METHOD FOR SEAT OCCUPANCY DETECTION

Abstract

A method for occupancy detection for at least one vehicle seat, using at least one transmit antenna and a plurality of receive antennas, includes: emitting a detection signal with each transmit antenna onto at least one vehicle seat, which detection signal is a frequency-modulated continuous-wave radar signal, and receiving with each receive antenna a reflected signal; recording sample data representing the reflected signal, the sample data having M channels, with M=N1.Math.N2, where N1 is the number of transmit antennas and N2 is the number of receive antennas; for each channel, removing a component from the sample data that corresponds to a reflection from a static object; and applying a frequency estimation method to the sample data to at least implicitly determine at least one angle of arrival θ.sub.i corresponding to a position of an occupant on a vehicle seat.

Claims

1. A method for occupancy detection for at least one vehicle seat, using at least one transmit antenna and a plurality of receive antennas, the method comprising: emitting a detection signal with each transmit antenna onto at least one vehicle seat, which detection signal is a frequency-modulated continuous-wave radar signal, and receiving with each receive antenna a reflected signal; recording sample data representing the reflected signal, the sample data having M channels, with M=N1.Math.N2, where N1 is the number of transmit antennas and N2 is the number of receive antennas; for each channel, removing a component from the sample data that corresponds to a reflection from a static object; and applying a frequency estimation method to the sample data to at least implicitly determine at least one angle of arrival θ.sub.i corresponding to a position of an occupant on a vehicle seat. wherein the frequency estimation method is a multiple signal classification (MUSIC) method and comprises: calculating a sample covariance matrix {circumflex over (R)} calculating eigenvalues and eigenvectors of the sample covariance matrix {circumflex over (R)}; sorting the eigenvalues in descending order and, with D being a number of targets, selecting M−D smallest eigenvalues and corresponding eigenvectors to determine a noise subspace G; and calculating roots z.sub.i of a root-MUSIC polynomial J(z) with
J(z)=z.sup.M-1p.sup.T(z.sup.−1)GG.sup.Hp(z) wherein $p\left(z\right)=\left[\begin{array}{c}{z}^0\\ .Math.\\ z^{M-1}\end{array}\right]$ and each root z.sub.i corresponds to an angle of arrival θ.sub.i.

2. The method according to claim 1, wherein the method uses only D roots z.sub.i located inside a unit circle in the complex plane and closest to the unit circle.

3. The method according to claim 1, wherein for each vehicle seat, an angle interval is defined and wherein the method uses only roots z.sub.i corresponding to an angle of arrival θ.sub.i within an angle interval.

4. The method according to claim 1, wherein an inner circle having a radius of less than 1 is defined and wherein the method uses only roots z.sub.i outside the inner circle and inside the unit circle.

5. The method according to claim 1, wherein 1 to 5 samples are used for calculating the covariance matrix {circumflex over (R)}.

6. The method according to claim 1, wherein for each seat, an associated area in the complex plane is defined and for each frame, which corresponds to a plurality of modulation periods of the detection signal, a counter for this associated area is increased if at least one root z.sub.i is located in this associated area and decreased if there is no root z.sub.i in this associated area and the seat is identified as occupied if the counter exceeds a predefined threshold.

7. The method according to claim 1, wherein before applying the frequency estimation method, a range gating is performed for each channel by: transforming the sample data into a range representation; and only considering a portion of the sample data corresponding to a predefined range interval, which includes a potential position of an occupant.

8. The method according to claim 1, wherein a plurality of transmit antennas is used.

9. The method according to claim 1, wherein transmit signals from different transmit antennas are separated by time division multiplexing.

