ou la fréquence respiratoire.
     
    14. Appareil pour estimer un signe vital périodique d'un sujet, comprenant :

    une caméra vidéo (5) pour capturer une séquence d'images vidéo du sujet ;

    un processeur de données d'image (7) configuré pour exécuter le procédé de l'une quelconque des revendications précédentes ; et

    un écran d'affichage (9) pour afficher l'estimation du signe vital périodique.


     
    15. Programme d'ordinateur, comprenant des moyens codes de programme adaptés pour faire en sorte qu'un système d'ordinateur réalise le procédé de l'une quelconque des revendications 1 à 13, lorsqu'ils sont exécutés sur ledit système d'ordinateur.
     




    Drawing

































    REFERENCES CITED IN THE DESCRIPTION



    This list of references cited by the applicant is for the reader's convenience only. It does not form part of the European patent document. Even though great care has been taken in compiling the references, errors or omissions cannot be excluded and the EPO disclaims all liability in this regard.

    Patent documents cited in the description




    Non-patent literature cited in the description