Radar sensor with self-learning for motion detection

Abstract

The present disclosure relates to a method for calibrating a radar motion sensor device in which a temporal sequence of samples of a radar data signal is obtained, each sample indicating a first power level in a first frequency band of a radar signal detected by the radar motion sensor device. A pattern recognition algorithm is applied to the samples to distinguish between motion in the room and a passing-by event comprising motion past the room. The samples which correspond to a passing-by event and when the room is vacant are identified. These identified samples are used (i) to update a threshold for identifying motion in the room and / or (ii) for identifying when a room has been vacated. The updated thresholds are used to perform motion detection and / or it is identified when a room has been vacated, in a manner which distinguishes between motion in the room and a passing-by event.

Description

[0001] Radar sensor with self-learning for motion detection

[0002] FIELD OF THE INVENTION

[0003] The present invention generally relates to a method for calibrating a radar sensor for motion detection and a radar sensor comprising a control module configured to perform the method.

[0004] BACKGROUND OF THE INVENTION

[0005] Motion detectors are used in many applications ranging from burglar alarms to automatic door openers or automatic control of illumination or HVAC systems. For example, motion detectors are often used in offices, e.g. in conference rooms, to detect when people are present to steer light modules and / or HVAC systems such that lights, ventilation, heating and air conditioning are turned on when people are present and automatically turned off when no people are present. Automatic control of these systems based on motion detection has the potential to save large amounts of energy since e.g. the risk of lights or air conditioning units in an office remaining on during the night, or over the weekend, when no people are present is reduced.

[0006] Typically, Passive Infrared Sensors, PIRs, are used to detect motion. PIRs operate by registering incident IR radiation and when e.g. people move in front of the PIR the intensity distribution on the PIR sensor changes which can be taken as an indication that people are present in front of the sensor. A drawback with PIRs is, however, that they are not very sensitive and small motions, e.g. small hand gestures or small head motions, made by people sitting around a conference table may not always be detected by the PIR whereby the lights, or others systems, are deactivated while people are still using the conference room. In other words, the PIRs are prone to high false negative rates meaning that PIRs often fail to sense motion when there in fact is motion to detect.

[0007] In view of the drawbacks of PIRs, radar sensors have been used instead. Radar sensors are active sensors which transmit a radar signal into the environment and measure the reflected radar signal. If an object (e.g. a human) moves this will alter the reflected radar signal and this alteration can be used as an indication of motion. Radar sensors tend to be much more sensitive compared to PIRs such that even small motions can be detected which reduced the false negative rate. At the same time, radar signals can penetrate non-metal materials such as glass, plastic or drywall meaning that it is possible to hide radar motion sensors in walls or integrate the radar motion sensor into devices without the radar sensor being visible.

[0008] On the other hand, radar sensors are in some scenarios too sensitive leading to an increase in false positive rates. That is, radar sensors are prone to detecting motion even when no people are present. For example, since radar signals can penetrate drywall and glass, it is possible that motion in a completely separate room is detected by the radar sensor or that a person passing-by a room equipped with a radar sensor is registered by the radar sensor. To circumvent the issues with high false positive rates one common approach is to use a sensitivity setting that allows users to modify the sensitivity of the radar sensor. However, manually adjusting the sensitivity setting is a trial-and-error method and thus not user friendly. For example, the user may need to try a number of sensitivity levels before determining the right sensitivity setting, and sometimes users cannot identify one level to achieve satisfying false positive and false negative rates.

[0009] US2017 / 123058A1 discloses using a radar sensor to detect a presence of a person in the area of coverage based on the analysis of reflected radio signal emitted by the radar sensor. The analysis of the processed signals can be performed in both time and frequency domain. In addition to radar, an input from an infrared sensor can also be used in conjunction with radar based detection.

[0010] EP3511736B1 discloses a motion sensing method includes monitoring for a first motion in a first region using a first antenna using a first motion detection parameter, when no first motion is sensed by the monitoring using the first antenna monitoring for a second motion in a second region using a second antenna using a second motion detection parameter, and when no second motion is sensed by monitoring using the second antenna, designating a space, which encompasses the second region, as unoccupied, wherein the first region and the second region overlap one another, and the first motion detection parameter is different from the second motion detection parameter.

[0011] EP3928122B9 discloses a motion detector configured to signal-process a motion signal to obtain multiple motion components, a motion component being correlated with a direction and velocity of a motion in the environment; cancel in motion components, two motion components that correspond to motions with opposite directions; and detect motion in the environment from the motion components that are not cancelled. In view of these shortcomings there is a need for an improved method for calibrating a radar motion sensor and an improved radar motion sensor.

[0012] The applicant has proposed, but not yet published, a self-learning method for a radar based motion sensor to largely reduce false triggers if the sensor is installed in a meeting room or small office room. For example, a radar sensor may be installed on the ceiling of a first room to realize automatic lighting control (i.e., light-on-demand). Light will be switched on automatically only when the radar sensor detects a valid motion, e.g., a person walking into the first room. On the other hand, a light should not be switched on if a person is walking into an adjacent second room.

[0013] The proposed self learning method involves application of a learning process after being installed. During the learning process, the radar sensor tries to capture a number of potential valid motions. Each potential valid motion contains the time series sensor signal corresponding to a person walking into either of the two rooms. After enough samples of potential valid motions are captured, they are feed into a clustering module and are separated into valid motions (i.e., walking into the first room) and invalid motions (walking into the second room). Finally, the radar sensor updates its motion detection algorithm (e.g., by updating relevant thresholds according to the output of the clustering module) so that only valid motions are detected.

[0014] However the proposed approach does not take account of the motions of passing-by the two rooms. In particular, a light should also not be switched on if a person is walking along a corridor outside the room (i.e., passing-by the room).

[0015] SUMMARY OF THE INVENTION

[0016] It is an object of the present invention to provide a new and improved method for calibrating a radar motion sensor. This and other objects are achieved by the method and radar motion sensor of the independent claims. Embodiments of the present invention are defined in the dependent claims.

[0017] According to an aspect of the present invention, there is provided a method for calibrating a radar motion sensor device located in a room, comprising: obtaining a temporal sequence of samples of a radar data signal from a radar motion sensor, each sample indicating a first power level in a first frequency band of a radar signal (6) detected by the radar motion sensor device; applying a pattern recognition algorithm to the samples to distinguish between motion in the room and a passing-by event comprising motion past the room; identifying those samples which correspond to a passing-by event when the room is vacant; using the identified samples to update a first threshold for identifying motion in the room or for identifying when a room has been vacated; and using the updated first threshold to perform motion detection and / or identifying when a room has been vacated with the radar motion sensor device, in a manner which distinguishes between motion in the room and a passing-by event.

[0018] This method is used either as part of a learning phase for setting one or more thresholds which will be used to identify movement in a room using a radar sensor or it is used to determine when a room has been vacated, for example for automatically turning off the lights. The one or more thresholds are selected such that passing-by events are not falsely recognized as movement in the room being monitored, so that these passing-by events do not result is false triggers, or delay the turning off of lights in a vacated room. This is achieved by detecting passing-by events during the learning phase and / or during use of the radar sensor by means of a pattern recognition algorithm. When used during the learning phase, this pattern recognition algorithm can take account of the radar data signal samples before and after the detection event, and this enables passing-by events to be detected.

[0019] A passing-by event may involve a person walking past an entrance to the room, e.g. along a corridor, or it may be a person walking along a glass wall or stud plasterboard wall of the room, This may also give a false trigger of the motion detection.

[0020] The radar signal levels associated with the passing-by event can then be used to adapt thresholds used for motion detection such that motion detection (for detecting motion in a room) is not triggered by a passing-by event. The thresholds for motion detection may for example comprise one or more thresholds applied to different frequency bands of the radar signal.