10. The method according to claim 1, wherein the receiving antennas are arranged as a uniform linear array.

11. The method according to claim 1, wherein the frequency estimation method is a Capon method, and comprises: calculating the sample covariance matrix {circumflex over (R)} with $\hat{R}=\frac{1}{K}\underset{k=0}{\overset{K-1}{.Math.}}x\left[k\right]x^H\left[k\right]$ wherein x[k] is an M-dimensional output vector representing the sample data in each channel in range representation and K is the number of data samples considered; calculating an inverse sample covariance matrix {circumflex over (R)}.sup.−1; generating a steering vector a(θ) for each of a plurality of scanning angles θ; and calculating a Capon power spectrum P.sub.Cap(θ) with $P_{\mathrm{Cap}}\left(\theta \right)=\frac{1}{a^H\left(\theta \right){\hat{R}}^{-1}a\left(\theta \right)}$

12. The method according to claim 1, wherein the method further comprises, after determining the noise subspace G: generating a steering vector a(θ) for each of a plurality of scanning angles θ; and calculating a MUSIC power spectrum P.sub.MUSIC(θ) with $P_{\mathrm{MUSIC}}\left(\theta \right)=\frac{a^H\left(\theta \right)a\left(\theta \right)}{a^H\left(\theta \right){\mathrm{GG}}^{H}a\left(\theta \right)}.$

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Further details and advantages of the present invention will be apparent from the following detailed description of not limiting embodiments with reference to the attached drawing, wherein:

[0040] FIG. 1 is a schematic front view of a vehicle interior with a system for performing embodiments of the inventive method for occupancy detection;

[0041] FIG. 2 is a schematic representation of two possible antenna arrays for the system from FIG. 1;

[0042] FIG. 3 shows a time evolution of the frequency of a detection signal from the antenna array from FIG. 2;

[0043] FIG. 4 illustrates the construction of a radar data cube;

[0044] FIG. 5 is a first comparative diagram of power spectra derived from different methods according to embodiments of the invention;

[0045] FIG. 6 is a second comparative diagram of power spectra derived from different methods according to embodiments of the invention;

[0046] FIG. 7 is a first diagram showing positions of roots of a MUSIC polynomial in the complex plane;

[0047] FIG. 8 is a first diagram showing associated areas in the complex plane;

[0048] FIG. 9 is a second diagram showing associated areas in the complex plane;

[0049] FIG. 10 is a diagram corresponding to FIG. 9 which also shows positions of roots of a MUSIC polynomial;

[0050] FIG. 11 is diagram showing a power spectrum as a function of range;

[0051] FIG. 12 is a diagram showing a power spectrum derived from a method according to embodiments of the invention;

[0052] FIG. 13 is a never diagram showing associated areas and positions of roots of a MUSIC polynomial in the complex plane;

[0053] FIG. 14 is a third comparative diagram of power spectra derived from different methods according to embodiments of the invention;

[0054] FIG. 15 is a flowchart illustrating a first embodiment of a method according to an embodiment of the invention;

[0055] FIG. 16 is a flowchart illustrating a second embodiment of a method according to an embodiment of the invention; and

[0056] FIG. 17 is a flowchart illustrating a third embodiment of a method according to an embodiment of the invention.

DETAILED DESCRIPTION

[0057] FIG. 1 schematically shows the interior of the vehicle 30 with a rear bench comprising a left seat 31, a middle seat 32 and a right seat 33. A system 1 for occupancy detection is installed in a ceiling above the rear seat region of the vehicle 30. It can e.g. be mounted behind the headliner in order to illuminate all rear seats 31, 32, 33 in the best possible way. The system 1 is adapted for occupancy detection, which may also include detection of unattended children that are left behind in the vehicle 30.

[0058] FIG. 2 schematically shows two versions of an antenna array that can be employed in the system 1. In each case, a plurality of receive antennas Rx are arranged as a uniform linear array (ULA) with a distance d between two receive antennas Rx. According to a first option, shown under letter a), the antenna array and is a SIMO system with a single transmit antenna Tx and eight receive antennas Rx. According to a second option, shown under letter b), the antenna array is a MIMO system with two transmit antennas Tx separated by a distance 4d and four receive antennas Rx. As illustrated in FIG. 2, in each case, eight different relative phase shifts from 0 to 7 co occur between the transmit antenna(s) Tx and the receive antennas Rx, with ω=4d.Math.sin θ. Each combination of a receive antenna Rx and a transmit antenna Tx corresponds to one (virtual) channel. The system 1 is configured to transmit an FMCW radar signal, which is herein referred to as a detection signal. If the MIMO configuration is used, it is important that the receive antennas Rx can separate reflected signals corresponding to the different transmit antennas Tx. In this example, the transmit signals of different transmit antennas Tx are separated by Time Division Multiplexing (TDM), which is illustrated in FIG. 3. Each of the transmit antennas Tx generates a series of alternating chirps, i.e. when one transmit antenna Tx generates a chirp, the other transmit antenna Tx is inactive. A pair of chirps forms a block of the transmit signal, and a frame comprises a plurality of blocks (e.g. 16 blocks, i.e. 16 chirps from each transmit antenna). During operation, a series of frames are emitted. In this example, the operating frequency is from 79 GHz to 109 GHz, corresponding to a base wavelength λ of 3.8 mm. The frequency slope of each chirp is 70 MHz/μs. The period of a frame is 62.5 ms, corresponding to 16 frames/sec. The distance d is 1.9 mm (i.e. λ/2).

