GB2564085A - Method and apparatus for digital signal processing of a multichannel radar - Google Patents

Abstract

A digital processing system 401 for radar comprises an Application Specific Integrated Circuit (ASIC) or Field Programmable Gate Array (FPGA), wherein the system receives samples from a plurality of channels. The channels are simultaneously sampled at a uniform sampling rate and the system assembles the samples during each sample period and interleaves them onto one or more time division multiplex (TDM) sequences (see figure 2) containing the channel samples and applies the sequences to a digital signal processing chain. The system may operate on samples derived from a radio frequency modulation scheme consisting of a sequence of fast chirp modulated pulses. A pre-scan may be carried out to determine a subset of range Fast Fourier Transform (FFT) bins to store during each cycle, where the subset has an upper limit less than the longest FFT.

Description

Embodiments of the patent disclosure generally relate to radar systems and more specifically to radar digital signal processing implemented in an Application Specific Integrated Circuit (ASIC) or Field Programmable Gate Array (FPGA) for advanced driver assistance systems (ADAS).

:\C: K kJ U LJ

New vehicles are being introduced that embody advanced driver assistance systems (ADAS) to enhance safety and reduce the driving burden. In some cases going as far as providing a self-drive capability where the driver is no longer in control. Such systems may employ several different technologies, for example radar, cameras and lidar feeding an Electronic Control Unit (ECU) that makes driving and safety decisions. Reference [1] provides an overview of the current state of the art in radar technology and provides background material explaining the terms and techniques described herein. Relevant prior art patents are given in [2] - [6],

Radar is an essential technology for providing accurate distance and speed for objects in a wide and deep field of view, in all weather and prevailing light conditions. These radar systems are usually based on millimetre wave frequency band transmission and reception. The conventional embodiment employs a Radio Frequency (RF) transceiver device that emits a sequence of frequency modulated continuous wave pulses (FMCW) with high output power and wide bandwidth. The pulses, often referred to as chirps, are usually sent from a plurality of transmit antennas to help shape the spatial beam pattern in the direction of interest, and to distribute the power across multiple power amplifiers. A plurality of receive antennas is employed to receive the back scattered echoes. The received signal from an individual antenna is mixed with a feed from the transmit modulation forming a new signal that contains beat frequencies related to each echo time delay.

Current automotive radar systems often have more than one RF transceiver, as described above, where each transceiver supports a small number of antenna. These are combined to support longer or more complex antenna geometries.

In one embodiment all transmit and receive antennas can be co-located forming a two dimensional antenna array for providing Azimuth angle, in which case each receiver has a common mixing feed from the transmit modulation.

In a second embodiment all transmit and receive antennas can be co-located forming a three dimensional antenna array for providing Azimuth and Elevation angles. As in the first embodiment each receiver has a common mixing feed from the transmit modulation.

In a third embodiment transmit and receive antennas may be distributed over physically separate modules and operate at different times, so as to avoid interference and mitigate shadowing and blocking.

Each demodulated antenna signal is converted from analogue to digital representation in an analogue to digital converters (ADCs). This converter can be co-located in the RF transceiver device

04 18 or separate to it. The digital signals are sent over a high speed bus to a Digital Signal Processing Device, which forms the subject of this disclosure.

The Digital Signal Processing Device analyses the beat frequencies, performs detection of signals in noise and interference, associates measurements to tracks, forms new tracks from un-associated measurements, deletes tracks where measurements have not been associated for some time, and finally forms a list of tracked object metrics, e.g. spatial position, velocity, reflectivity, signal-to-noise ratio together with associated confidence metrics.

Automotive radar has traditionally been classified in operating modes as parking radar (PR), short range radar (SRR), medium range radar (MRR) and long range radar (LRR), and have differed in the type of modulation, bandwidth and transmit beam width employed. Typical SRR and MRR systems are using 4 receive antennas, whereas LRR systems are more likely to use 8 or 12 antennas. Recent advances in RF technology have enabled a modulation scheme sometimes called Fast Chirp, which consists of short duration, high bandwidth frequency modulation, such that the fastest practical movement of any observed vehicle, relative to the radar platform, is negligible during the time it is illuminated by an individual pulse. One immediate advantage of the fast chirp modulation is that the same modulation scheme can be employed in PR, SRR, MRR, and LRR by just changing the modulation parameters and tracking variables. A second advantage is that the object speed and range may now be independently estimated, which is an important consideration when the driving environment is complex and cluttered.