[0021] For example, distinguishing or detecting a passing-by event may comprise detecting a decreasing distance to the radar motion sensor device followed by an increasing distance to the radar motion sensor device. It is noted that for a single channel Doppler radar, distance is not measured directly. Rather, the Doppler shift caused by the motion is calculated by applying a FFT to the raw sensor data. Either the power of the raw sensor signal or the Doppler shift profile in the frequency-time domain indicates distance changes caused by a passing-by event. In this way, analysis of radar signal is used to identify that a passing-by event has been detected. Detecting a passing-by event may comprise detecting a symmetrically decreasing distance and increasing distance with regarding to time. This corresponds to an object (person) moving at constant speed past the radar sensor.

[0022] Detecting a passing-by event may comprise: converting the temporal sequence to the frequency domain; calculating a power in a first frequency range; deriving a first smoothed power signal for the first frequency range; and checking if the first smoothed power signal rises to cross a second threshold and then falls to cross the second threshold within a predefined duration, thereby defining a crossing window.

[0023] This crossing window indicates the presence of the decreasing then increasing distance discussed above.

[0024] Detecting a passing-by event may further comprise: calculating the power in a second frequency range, higher than the first frequency range; deriving a second smoothed power signal for the second frequency range; comparing a time delay, between a maximum of the first smoothed power signal and a maximum of the second smoothed power signal within the crossing window, with a time delay threshold, wherein a passing-by event is detected when the time delay is below the time delay threshold.

[0025] Detecting a passing-by event for example comprises detecting the room is vacant before the passing-by event by determining that the first and second smoothed power signals are below a third threshold.

[0026] As explained above, one use of the passing-by event detection is to use the identified samples for identifying when a room has been vacated. This may involve: deriving a silence measure which increases or decreases when there is no motion detection and decreases or increases, respectively, when there is motion detection; and comparing the silence measure with a fourth threshold to determine when there has been sufficiently low motion that the room can be determined to be vacated, wherein the method comprises discounting motion detection when it is determined to be detection of a passing-by event.

[0027] In this way, a silence detection approach is improved by identifying when interruptions to the silence are caused by passing-by events rather than movement in the room. By ignoring those passing-by events, it can be determined more quickly that a room has been vacated so that the lights can be turned off in an efficient manner.

[0028] The silence measure is for example a sum of binary silence indices each for an associated time window, wherein each silence index has a first value when there is motion detection and a second value when there is no motion detection, wherein the method comprises swapping the first value for the second value before performing the sum, when motion of a passing-by event is detected. Thus, the standard silence detection method can be used, but passing-by events are interpreted as no motion.

[0029] As explained above, another use of the passing-by event detection is to set or update the threshold (or multiple thresholds) used for detecting motion in the room. The method may for example comprise updating the threshold for identifying motion in the room by: identifying a plurality of valid trigger samples by comparing the first power level of each sample to a first predetermined threshold, wherein each valid trigger sample is identified as a sample having a first power level that exceeds the first predetermined threshold and is followed by a trailing event in a predetermined time duration after the sample; analyzing samples of a trigger context window associated with each of the plurality of valid trigger samples to determine a first updated threshold (T F,f).

[0030] This approach is based on detecting valid trigger samples, for use in setting detection thresholds, as trigger samples associated with a trailing event indicating some level of activity after the trigger sample which exceeds a first predetermined threshold.

[0031] The first updated threshold may be equal to or higher than the first predetermined threshold. In some implementations, the first updated threshold is strictly higher than the first predetermined threshold.

[0032] Analyzing the samples of a trigger context window associated with each of the plurality of valid trigger samples may comprise determining a first candidate threshold for each valid trigger sample based on the samples of the trigger context window; and clustering the first candidate thresholds into at least one group, each group being associated with a group center level. Wherein the first updated threshold is set to a value that separates the group associated with a highest group center level. In some cases, the value may exceed the first predetermined threshold. In some cases, the value may equal to the first predetermined threshold. Preferably, the candidate threshold is set higher than the predetermined threshold so as to decrease false positives meaning that when the updated threshold is set it can be set higher than the predetermined first threshold but still at an appropriate level that separates the cluster of all collected candidate thresholds.

[0033] The clustering may comprise clustering the plurality of first candidate thresholds into at least two groups and setting the first updated threshold to separate the group associated with the highest group center level from the group associated with a second highest group center level.

[0034] By using two or more groups in the clustering it is possible to set accurate updated threshold levels even for radar motion sensors installed in places that are challenging for accurate motion detection by radar. For example, if the radar motion sensor is installed in a first conference room having a transparent wall or window into a neighboring second conference room the radar motion sensor may detect valid trigger samples also when people are moving the second conference room. However, by clustering the candidate thresholds into at least two groups an updated threshold can be identified to separate detected motions in the first room from detected motions in the second room. In general, the radar signals resulting from motion in the first room will be more powerful compared to the radar signals resulting from motion in the second room due to e.g. transmission losses through the transparent wall or window and increased propagation distance for the radar signals.

[0035] The first candidate threshold is for example greater than the first predetermined threshold and smaller than the maximum first power level in the trigger context window.

[0036] That is, the first candidate threshold is set such that at least one trigger sample can still be identified in the trigger context window.

[0037] The method may further comprise identifying a plurality of trigger samples by comparing the first power level of each sample to a first predetermined threshold, wherein each trigger sample is identified as a sample having a first power level that exceeds the first predetermined threshold. For each identified trigger sample the method further comprises defining a trailing context window associated with the trigger sample comprising a predetermined number of samples temporally later than the trigger sample and determining, for each sample in the trailing context window associated with the trigger sample whether the first power level of the sample exceeds a predetermined trailing threshold. The method further comprises classifying the trigger sample as a valid trigger sample when the percentage of samples in the trailing context window exceed a percentage threshold. Accordingly, trigger samples are first detected and, subsequently, it is determined if they are to be treated as valid trigger samples if the trailing event has occurred after the trigger sample. The predetermined trailing threshold may equal the first predetermined threshold.

[0038] It has been realized that the first predetermined threshold can be used as the trailing threshold. Accordingly, if the first power level remains higher (for at least a predetermined percentage of the samples) than the first predetermined threshold also during the trailing context window it can be established that a trailing event has occurred. The predetermined percentage may e.g. be 80% or 90%, although other percentages can also be used, as will be described in the below.

[0039] The trigger context window for example comprises a predetermined number of samples temporally earlier and / or later than the valid trigger sample. For example, the trigger context window comprises temporally earlier and / or later samples extending to a point in time at least 1 second earlier and / or later than the trigger sample, preferably at least 5 second earlier and / or later than the trigger sample, more preferably at least 10 seconds earlier and / or later than the trigger sample and most preferably at least 20 seconds earlier and / or later than the trigger sample.

[0040] That is, the trigger context window captures the temporal context surrounding the trigger sample and includes at least one of samples that are temporally earlier than the trigger sample and samples that are temporally later than the trigger sample. When determining the candidate threshold many samples can thus be taken into account whereby candidate thresholds can be found that, if used as replacement for the first predetermined threshold, would still enable a trigger sample to be identified in the trigger context window.

[0041] Each sample may further indicate a second power level in a second frequency band of the radar signal detected by the radar motion sensor device, the second frequency band being different from the first frequency band. Accordingly, two frequency bands, a first frequency band and a second frequency band may be considered at the same time for enhanced accuracy. Optionally, the second frequency band contains higher frequencies than the first frequency band whereby the first frequency band may be referred to as a low frequency band and the second frequency band may be referred to as a high frequency band.

[0042] The method may further comprise identifying a plurality of trigger samples by comparing the first power level and the second power level of each sample to a first predetermined threshold and second predetermined threshold respectively, wherein each trigger sample is identified as a sample having a first power level and second power level that exceeds both the first predetermined threshold and the second predetermined threshold.