[0059] The detection signal is at least partially reflected by the interior of the vehicle 30 and, apart from the vehicle seats 31, 32, 33, it may be reflected in by an occupant 20 or an in animate object like a backpack 21 disposed on one of the vehicle seats 31, 32, 33. At least a portion of the reflected signal is received by the receive antennas Rx. This is recorded as sample data, the structure of which may be illustrated in form of a radar data cube and as shown in FIG. 4. The radar data cube intuitively illustrates the processing data of the radar. It can be described as a three-dimensional arrangement of sample bins. One dimension of the cube corresponds to a number M of channels, in this case M=8. A second dimension corresponds to the number K of data samples (or data snapshots) recorded during one block of the signal, in this case K=256. The third dimension corresponds to total number of blocks of the recording. One could also say that the second dimension corresponds to fast time, while the third dimension corresponds to slow time. During the transmission of two chirps, one column of the cube is filled for all M channels. Over the duration of one frame, the remaining columns are filled. It is possible, though, that the recording comprises a plurality of frames, each of which comprises a certain number of blocks.

[0060] Performing an FFT over all channels would result in an angle information of the scene (Angle FFT). This is normally done according to prior art, but yields unsatisfactory angular resolution. Performing an FFT over all columns (corresponding to slow time) would yield velocity information (Doppler FFT). Performing an FFT over a single column yields range information (Range FFT).

[0061] Embodiments of the inventive method will now be explained with reference to the flowcharts in FIGS. 15 to 17. It should be noted that the sequence of the steps in each flowchart could be changed or that some steps that are shown to happen sequentially could be performed simultaneously.

[0062] According to a first embodiment, which is explained with reference to the flowchart in FIG. 15, a Capon method (also known as the Minimum Variance Distortionless Response (MVDR) algorithm) is used. In a first step, at 100, the detection signal is emitted and the reflected signal is received. At 105, the sample data is recorded. Then, at 110, a portion of the sample data corresponding to reflections by static objects is removed. This may be done by taking the average of a specific sample bin (e.g. the first sample bin in each column) over all columns and subtracting this average from every sample bin. At 115, a range FFT is performed to transfer the sample data into range representation.

[0063] At 120, a range gating is performed, which means that all sample data not corresponding to a certain range interval are discarded. The reason for this can be seen in FIG. 11, which illustrates a typical power spectrum of the reflected signal in range representation. Most of the relevant spectral density is located in a range from 30 cm to 110 cm, wherefore the range interval may be chosen accordingly. Also, in this example, any object further away than 110 cm would be located inside or even behind a vehicle seat 31, 32, 33. Range gating greatly reduces the number of (relevant) sample bins in each column, e.g. from 256 to 21.

[0064] Afterwards, the Capon method is performed (at 130). At 135, a sample covariance matrix {circumflex over (R)} is calculated as:

[00008]$\hat{R}=\frac{1}{K}\underset{k=0}{\overset{K-1}{.Math.}}x\left[k\right]x^H\left[k\right]$

wherein x[k] is an M-dimensional output vector representing the sample data in each channel in range representation and K is the number of data samples considered (e.g. K=21 after range gating). At 140, an inverse sample covariance matrix {circumflex over (R)}.sup.−1 is calculated.