The main challenges of ADAS radar supporting fast chirp are how to implement the digital signal processing in the most area and power efficient way on a single chip ASIC which may or may not include the RF transceivers and ADCs in a mixed signal chip. The current state of the art is not well suited to achieve a single chip implementation. This is because of the very high digital signal processing requirements that arise from performing Fast Fourier Transforms (FFT) in Range, Doppler, Azimuth and Elevation. This is compounded by performing these transforms for each of a high number of receive antennas, typically 12, necessary to get good spatial resolution. A second problem that is faced by the current state of the art is the storage requirements necessary to hold all the samples in a measurement cycle, see for example Patent US20160018511 para 0033, which has necessitated using RAM external to the ASIC. Finally the functional safety requirement of taking an appropriate emergency response in under 100 ms, means that the time allocated to the radar processing should be under 20 ms.

Figure 1 shows the context of a multichannel radar having multiple antennas, a Radio Frequency (RF) transceiver, analogue-to-digital conversion (ADC) and digital signal processing (DSP).

Figure 2 explains the concepts of single channel, multichannel and time division multiplex (TDM)

Figure 3 illustrates the high level operation of the digital signal processing having phases operating to recover range, Doppler and Azimuth information in cycles of the same name and how these relate in time to the generation of pulses in pre-scan and measurement cycles.

Figure 4 provides detail on the inputs, output and processing blocks for the digital signal processing

With reference to Figure 1, the present disclosure is of methods and apparatus for radar digital signal processing (DSP) (103) shown in the context of a system comprising of a Radio Frequency (RF) transceiver block (101) and an analogue-to-digital (ADC) conversion block (102) and the Digital Signal

04 18

Processor (103) providing at least an object list (105) in the form of output. The blocks 101,102 and 103 may be combined in any way on a single Application Specific Integrated Circuit (ASIC) and be combined with power supply regulation circuitry on the same ASIC or they can be contained in separate ASICs.

Hereinafter the term channel is used to represent the signal, in whatever form analogue or digital, that was received by an individual antenna. The term multichannel is used to represent a plurality of channels. The term Time Division Multiplex (TDM) stream is the arrangement of multichannel samples interleaved into a single stream of samples at a sample rate multiple at least equal to the number of channels. These concepts are shown in Figure 2.

The present disclosure is of methods and apparatus for radar digital signal processing that operates on one or more TDM streams to perform fast chirp frequency and spatial analysis, detection, object association and tracking for a system that may be configured for operation modes PR, SRR, MRR and LRR.

In one aspect of the system the digital signal processing may be configured to operate on one or more TDM streams originating from a system configured for cycling or alternating between operation modes, one example of which is shown in Figure 3 where the first measurement cycle is from SRR and the next is from LRR. The reader will recognise that this example does not preclude alternation or cycling within other operation modes.

In one aspect of the system the digital signal processing may be configured, according to a programmed operation, to operate on one or more TDM streams originating from any one of these antenna configurations (a) a co-located long aperture antenna array providing only Azimuth spatial angle (b) a co-located but partitioned three dimensional antenna array where some of the antennas provide Azimuth spatial angle and the remaining provide Elevation angle (c) a distributed set of antenna sub-arrays

In one aspect of the system, as shown in Figure 3, the RF device generates a sequence of P pulses prior to a new measurement cycle in a mode known as pre-scan. A short gap of several pulse durations may be necessary to separate the pre-scan cycle from the main measurement cycle. The processing proceeds in series to operate over Range, Doppler, Azimuth/Elevation cycles such that the processing is complete before the start of the next pre-scan. A tracking cycle follows the Azimuth/Elevation cycle, and it must complete in the time of a measurement cycle, but can overlap the start of the next pre-scan cycle.

In one aspect of the system, one or more RF transceiver devices generate a sequence of L pulses in a measurement frame. Echoes from each pulse are received by K antennas, each demodulated and uniformly sampled to derive N samples of M-bits over the pulse duration. The ADC system (102) delivers the set of K*N*L by M-bit samples for a measurement frame to the digital signal processing (103) as a stream K*M bits wide by N*L time samples. N and L are conventionally a power of 2, but this is not a restriction of this embodiment, to enable Range and Doppler processing by FFTs of length N and L respectively. In our exemplar system the maximum value of K is 12, the maximum value of L is 256, the maximum value of N is 2048, the maximum value of M is 12 and the sampling rate is 40 MHz at each antenna.