[0043] That is, the trigger sample may be detected with a stricter criteria requiring the power in both frequency bands to exceed a respective predetermined threshold. It has been realized that when people enter a room there are both low and high frequency components registered by the radar motion sensor. Accordingly, by using this stricter criteria the false positive rate can be reduced further compared to using only one frequency band and predetermined threshold for identifying the trigger samples.

[0044] The method may further comprise, for each valid trigger sample, determining, a second candidate threshold based on the samples of the trigger context window, wherein the second candidate threshold is greater than the second predetermined threshold and smaller than the maximum second power level in the trigger context window. The method further comprising determining a second updated threshold by clustering a plurality of second candidate thresholds into at least one second group of second candidate thresholds, wherein each second group of second candidate thresholds is associated with a respective second group center level, and setting the second updated threshold to a value that separates the second group associated with a highest second group center level and using the first and second updated threshold to perform motion detection with the radar motion sensor device. In some cases, the value may exceed the second predetermined threshold. In some cases, the value may equal to the second predetermined threshold.

[0045] That is, the second candidate thresholds and the second updated threshold is determined analogously to the first candidate thresholds and the first updated threshold. In this way, an updated threshold which offers a reduced false positive rate, without increased false negatives, may be determined automatically for both frequency bands. Optionally, determining a second updated threshold comprises clustering the plurality of second candidate thresholds into at least two second groups and setting the second updated threshold to separate the second group associated with the highest second group center level from the second group associated with a second highest second group center level.

[0046] Using the first updated threshold and second updated threshold to perform motion detection may comprise obtaining a sample of a second radar data signal, the sample indicating the first power level and second power level, comparing the first power level to the first updated threshold, comparing the second power level to the second updated threshold, and outputting a signal indicating that motion has been detected in response to the first power level exceeding the first updated threshold and the second power level of the sample exceeding the second updated threshold. The second radar data signal may be a temporally later portion of an original radar data signal comprising both the radar data signal and the second radar data signal. It is understood that the radar motion sensor may continuously record a radar data signal and, once the updated threshold has been determined the updated threshold is used to detect that motion has occurred for future segments of the continuously recorded radar data signal.

[0047] In embodiments where only one frequency band is used and the first updated threshold is used to perform motion detection, the method may further comprise: obtaining a sample of a second radar data signal, the sample indicating the first power level, comparing the first power level to the first updated threshold and outputting a signal indicating that motion has been detected in response to the first power level of the sample exceeding the first updated threshold.

[0048] According to a second aspect of the invention there is provided a computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of the method according to the first aspect.

[0049] According to a third aspect of the invention, a controller for a radar motion sensor, the controller comprising a processor and a memory, wherein the processor is configured to perform the method according to the first aspect.

[0050] In some implementations, two consecutive trigger samples are separated by at least one trailing context window.

[0051] Accordingly, after one trigger sample has been identified, the method starts to look for the next trigger sample starting with samples that are at least the duration of the trailing context window temporally later than the previous trigger sample. As a consequence, the computational complexity is reduced since some samples may be ignored when looking for a next trigger sample after a previous trigger sample has been identified. For each trigger sample it is determined if it is a valid trigger sample by determining if a trailing event has occurred in the trailing context window.

[0052] In embodiments, when both frequency bands are used, the first and second candidate threshold is determined such that at least one sample in the trailing context window has a first and second power lever exceeding the first and second candidate threshold respectively.

[0053] The first and second candidate thresholds are hereby determined together such that at least one sample in the trigger context window fulfills the stricter criteria of being associated with a first and second power level that exceeds both the first and second candidate threshold. In this way, the first and second candidate thresholds can be set as high as possible while still ensuring that the trigger sample is identified in the trigger context window.

[0054] The invention according to the second and third aspect features the same or equivalent benefits as the invention according to the first aspect. Any functions described in relation to the method, may have corresponding features in a controller or a computer program product.

[0055] BRIEF DESCRIPTION OF THE DRAWINGS

[0056] This and other aspects of the present invention will now be described in more detail, with reference to the appended drawings showing embodiments of the present invention.

[0057] Fig. 1 shows a top-down view of an exemplary room-layout that could increase the likelihood for false positives for a radar motion sensor.

[0058] Fig. 2 is a flowchart illustrating the first learning phase, second learning phase and use phase for calibrating a radar motion sensor according to some implementations.

[0059] Fig. 3 is a graph showing an example of a raw radar signal from a radar motion sensor.

[0060] Fig. 4 is a graph showing the lower and high frequency spectral power of the raw radar signal.

[0061] Fig. 5 is a block diagram showing the processing performed in the second learning phase.

[0062] Fig. 6a-d are graphs comparing the candidate thresholds to the predetermined thresholds for the low and high frequency band respectively.

[0063] Fig. 7a and 7b are graphs illustrating the resulting updated thresholds levels obtained for the low and high frequency band respectively using clustering.

[0064] Fig. 8 shows a method detecting passing-by events;

[0065] Fig. 9 shows graphically how passing-by events are detected;

[0066] Fig. 10 shows the signals associated with passing-by event.

[0067] Fig. 11 shows how the turning off of a lamp may be controlled when a room has been vacated;

[0068] Fig. 12 shows how passing-by events may delay the turning off of a lamp;

[0069] Fig. 13 shows how passing-by event detection may be used to modify the turning off of a lamp; and Fig. 14 is a flowchart to explain a lamp turn off strategy.

[0070] DETAILED DESCRIPTION

[0071] The present disclosure provides a method for calibrating a radar motion sensor device in which a temporal sequence of samples of a radar data signal is obtained, each sample indicating a first power level in a first frequency band of a radar signal detected by the radar motion sensor device. A pattern recognition algorithm is applied to the samples to distinguish between motion in the room and a passing-by event comprising motion past the room. The samples which correspond to a passing-by event and when the room is vacant are identified. These identified samples are used (i) to update a first threshold for identifying motion in the room and / or (ii) for identifying when a room has been vacated. The updated first thresholds are used to perform motion detection and / or it is identified when a room has been vacated, in a manner which distinguishes between motion in the room and a passing-by event.

[0072] Fig. 1 shows a top down view of an exemplary indoor environment that introduces an extra challenging environment for radar motion sensors. Two rooms la, lb (e.g. conference rooms) are located next to each other and delimited by a transparent glass wall or window 3. In the first room la there is a radar motion sensor 2 installed with the purpose of detecting motion in the first room la. For example, upon detection of motion in the first room la the lights are turned on in the first room la and / or the HVAC system associated with the first room la is activated. On the other hand, if no motion has been detected in the first room la for a predetermined period of time, the lights and / or HVAC system associated with the first room la is deactivated.

[0073] However, since the radar motion sensor 2 is sensitive and utilizes electromagnetic signals capable of penetrating the transparent glass wall or window 3 the radar motion sensor 2 may detect motion associated with a first person 5a walking by in a hallway 4 or motion associated with a second person 5b moving the second room lb on the other side of the transparent glass wall or window 3. Detection of these motions may increase the number of false positives wherein the lights and / or HVAC system of the first room la is erroneously activated because a motion in the second room lb or hallway 4 is registered.

[0074] The applicant has proposed a method for automatically calibrating a radar motion sensor 2 which will tune the sensitivity of the radar motion sensor 2 such that the false positive rate can be reduced. This invention adds analysis of passing-by events (i.e., a person walking past an entrance to a room, when the motion detection is for detecting motion in the room), to further reduce the false positive rate, by preventing that passing-by events are falsely detected as in-room motion.

[0075] With reference to the flowchart in Fig. 2 the method proposed by the applicant, but not yet published, is first described in detail, before the modification proposed in this disclosure. This modification is also represented in Fig. 2, in particular as step S3a.