[0065] Then, at 145, a steering vector a(θ) for each of a plurality of scanning angles θ is generated. The steering vector a(θ) corresponding to a scanning angle θ is given by a column vector and for a uniform linear array with M channels it is defined as

[00009]$a\left(\theta \right)=\left[\begin{array}{c}1\\ e^{j\left(\frac{2\pi \mathrm{dsin}\theta }{\lambda }\right)}\\ .Math.\\ e^{j\left(\frac{2\pi \left(M-1\right)\mathrm{dsin}\theta }{\lambda }\right)}\end{array}\right]\in {ℂ}^{M\times 1}$

wherein λ is the wavelength of the reflected signal and d is the distance between two receiving antennas.

[0066] When the covariance matrix and the steering vector have been determined, a Capon power spectrum P.sub.Cap(θ) is calculated (at 150) with

[00010]$P_{\mathrm{Cap}}\left(\theta \right)=\frac{1}{a^H\left(\theta \right){\hat{R}}^{-1}a\left(\theta \right)}$

Then, at 155, at least one angle of arrival θ.sub.i can be determined by examining the peaks of the Capon power spectrum P.sub.Cap(θ). In FIG. 5 and FIG. 6, which respectively show the power spectrum for the Bartlett method and the Capon method, it can be seen that the latter shows an improved signal-to-noise ratio. Also, in FIG. 5, which represents a simulation based on two targets disposed at 0° and 10°, respectively, the Bartlett method is unable to resolve two separate peaks, while the Capon method shows these peaks clearly. In FIG. 6, which represents the situation illustrated in FIG. 1 with the occupant 20 on the left seat 31 and the backpack 21 on the right seat 33, the peak resulting from the backpack 21 at around 25° has about half the height of the peak from the occupant 20 at about −15°. With the Bartlett method, the respective peaks have a similar height, making it nearly impossible to distinguish between an occupant 20 and an inanimate object 21.

[0067] According to another embodiment, which will now be explained with reference to FIG. 16, a MUSIC (multiple signal classification) method is employed. In a first step, at 200, the detection signal is emitted and the reflected signal is received. At 205, the sample data are recorded. Then, at 210, a portion of the sample data corresponding to reflections by static objects is removed. At 215, a range FFT is performed to transfer the sample data into range representation. At 220, range gating is performed.

[0068] At 230, the MUSIC method starts. At 235, the sample covariance matrix {circumflex over (R)} is calculated as described above with respect to the Capon method. At 240, the eigenvalues and eigenvectors of the sample covariance matrix {circumflex over (R)} are calculated. Assuming that all columns of {circumflex over (R)} are linearly independent, there will be M eigenvectors with corresponding eigenvalues, some of which eigenvalues could possibly be identical.

[0069] Then, at 245, the eigenvalues are sorted in descending order and the eigenvectors are adjusted. With D being a number of targets, M−D smallest eigenvalues and corresponding eigenvectors are selected to determine a noise subspace G (at 250). At 255, the product GG.sup.H is calculated.

[0070] At 260, a steering vector a(θ) for each of a plurality of scanning angles θ is generated as explained above with reference to the Capon method. Then, at 265, a MUSIC power spectrum P.sub.MUSIC(θ) is calculated with

[00011]$P_{\mathrm{MUSIC}}\left(\theta \right)=\frac{a^H\left(\theta \right)a\left(\theta \right)}{a^H\left(\theta \right){\mathrm{GG}}^{H}a\left(\theta \right)}$

If a(θ) is normalized, this can be simplified as:

[00012]$P_{\mathrm{MUSIC}}\left(\theta \right)=\frac{1}{a^H\left(\theta \right){\mathrm{GG}}^{H}a\left(\theta \right)}$

[0071] At 270, at least one angle of arrival θ.sub.i can be deduced by examining the peaks of the MUSIC power spectrum P.sub.MUSIC(θ). As can be seen from both FIGS. 5 and 6, results can still be improved compared to the Capon method. In the scenario of FIG. 5, the signal-to-noise ratio is improved by about 17 dB as compared to Capon. The two peaks are more clearly separable. In FIG. 6, a single peak (at −15°) representing the occupant 20 is much more pronounced than in the Capon power spectrum, with a side peak at about −40° being reduced and the peak at about 25° corresponding to the backpack 21 being even lower. It should be noted that the diagram in FIG. 6 corresponds to a variant where no range gating has been performed as indicated by the dotted arrow in FIGS. 15-17, i.e. steps 120 and 220, respectively, have been omitted. FIG. 12 shows a MUSIC power spectrum P.sub.MUSIC(θ) where range gating has been performed. Although the difference between the main peak at −15° and the side peak at −40° is not as pronounced as in FIG. 6, the peak corresponding to the backpack 21 has almost disappeared.