[0076] At step SI a radar data signal is obtained. The radar data signal may be extracted from a “raw” radar signal (sometimes referred to as simply the radar signal) being the signal recorded directly by the radar sensor. The radar sensor may use any type of Doppler radar, such as a pulsed Doppler radar or an unmodulated continuous wave (CW) Doppler radar. In the latter example, a constant frequency carrier wave in the GHz range is emitted and its reflection received. Transmitted and reflected waves are mixed to form an intermediate frequency (IF) signal in the time domain which is outputted as the “raw” radar signal. In available radar motion sensors, the carrier frequency may be for example 5.8 GHz, 24 GHz or 60 GHz. Accordingly, the radar signal comprises a plurality of successive samples, with each sample representing the radar signal at a respective point in time.

[0077] With further reference to Fig. 3, an exemplary (IF) radar signal 6 is shown with the horizontal axis denoting time and the vertical denoting received signal power. Around time to the radar signal 6 is relatively stable and comprises mainly noise. At around time ti a movement occurs in the vicinity of the radar motion sensor which causes the power level to fluctuate rapidly. The rapid fluctuation continues until time t2, and after time t2 some fluctuations are still present, although less rapid. After time t3 the fluctuations are reduced even further.

[0078] The frequency content of the radar signal 6 may be extracted in one or more frequency bands forming a radar data signal having a temporal sequence of samples with each sample representing the spectral power in the one or more frequency bands. For example, the radar data signal may be transformed from time domain (as shown in Fig. 3) to a time-frequency domain (e.g. using an FFT) so that the frequency content of the radar signal 6 can be analyzed.

[0079] With further reference to Fig. 4 the spectral power, over time, in a first frequency band of the radar signal 6 is indicated with line 7 and the spectral power, over time, in a second frequency band of the radar signal 6 is indicated with line 8. In this example, the first frequency band is a low frequency band with frequencies 0 - 4 Hz and the second frequency band is a high frequency with frequencies 6 - 10 Hz. However, it is understood that these frequency bands are merely exemplary and while these frequency bands are suitable for radar motion sensors using 5.8 GHz carrier signals, other frequency bands may be used if other carrier frequencies are used.

[0080] The respective power levels 7, 8 in different frequency bands are directly extracted from the radar signal 6 and, in general, a signal indicating for each time sample the power level in one or more frequency bands will be referred to as a radar data signal. In some implementations, two frequency bands are used and each sample in the radar data signal will represent the power level 7, 8 in each frequency band respectively. Accordingly, each sample in the radar data signal may be represented with a vector associated with an individual timestamp, wherein each vector element representing the spectral power in a respective frequency band. In a completely analogous alternative definition, each radar data signal indicates the spectral power level of only one frequency band whereby in embodiments where two or more frequency bands are used there are two or more radar data signals with samples that represent only one power level. The temporal sequence of samples in each radar signal will in such cases be corresponding in time, meaning that sample number A in one radar data signal will have the same timestamp as sample number A in another radar data signal.

[0081] In the below, it will be assumed that one radar data signal with samples representing one or more power levels will be used.

[0082] In Fig. 4, the low frequency spectral power level 7 increases rapidly from time to to time ti together with the high frequency spectral power level represented with line 8. Between times t2 and t3, the low frequency spectral power 7 is consistently higher compared to line 8 indicating that between t2 and t3 the low frequency band contains more spectral power than the high frequency band.

[0083] It has been realized that this general pattern, i.e. a rapid increase in both high and low frequency spectral power followed by a comparatively high level of low frequency spectral power and comparatively low level of high frequency spectral power is typical for when people enter a room (e.g. a conference room or office) and sit down (at e.g. a conference table or desk). First, when people enter the room and approach the radar sensor there are many types of motion registered by the radar motions sensor. Then, when the people sit down there are mostly small motions (e.g. hand gestures or head moving) which are mainly associated with low frequency components in the radar signal 6. At step S2 a trigger sample is determined. The trigger sample is determined by identifying the first sample which is associated with a first or second power level 6, 7 that exceeds one or more of the predetermined thresholds TLF, THF.

[0084] In some implementations, the trigger sample is determined as the first sample which is associated with a power level in the first frequency band that exceeds a first predetermined threshold TLF. In the example shown in Fig. 4 it is seen that the first power level (i.e. the spectral power in the first frequency band represented with line 7) reaches the first predetermined threshold TLF at time ti. Accordingly, the sample at ti may be identified as the trigger sample.

[0085] In some implementations, the trigger sample is determined as the first sample which is associated with a first and a second power level 7, 8 which both exceed a respective first and second predetermined threshold TLF, THF. AS seen in Fig. 4 the second power level 8 exceeded the second predetermined threshold level THF some time prior to time ti and remained above the second predetermined threshold level THF until time ti and beyond. Accordingly, the sample at ti is again identified as the trigger sample although it is understood that with this stricter criteria of requiring both power levels 7, 8 to exceed their respective predetermined thresholds TLF, THF which sample that is identified as trigger sample will generally be different from the sample identified with the more relaxed criteria (i.e. using only one power level and threshold).

[0086] The first and second predetermined threshold TLF, THF may be the same or different.

[0087] When the trigger sample (in this example the sample at time ti) has been identified the method continues to step S3 and determines if a trailing event has occurred. The trailing event will provide an indication whether people whose movement caused the criteria for identification of the trigger sample to be fulfilled continues to move and remains in proximity to the radar motion sensor.

[0088] When determining whether a trailing event in a predetermined time duration following the has occurred, a set of samples in a trailing window following the trigger sample may be analyzed. The trailing window may comprise e.g. all samples in a predetermined time duration following the trigger sample that was identified at time ti. However, this definition of the trailing window is merely exemplary, and many types of trailing windows can be used. For example, the trailing window or predetermined time duration is at least 20 seconds long, at least 40 seconds long or at least 60 seconds long and starts immediately after the trigger sample or after predetermined waiting time (e.g. 2, 5 or 10 seconds) after the trigger sample. In the depicted embodiment, the trailing window comprises the first sample after the trigger sample (at time ti) and all samples until time t3. To determine if a trailing event has occurred it is determined whether (or at least to what extent) at least one of the power levels continues to exceed their respective predetermined threshold TLF, THF.

[0089] If the first and / or second power level exceeds the first and / or second predetermined threshold level for a predetermined percentage of the samples in the trailing context window a trailing event has been identified. Preferably, in the trailing window people are often expected to sit down and exhibit small movements whereby at least the low frequency power level with its respective predetermined threshold TLF is used (optionally in combination with the high frequency power level) to determine if the trailing event has occurred. The predetermined percentage may be between 50% and 95%, and preferably between 70% and 95% such as 95%, 90% or 80%. Preferably, the percentage is adjusted based on the length of the trailing context window with higher percentages being used for shorter trailing context windows and vice versa. For example, if a 30 second trailing context window is used the predetermined percentage may be 90% whereas if a 60 second trailing context window is used the predetermined percentage may be 80%.

[0090] If no trailing event has been identified the method goes back to step S2 and monitors the radar data signal until the next trigger sample can be identified. Preferably, the method resumes its analysis of the radar data signal at a sample which occurs later than the trigger sample at time ti, such as at sample which is 10 seconds later than ti or at least the length of one trailing window later than ti. In some implementations, the method resumes its analysis to find the next trigger sample after the trigger context window and trailing context window has passed. However, it is also envisaged that the method resumes its analysis of the radar data signal at the very next sample after ti.

[0091] If a trailing event has been identified the trigger sample is classified as a valid trigger sample. The method continues to step S4 and stores a record of the samples in a trigger context window surrounding the valid trigger sample. The trigger context window comprises the valid trigger sample (at ti) and a plurality of samples temporally earlier and / or temporally later than the valid trigger sample. For example, the trigger context window extends ± 1 second, ± 5 seconds, ± 10 seconds, or ± 20 seconds from the valid trigger sample at time ti. The trigger context window is not necessarily symmetrical and may for example contain a longer duration of samples after ti compared to before ti or vice versa. In one example implementation, the trigger context window includes ti as well as 10 seconds of samples temporally earlier than ti and 20 seconds of samples temporally later than ti. In Fig. 4, the trigger context window extends from to to t2.

[0092] The record of the trigger context window includes a record of the first and / or the second power level for each sample in the trigger context window.

[0093] Steps S2, S3 and S4 are repeated a plurality of times until a plurality of valid trigger samples have been detected and a database of records with trigger context windows has been collected. These trigger context windows indicate portions of the radar data signal that are followed by a trailing event and therefore represent power level sequences associated with potentially valid movements associated with people actually entering the room of the radar sensor. The fact that a valid trailing event has been registered for each record indicates that these trigger context windows likely indicate people entering the room of the radar motion sensor and not e.g. merely walking by in the hallway.

[0094] Once a sufficient number of (e.g. a predetermined number of) records with trigger context windows have been collected in the database the first training phase El is completed, and the method continues to step S5 and S6 of the second training phase E2.

[0095] Step S5 involves, for each trigger context window in the database, determining a candidate threshold TLF.U, THF.U that in general is greater than or equal to the static predetermined threshold TLF, THF used in step S2 to determine the trigger sample. The process of determining candidate threshold levels TLF.U, THF.U, will now be described with further reference to Fig. 5 where eight records, labeled record 1-8, have been collected during the first training phase El. The number of records is merely exemplary, and it is envisaged that as few as two records may be used but often more than ten or even hundreds of records will be available after the first training phase El.

[0096] Each record comprises at least the first power level of each sample in one trigger context window. Optionally, each record comprises the first and second power level of each sample of the same trigger context window. The records are obtained by a threshold updating module 9 which is configured to determine, for each record and frequency band, a candidate threshold TLF.U, THF.U. The threshold updating module 9 accomplishes this by finding a candidate threshold TLF.U, THF.U which is greater than or equal to the predetermined threshold TLF.U, THF.U but still allows a trigger sample to be identified in the trigger context window of the respective record.

[0097] With further reference to Fig. 6a and 6b an example of the updated threshold TLF, ui for the first frequency band with respect to the predetermined threshold TLF is shown. In Fig. 6a and 6b, the power level 7 of the first frequency band of record 1 is plotted with respect to time along the horizontal axis with an arbitrary power unit a.u. on the vertical axis. In Fig. 6a it is seen that the power level 7 of the first frequency band is above the first predetermined threshold TLF set to 45 a.u. at some time prior to tawhereby a trigger sample is identified.

[0098] The threshold updating module 9 finds a first candidate threshold TLF, ui (see Fig. 6b) that is higher (equal to about 50 a.u.) compared to the first predetermined threshold TLF wherein at least one sample of the first power level 7 is above the first candidate threshold TLF, ui. Accordingly, this candidate threshold TLF, ui can be used to replace the first threshold TLF while still enabling a trigger sample to be identified in the same trigger context window.

[0099] The threshold updating module 9 performs the same processing for the remaining records and determines one first candidate threshold TLF, ui, . . . ., TLF,U8 for each record 1, 2, . . ., 8 wherein each first candidate threshold is greater than the first predetermined threshold TLF.

[0100] An updated threshold TLFJ is then determined at step S6 based on one or more of the candidate thresholds TLF, ui, ... ., TLF, U8.

[0101] In some implementations, a plurality of candidate thresholds TLF, ui, ... ., TLF,U8 are available and provided to a clustering module 10 which performs clustering (K-mean clustering) into K groups wherein K > 1 wherein the updated threshold TLFJ separates the group with the highest center level (group centroid) from the other groups (if present). An example of clustering with K = 2 is shown in Fig. 7a where the first and second candidate thresholds TLF, U1, TLF, u2 have been assigned to different groups and the updated threshold TLF,F has been determined to separate the two groups.

[0102] The processing for the first frequency band may be repeated also for the second frequency band. With reference to Fig. 6c and 6d an example of a second updated threshold THF, ui for the second frequency band with respect to the predetermined second threshold THF is shown. In Fig. 6c and 6d, the power level 8 of the second frequency band of record 1 is plotted with respect to time along the horizontal axis and with respect to the arbitrary power unit a.u. on the vertical axis. In Fig. 6c it is seen that the power level 8 of the second frequency band is above the second predetermined threshold THF set to 45 a.u. at ta. The threshold updating module 9 finds a second candidate threshold THF, ui (see Fig. 6d) that is higher (equal to about 50 a.u.) compared to the second predetermined threshold THF wherein at least one sample of the second power level 8 is above the second candidate threshold THF. UI .

[0103] The threshold updating module 9 also processes the other records and determines one second candidate threshold THF, ui, ...., THF, u8 for each record 1, 2, . . . , 8 wherein each second candidate threshold THF, ui, ...., THF, u8 is greater than the second predetermined threshold THF.

[0104] An updated second threshold THFJ is then determined at step S6 based on one or more of the candidate thresholds THF, ui, ... ., THF, U8. In some implementations, a plurality of candidate thresholds THF, ui, ... ., THF, U8 are provided to the clustering module 10 which performs clustering (K-mean clustering) into K groups wherein K > 1 wherein the updated threshold THF, f so as to separate the group with the highest center level (group centroid) from the other groups. An example of clustering with K = 2 is shown in Fig. 7b where the seventh and eight second candidate thresholds THF, U7, THF, U8 have been assigned to different groups and the updated threshold THF, f has been determined to separate the two groups.

[0105] Accordingly, in some implementations the processing at step S5 and S6 is separate for the two frequency bands. However it is also envisaged that the method is performed with only one frequency band, such as only for the first frequency band or only for the second frequency band.

[0106] In some implementations, the processing at step S5 (performed by the threshold updating module 9) for determining first and second candidate thresholds TLF, ui, THF, ui is linked between both frequency bands such that at least one sample which exceeds both the first and second candidate threshold TLF, ui, THF, ui is present in the trigger context window of each record.

[0107] As an example, the threshold updating module 9 determines the first candidate thresholds TLF, U1, THF, U1 using an iterative process. The iterative process is initiated with the threshold updating module 9 setting a respective attempt threshold level to equal the first and second predetermined threshold level. At iteration 1 the threshold updating module 9 then incrementally increases the first attempt threshold level and checks if a trigger sample having power levels above both attempt thresholds levels can be detected in the trigger context window. If this is the case, the threshold updating module 9 continues to iteration 2 and incrementally increases the second attempt threshold level and checks if a trigger sample having power levels above both attempt thresholds level can be detected in the trigger context window. If this is the case, the threshold updating module 9 continues to iteration 3 and again incrementally increases the first attempt threshold and so on. The process repeated by alternating incremental increase of the first and second attempt thresholds until after N iterations it is found that no sample in the trigger context window has a first and second power level above the first and second attempt threshold level. The threshold updating module 9 then assigns the attempt threshold levels of iteration N - 1 to the candidate threshold levels TLF. UI, THF. UI respectively.

[0108] Accordingly, using this iterative process candidate threshold levels TLF, U1, THF, ui can be maximized while still allowing at least one trigger sample to be identified in the trigger context window even with the stricter criteria for trigger sample identification requiring that both power levels should be above the respective candidate threshold TLF, ui, THF, ui.

[0109] It is further envisaged that as an alternative to using clustering for determining the updated threshold levels TLF, F, THF, f respectively different operations can be used. For example, the updated threshold levels TLF, F, THF, fare set as the average or median of the respective candidate threshold levels TLF, U, THF, U.