[0072] According to a third embodiment, which will be described with reference to FIG. 17, a different method is used which can be referred to as a root-MUSIC method, which starts at 330. The steps 300 to 355 are identical to the steps 200 to 255 described above. After determining the noise subspace, the roots z.sub.i of a root-MUSIC polynomial J(z) with

J(z)=z.sup.M-1p.sup.T(z.sup.−1)GG.sup.Hp(z)

are determined at 360, where

[00013]$p\left(z\right)=\left[\begin{array}{c}{z}^0\\ .Math.\\ z^{M-1}\end{array}\right].$

FIG. 7 is a diagram illustrating the positions of the roots z.sub.i in relation to a unit circle C1 in the complex plane. Generally, the polynomial J(z) has a total of 2(M−1) roots, i.e. in this case 14 roots z.sub.i. As can be seen in FIG. 7, the roots occur pairwise with 7 roots z.sub.i inside the unit circle C1 and 7 roots z.sub.i outside the unit circle C1. At 365, the M−1 roots z.sub.i inside the unit circle are selected. Furthermore, at 370, an inner radius for an inner circle C2 is defined. This inner radius is smaller than 1 and may e.g. be 0.7. Also, an angle interval is defined for each of the seats 31, 32, 33. For example, the angle intervals could be [−75°; −18° ] for the left seat 31, [−12°; 12° ] for the middle seat 32 and [18°; 75] for the right seat 33. With the boundaries defined by the unit circle C1, the inner circle C2 and the angle intervals, an associated area A1, A2, A3 can be defined for each of the vehicle seat 31, 32, 33, as illustrated in FIG. 9. In a simpler version of the method, the inner circle C2 could be omitted as a criterion, as illustrated in FIG. 8. Now, at 375, only roots z.sub.i inside the respective associated area A1, A2, A3 are considered. In the example illustrated in FIG. 13, two roots z.sub.4, z.sub.5 are located in the associated area A1 of the left seat 31. Therefore, a counter of buffer for the left seat 31 is increased by 2. In corresponding buffers for the middle seat 32 and the right seat 33 are decreased by 1, because no roots are located in their respective associated area A2, A3. At 380, each buffer is compared to a threshold, in this case 40, and when the threshold is exceeded, the corresponding seat 31, 32, 33 is considered occupied. In this case, some safety-relevant system like a seatbelt reminder for the respective seat 31, 32, 33 can be triggered.

[0073] When the roots z.sub.i in have been determined, an angle of arrival θ.sub.i can also be determined explicitly for each root z.sub.i, as

[00014]${\theta }_i=\arcsin \left(\frac{\lambda }{2\pi d}\arg \left(z_i\right)\right)$

[0074] FIGS. 13 and 14 refer to an embodiment where the range gating has been restricted so that only a single sample, corresponding to a single range value, has been used. In this case, the sample corresponds to a range value of 70 cm. As can be seen in FIG. 13, using the root-MUSIC method, the occupant 20 can be successfully identified by the two roots z.sub.4, z.sub.5 that are located in the associated area A1 of the left seat 31. Using the spectral MUSIC method, the result of which is shown as the black line in FIG. 14, a single peak can be identified at −15°, leading to a successful identification of the occupant 20. Using a single sample for the Capon method, on the other hand, yields a power spectrum that does not allow for any identification, as shown by the grey line in FIG. 14. The spectrum contains a large number of peaks which are spread over the entire angle range. The reason for this can be seen in that the Capon method requires calculation of the inverse covariance matrix {circumflex over (R)} to calculate the spectrum. With a single sample, however, the columns of the covariance matrix k become correlated, wherefore the inverse covariance matrix {circumflex over (R)}.sup.−1 cannot be successfully calculated. Accordingly, the Capon algorithm cannot be used if the columns of the covariance matrix {circumflex over (R)} are correlated.