[0110] After step S6 has been completed a first updated threshold (and optionally also a second updated threshold) is available and the second learning phase E2 has been completed.

[0111] The method then goes to the use phase E3 and performs motion detection at step S7 for a second, future, radar data signal having samples using the first updated threshold and optionally also the second updated threshold. A motion is detected if the first, and optionally second, updated threshold is exceeded in a single sample by the first, and optionally the second, power level. Preferably, during phase E3, the processing related to determining a trailing event is skipped which makes the implantation less computationally intensive. Optionally, the radar motion sensor device can be periodically or manually (e.g. when the radar motion sensor is moved to a new room) reset whereby learning phase El and E2 are repeated with the default predetermined threshold values TLF, THF.

[0112] The radar motion detector may still be used during the first and second learning phase El, E2 e.g. while records are collected in learning phase El. For example, the first and second predetermined thresholds are used to identify a trigger sample which is used as an indication that a movement has been detected. While the first and second predetermined threshold may exhibit a relatively high false positive rate this is only a temporary solution until enough records have been collected and the updated threshold level(s) have been determined. Turning back to the exemplary room layout of Fig. 1 it is understood how the calibration process for determining updated thresholds described above will improve motion detection accuracy.

[0113] Firstly, since only trigger context windows for which a trailing event is detected after the trigger sample it is likely that no, or very few, trigger context windows associated with a person 5a walking by in the hallway are stored among the records. This is because the duration of the movement of the person 5a walking by is likely too short to fulfill the requirements of the trailing event. This means that the trigger context windows associated with e.g. a person 5a walking by in the hallway will not be stored as a record in the first learning phase El and therefore also ignored in the second learning phase E2 so as to not prevent setting higher updated threshold levels.

[0114] Secondly, if clustering is performed with at least two groups, it is possible to identify updated thresholds which separate motion by a person 5b in an adjacent room lb from motion by a person in the room la where the radar motion sensor 2 is located. The radar motion sensor 2 may detect both a trigger sample and a trailing event when a person 5b enters the adjacent room lb, however by finding candidate thresholds and performing clustering of these candidate thresholds into at least two groups it will in many applications be possible to distinguish between a person 5b entering the adjacent second room lb and a person entering the first room la where the radar motion sensor 2 is located. The clustering may optionally be performed so as to form more than two groups. For example, some conference rooms will have a transparent wall or window 3 facing two or more other conference rooms whereby clustering into three or more groups will be beneficial for accurate detection of motion in first room where the radar motion sensor 2 is located.

[0115] One limitation of the approach described above is that it does not take the motions of passing-by into consideration. Motions that transpire over a very short time (e.g. due to someone walking by in the hall) are not considered when determining the updated threshold or thresholds. This limitation may lead to false triggers in certain circumstances.

[0116] For example, with reference to Fig. 1, a motion detection algorithm may well distinguish walking into room la from walking into room lb, may fail to distinguish walking into room la from walking along the corridor, and thus generate false triggers.

[0117] This invention proposes an improved self-learning method by further capturing passing-by information, in order to update the motion detection algorithm (e.g., to update relevant thresholds). As described above, the radar sensor starts a learning process after installation. During the learning process, it keeps recording a certain duration of raw or processed radar signal into a buffer.

[0118] In accordance with the modification to the above method as provided by this disclosure, a pattern recognition algorithm is additionally used to detect passing-by events from the radar sensor data. The pattern recognition algorithm is pre-developed to detect a passing-by event, i.e., a person walks continuously outside the room (i.e., the target room where the radar is installed) in parallel with one wall (e.g., the wall near the corridor) of the target room.

[0119] This is shown as step S3a in Fig. 2. It takes place after step the identification of trigger samples in step S2 and results in additional recording of samples of interest in step S4, which are then used when setting candidate thresholds in step S5 and / or adaptation of thresholds in step S6.

[0120] In addition to detecting a passing-by event, the algorithm further determines as part of step S3a if the room is vacant when the passing-by happens. For example, with reference to Fig. 1, room la may have been vacant for a while and a person is now walking along the corridor from left to right in parallel with the wall of room la in which the door is located (which may be a glass wall so that there is detection of movement in the corridor), crossing line LI, line L2 and line L3 in sequence. The pattern recognition algorithm will detect a passing-by event and further determine that room la is vacant when the passing-by event happens.

[0121] If the person stops walking or changes the walking direction in front of line L2, the pattern recognition algorithm will not detect a passing-by event. If room la is occupied while a person is walking along the corridor passing-by room la, a passing-by event will be detected but will not be saved in the database (because it is of no use for setting thresholds).

[0122] As the learning continues, the radar sensor will capture (in the database) a number of records of raw or processed radar signals corresponding to a person passing-by room la in addition to the captured samples described above based on the trailing characteristics.

[0123] As described above, the stored records, which in accordance with this modification to the method described above include passing-by events, are then used to update the motion detection algorithm so that it can distinguish walking into room la from passing-by room la, as well as distinguishing between walking in room la and walking in room lb as described above.

[0124] In the same way as explained above, this is achieved by continuously increasing the thresholds for detecting a valid motion (i.e., a person walking into the target room) until no passing-by event is detected as a valid motion.

[0125] Fig. 8 illustrates in more detail the detection of valid passing-by events, shown as step S3a in Fig. 2.

[0126] The raw radar signal is received by a pattern recognition algorithm 20 and it is buffer in buffer 22 (as mentioned above). The pattern recognition algorithm 20 keeps monitoring if there is a person passing-by the target room in step S10. During a passing-by event, the distance between the radar and the walking person keeps decreasing to a minimum then keeps increasing.

[0127] For example, with reference to Fig. 1, a person walking along the corridor from left to right will firstly cross line LI and enters the signal coverage of the radar sensor. The distance between the radar and the person will keep decreasing until the person is at line L2 where the distance is at its minimal. After crossing line L2, the distance starts increasing. Accordingly, the power of the radar signal reflecting the motion of the walking person firstly increases to a maximum then decreases, and the increasing and the decreasing are roughly symmetrical. This is the foundation of the pattern recognition algorithm to realize passing-by detection.

[0128] Fig. 9 shows the power signal derived from a Fast-Fourier transformation, FFT, in a lower frequency range as plot 30 (e.g., 5-25Hz) and in a higher frequency range as plot 32 (e.g., 26-50Hz). For each sample of the power signal, a moving average is calculated which is the average value of a number of samples (such as 10) preceding the sample, the sample itself, and a number of samples (such as 10) following the sample. In this way, two smoothed power signals are created: a smoothed low frequency power signal shown as plot 34 and a smoothed high frequency power signal shown as plot 36.

[0129] The algorithm keeps checking if the smoothed low frequency power signal rises to and crosses a pre-defined threshold or a second threshold (e.g., TH_P) then falls to and crosses the pre-defined threshold or the second threshold within a pre-defined duration (e.g., 3 seconds). The pre-defined duration may be a maximum duration or it may be a range of permitted durations (e.g., 1.6 to 3 seconds).

[0130] If such a period of the smoothed low frequency power signal is identified, e.g., the signal between T1 and T2 (T2 - T1 < 3s), a crossing window 40 is identified. The algorithm further determines the moment (e.g., T SLF) when the smoothed low frequency power signal 34 is at its maximum within that crossing window 40 (i.e., from T1 to T2). If T SLF is roughly at the middle of the crossing window (indicating a symmetrical motion towards and away from the radar sensor), the algorithm further determines the moment (e.g., T SHF) when the smoothed high frequency power signal 36 is at its maximum within the crossing window 40. If the time delay interval between T SLF and T SHF is small enough (e.g., <0.5 second or <1 second), a passing-by is detected.

[0131] Fig. 10 shows the signals associated with passing-by event. Fig. 10A shows a person 5 walking past a room with a ceiling radar sensor 2.

[0132] Fig. 10B shows the frequency -time characteristics of the FFT of the raw sensor data, for the passing-by event. The frequency-time domain profile is symmetric with a trough shape. This symmetric profile corresponds to a uniform speed of motion past the radar sensor. When the person is far from the sensor 2 and is approaching the sensor, there is a large component of their movement direction which is towards the sensor, hence a high Doppler frequency. When the person reaches the shortest distance from the sensor (line L2), there is no component of their movement direction that is towards the sensor (their movement is then tangential), hence a zero Doppler frequency.

[0133] Fig. 10C shows the power-time characteristics for the high frequency and low frequency components and Fig. 10D shows the smoothed power-time characteristics. The smoothing is performed by obtaining moving average results over time.

[0134] When the peaks of the (smoothed) power signals are near the center of the crossing window, this indicates symmetry in the power-time characteristic for each of the different frequencies. When all of the power-time characteristics are determined to be symmetrical, this is indicative of the symmetric frequency-time domain profile of the passing-by event.

[0135] After a passing-by event is detected (e.g., between T1 and T2), the pattern recognition algorithm further checks the power signals of a period 50 shortly preceding the passing-by event, e.g., a period of 30 seconds preceding the moment T1 - 2s), i.e., between TO and (Tl-2s).

[0136] If both power signals 30, 32 remain below a pre-defined threshold or a third threshold (e.g., TH_V) for the whole period, the target room is determined as vacant when the passing-by happens. Then, the raw radar signal and / or the power signals corresponding to the passing-by (i.e., between T1 and T2) are retrieved from the buffer, and saved in the database 24 in step S4. Fig. 8 shows the vacancy detection explained above as step SI 1. Only when samples indicate a passing -by event (step S10) and vacancy (step SI 1) are the samples stored in step S4 in the database 24, since those are the samples that are used to test and adapt the motion detection thresholds.

[0137] As described above, after enough records of passing-by events are captured, the learning process is completed and the thresholds for the motion detection algorithm can be updated.

[0138] In one example, the thresholds for detecting a valid motion (e.g., a person walking into the target room) may be first determined using the method described above (without detecting passing-by events), and the thresholds may then be updated further by continuously increasing them until no passing-by event is detected as a valid motion. Thus, the updating of the motion detection thresholds may comprise a two-phase updating process.

[0139] The thresholds could however be updated in the opposite order. Alternatively, an update process may be provided in which all of the sensor waveform analysis data is processed together to derive the suitable threshold levels.

[0140] It is noted that the strength of the motion signal resulting from a passing-by event compared to motion into a neighboring room will depend on the room layout. For example, for a room which is narrow (i.e., extends along a short corridor distance) and deep (i.e. shares a long wall with a neighboring room), the motion signal generated by a passing- by event may be weaker than the motion signal generated by a person entering the neighboring room. In such a case, there may be no false triggers after the thresholds have been set based on clustering between entering the target room and the neighboring room. Thus, no further updates to the thresholds may be needed. However, in other cases, further updates to the thresholds may be needed to avoid false triggers. For example, for a room which is wide (i.e., extends along a long corridor distance) and shallow (i.e. shares a short wall with a neighboring room), the motion signal generated by a passing-by event may be stronger than the motion signal generated by a person entering the neighboring room. This is also particularly the case when the room is in the middle of a series of rooms rather than at the end of a corridor.

[0141] High frequency and low frequency power thresholds may be adapted, such that all room entry events exceed the power threshold and all passing-by events are below the power threshold.

[0142] A further refinement of the invention relates to the method used for turning a lamp off when a room has been detected as vacated. A known approach involves the radar sensor constantly waiting for a period of silent background to switch off the lamp in a room (wherein "silent" is used to denote lack of movement).

[0143] Fig. 11 shows in the top plot a raw sensor signal and in the second plot the FFT result. The third plot shows a "silence measure". The silence measure increases during periods of no detected movement and flattens or decreases, during periods of movement. As explained more fully below, the silence measure is a sum of binary silence indices allocated to sequential (e.g., 15 second) time slots.

[0144] More generally, there is a silence measure which increases or decreases when there is no motion detection and decreases or increases, respectively, when there is motion detection. The silence measure can then be compared with a threshold or a fourth threshold to determine when there has been sufficiently low motion that the room can be determined to be vacated.

[0145] In the example of Fig. 11, from time tl there is no detected motion, so the silence measure ramps up over a period of 180 seconds. A continuous silence detection of 180 seconds will result in a silence measure which reaches a threshold to turn the lamp off.

[0146] The bottom plot shows the lamp status. It turns off when the silence measure reaches its threshold.

[0147] Fig. 12 shows the same plots as Fig. 11 for a small meeting room in which there are people passing by along a corridor. It shows that when the meeting has finished at time tl, even though it is totally empty inside the room, passing-by events outside the room may generate motion signals which prevent the lamp from being turned off in the desired 180s. In this example, the lamp switch off is delayed to 676 seconds, caused by detected passing-by events which cause the silence measure to increase more gradually.

[0148] To address this problem, the passing-by event detection described above may be applied to the silence measure calculation.

[0149] Once a passing-by event is detected, it is not used as a negative factor for the silence measure calculation, and thus does not result in the delay shown in Fig. 12. In this way, the silence measure will increase even when passing-by events occur outside the room, so that as a result the lamp will turn off in a more timely manner.

[0150] Fig. 13 shows the same raw sensor data and shows that passing-by events 30 are detected. By taking account of these, the turn off time is returned to 180 seconds. The original silence measure curve (of Fig. 12) is shown in Fig. 13 together with the silence measure curve 40 resulting from the processing outlined above, and explained more fully below. Fig. 14 is a flowchart to explain the lamp off strategy. New radar data is collected in step S20. The radar data is for example collected over periods of 15 seconds as also explained above.

[0151] A silent index bit SI is updated in step S21. For this purpose, the power level is compared to a preset background power threshold P_bg. If most of the time, e.g., 90%, the collected power is less than or equal to P_bg, a current silent index bit is set to 1, otherwise, the current silent index bit is set to 0. It indicates silence or movement in that individual time period (e.g., 15 seconds).

[0152] A passing-by event outside the room triggers a higher power than the background power P_bg and for a period of more than 10% of the time period (i.e. 1.5s in this example). Such a passing-by event may thus generate a silent index bit of 0 which indicates movement.

[0153] A FIFO (or double ended queue) is used to hold all the silent index bits during a hold on period, e.g., 3 minutes (180 seconds), as a result, there will be 12 bits in the FIFO.

[0154] It is determined if the lamp is on in step S22.

[0155] If the lamp is off, there is no need for silence analysis, and the method returns to the start.

[0156] If lamp is on, the 12-bit sequence is reviewed, and a trend value dS is obtained. This trend value is for example based on a difference between (i) the maximum, or summation, of the latest certain number of bits, and (ii) the minimum, or summation, of the preceding certain number of bits. The certain number of bits is for example six. Thus, it represents a change in movement characteristics over time. Any suitable measure of change over time may be used.

[0157] In step S23, the trend value dS is compared with a threshold th_O. If the threshold is exceeded, the collected motion signals are re-checked against the passing-by event characteristic. The threshold th_O is used to indicate a certain confidence level of the unoccupied status of the meeting room. Thus, if the value dS is larger than the threshold th_O this signifies that the room has potentially been vacated. If the room has not potentially been vacated, movement analysis (to check for passing-by events) is not needed so that processing power can be saved, and the method advances to step S28 described below.

[0158] For example, Fig. 13 shows an example of how an unoccupied status is assessed in this way. When the maximum value for a second half H2 of a review period (i.e., a latest 6 bits) reaches 6 (at point B), and the minimum of the first half Hl of the review period (i.e., the preceding 6 bits) was zero, then the trend value dS exceeds the preset threshold th_O, e.g., 5. This confirms potential vacancy.

[0159] At this point, the algorithm will review the stored silent index bit, and calibrate the silent index bit according to the identified passing-by event(s) during the past hold on period. As a result of the calibration, the silence measure curve will evolve from the previous plot to plot 40. For example, point A has been re-calibrated so that the divergence from plot 40 is corrected.

[0160] In particular, in step S24 it is checked if the current silent index bit is zero. A zero silent index bit indicates movement. If there is detected movement the movement is analyzed to determine if it was caused by a passing-by event. If the silent index bit is one, the movement analysis is not needed, so the method advances to step S28 described below.

[0161] The passing-by event detection explained above is used in step S25 to identify whether a passing-by event happened in the last 15 seconds. In step S26 it is determined if ONLY a passing-by event was detected, i.e., there was motion detected but it was not motion in the room but was a passing-by event motion.

[0162] If so, in step S27 the silent index bit is flipped from 0 (busy) to 1 (silent). This means a passing-by event outside the room will not create a delay for the lamp off decisions. Thus, the method involves swapping a first value (indicating motion detection) for a second value (indicating no motion detection), when motion of a passing-by event is detected.

[0163] Finally, a sum of all the silent index bits in the FIFO is measured in step S28. If the sum of all the silent index bits in the FIFO exceeds a threshold th_l then the lamp is turned off in step S29.

[0164] The th_l is for example 12. It indicates that the space has been silent enough during the past hold on period, such that a high probability of space unoccupancy is confirmed so that the lamp can be turned off. Otherwise, the device will wait for more radar data of another 15 seconds to update the silent bit for the next decision.

[0165] Thus, motion detection means the silent index bit will be 0 (busy) and this would normally delay the lamp turning off. However, if the silent index bit was 0 because of a passing-by event, the silent index bit is converted to a 1 so that the motion detection resulting from the passing-by event may be considered to be overruled, and thereby does not delay the lamp turn off function.

[0166] The example above makes use of two frequency bands. There may instead be three or more frequency bands that are analyzed. The person skilled in the art realizes that the present invention by no means is limited to the preferred embodiments described above. On the contrary, many modifications and variations are possible within the scope of the appended claims. For example, it is possible to use one the spectral power in one frequency band or two frequency bands. Additionally, it is envisaged that the spectral power in more than two frequency bands may be used to perform the calibration.

Claims

CLAIMS:

1. A method for calibrating a radar motion sensor device (2) located in a room, comprising: obtaining (SI) a temporal sequence of samples of a radar data signal from the radar motion sensor device (2), each sample indicating a first power level (7) in a first frequency band of the radar signal (6) detected by the radar motion sensor device (2); applying (S10) a pattern recognition algorithm to the samples to distinguish between motion in the room and a passing-by event comprising motion past the room; identifying (SI 1) those samples which correspond to a passing-by event when the room is vacant; using (S5, S6) the identified samples to update a first threshold for identifying motion in the room and / or for identifying when a room has been vacated; and using (S7) the updated first threshold to perform motion detection and / or identifying when a room has been vacated with the radar motion sensor device (2), in a manner which distinguishes between motion in the room and a passing-by event.

2. The method of claim 1, wherein distinguishing a passing-by event comprises detecting a decreasing distance to the radar motion sensor device followed by an increasing distance to the radar motion sensor device.

3. The method of claim 2, wherein distinguishing a passing-by event comprises detecting a symmetrically decreasing distance and increasing distance regarding to time.

4. The method of any one of claims 1 to 3, wherein distinguishing a passing-by event comprises: converting the temporal sequence to the frequency domain; calculating a power in a first frequency range; deriving a first smoothed power signal for the first frequency range; andchecking if the first smoothed power signal rises to cross a second threshold and then falls to cross the second threshold within a predefined duration, thereby defining a crossing window.

5. The method of claim 4, wherein distinguishing a passing-by event further comprises: calculating a power in a second frequency range, higher than the first frequency range; deriving a second smoothed power signal for the second frequency range; comparing a time delay, between a maximum of the first smoothed power signal and a maximum of the second smoothed power signal within the crossing window, with a time delay threshold, wherein a passing-by event is detected when the time delay is below the time delay threshold.

6. The method of claim 5, wherein distinguishing a passing-by event comprises detecting the room is vacant before the passing-by event by determining that the first and second smoothed power signals are below a third threshold.

7. The method of any one of claims 1 to 6 comprising using the identified samples for identifying when a room has been vacated by:(S28) deriving a silence measure which increases or decreases when there is no motion detection and decreases or increases, respectively, when there is motion detection; and(S28) comparing the silence measure with a fourth threshold to determine when there has been sufficiently low motion that the room can be determined to be vacated, wherein the method comprises (S27) discounting motion detection when it is determined to be detection of a passing-by event.

8. The method of claim 7, wherein the silence measure is a sum of binary silence indices (SI) each for an associated time window, wherein each silence index has a first value when there is motion detection and a second value when there is no motion detection, wherein the method comprises swapping the first value for the second value before performing the sum, when motion of a passing-by event is detected.

9. The method of any one of claims 1 to 8, wherein the method further comprises updating the threshold for identifying motion in the room by: identifying (S3) a plurality of valid trigger samples by comparing the first power level of each sample to a first predetermined threshold (TLF), wherein each valid trigger sample is identified as a sample having a first power level (7) that exceeds the first predetermined threshold (TLF) and is followed by a trailing event in a predetermined time duration after the sample; analyzing samples of a trigger context window associated with each of the plurality of valid trigger samples to determine a first updated threshold (TLF, I).

10. The method according to claim 9, wherein analyzing the samples of a trigger context window associated with each of the plurality of valid trigger samples comprises: determining a first candidate threshold (TLF. U) for each valid trigger sample based on the samples of the trigger context window; and clustering the first candidate thresholds (TLF. U) into at least one group, each group being associated with a group center level; wherein the first updated threshold (TLF, I) is set to a value that separates the group associated with a highest group center level.

11. The method according to any of claims 9 to 10, further comprising: identifying (S2) a plurality of trigger samples by comparing the first power level of each sample to the first predetermined threshold (TLF), wherein each trigger sample is identified as a sample having a first power level (7) that exceeds the first predetermined threshold (TLF); for each identified trigger sample: defining a trailing context window associated with the trigger sample comprising a predetermined number of samples temporally later than the trigger sample; determining (S3), for each sample in the trailing context window associated with the trigger sample whether the first power level (7) of the sample exceeds a predetermined trailing threshold; classifying the trigger sample as a valid trigger sample when the percentage of samples in the trailing context window exceed a percentage threshold.

12. The method according to any of claims 9 to 11, wherein using the first updated threshold (TLF. I) to perform motion detection comprises: obtaining a sample of a second radar data signal, the sample indicating the first power level (7); comparing the first power level (7) to the first updated threshold (TLF, F); outputting a signal indicating that motion has been detected in response to the first power level (7) of the sample exceeding the first updated threshold (T F,f).

13. The method according to any of claims 9 to 12, further comprising: identifying (S2) a plurality of trigger samples by comparing a first power level of each sample in a first frequency range and a second power level of each sample in a second frequency range to a first predetermined threshold (TLF) and second predetermined threshold (THF) respectively, wherein each trigger sample is identified as a sample having a first power level and second power level that exceeds both the first predetermined threshold (TLF) and the second predetermined threshold (THF).

14. A computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of the method according to any of the preceding claims.

15. A controller for a radar motion sensor (2), the controller comprising a processor and a memory, wherein the processor is configured to perform the method according to any of claims 1-13.

Images