In one aspect of the system, as shown in Figure 4, K antenna samples at each time instant are applied to the input of the digital signal processing system (401). This multichannel sample stream is

04 18 converted to one or more time division multiplex sample streams by (TDM) module (402) and applied to a pipelined time division multiplex FFT (403) operating at K times the antenna sample rate for producing K Range FFT outputs interleaved in a time division multiplex stream. This means that the time domain samples are applied in an interleaved streaming arrangement without the requirement to store samples for any antenna prior to starting FFT processing. A second advantage of this approach is that the logic requirement does not scale by K, only the internal FFT working memory requirement scales by K. This is a significant improvement over the prior art.

FFT architectures supporting time division multiplex and pipelined streaming samples are well known in prior art.

In one aspect of the system it has a time division multiplex bit-reversed index to natural order module (404). The pipelined time division multiplex FFT outputs frequency domain bins for each interleaved channel in bit-reversed index format and the system requires the frequency domain bins in a natural order. The implementation allows for the natural order to be either 0 ... N/2-1 or -N/2 ... N/2-1. The first natural order is used during Range FFT processing. The second type of natural order is used for Doppler FFT, Azimuth/Elevation FFT processing.

In one aspect of the system it has a time division multiplex windowing module (405). The need for windowing functions in FFT processing is well known in the prior art. This module applies the same window function to each channel. In this implementation the window is applied in the frequency domain (after the FFT) by convolution instead of in the time domain (before the FFT) by multiplication, however both methods are equally applicable and within the scope of this embodiment.

In one aspect of the system it has a beamforming module (406). The beamforming module can coherently combine all, or a subset of, K antenna signals in a time division multiplex stream into an aggregate channel stream. The aggregate stream is derived by combining the outputs from a subset-point spatial FFT to achieve a desired look-direction enhancement. The subsets can be the receiver channels that correspond to antennas arranged for Azimuth reception and those arranged for Elevation reception.

In one aspect of the system it has a power spectral density module (407). This module integrates linear power samples from the beamformer at each Range bin over the P pulses in a pre-scan and restarting again over the L pulses of a measurement cycle, to provide an incoherent signal to noise ratio gain in the beamformer look direction.

In one aspect the system it uses two simultaneous constant false alarm rate (CFAR) modules (408) with one optimised for interference rejection and the other optimised for white noise rejection and a decision rule to combine the detector outputs. In another embodiment it could just as well use a single CFAR module. The CFAR module(s) have run-time configuration of guard cells, pre/postwindow length, pre/post-window sub-window length (CASH-CFAR), pre/post-window ordered statistic index (GO/SO-CFAR), threshold and threshold modifier. The CFAR outputs at least the sample, a detection flag and an estimate of its signal-to-noise-plus-interference ratio.

In one aspect of the system it has a peak detection (reduction) module (409) that operates a sliding window of samples from the CFAR detected sample output and only outputs the central detected sample if it has the highest magnitude.

In one aspect of the system, Range FFT bin indices identified by CFAR detection during pre-scan are stored (419) in preparation for the main measurement cycle. During the main measurement cycle,

04 18 complex samples prior to the beamforming module (406) from Range FFT bin indices identified during pre-scan are the only samples stored (420) by only routing these through the data selector (412) . The same indexed samples are stored for each of the L pulses and for all K antennas. The prescan mode is applicable for SRR using long FFT lengths and fast moving objects. For typical LRR usage then pre-scan is unnecessary and the Range FFT bins identified from the L pulses on the previous measurement cycle can be used to qualify complex sample storage on the current measurement cycle. An advantage of this method is that there's no need to store the FFT output samples for all Range bins and antennas. We only store up to a fixed maximum, for example in our exemplar system this is 128. Since storage comprises the most area in an ASIC implementation of this digital signal processing architecture then this provides a significant advance over prior art.

In one aspect of the system, at the end of an L chirp cycle, the stored samples corresponding to the same Range bin on each pulse are read out of the store (420) and applied back via path (417) to the TDM module (402). This now constitutes the start of a Doppler processing cycle and this continues for every Range bin stored during the Range processing cycle. The Doppler cycle applies samples in TDM format through modules 403, 404, 405, 406, 408, 409, 411, 412 and into storage in 420, where unused modules are bypassed. The CFAR (408) is used to identify the indices of beamformed samples that exceed a threshold above the noise and interference in the Doppler FFT. The beamforming can be on subsets of the K antennas depending on whether the antennas are arranged for two or three dimensional reception. Only K samples corresponding to bin indices identified by the CFAR are stored (420) together with their Doppler FFT index and SNR estimate. Since the CFAR operates on streaming data there is no need to buffer-up or store data other than detected samples, which provides a significant improvement on the prior art.

In one aspect of the system, at the end of a Doppler processing cycle, the stored samples corresponding to each antenna subset are read out of the store 420 and applied back via path (417) to the TDM module (402). The TDM module interpolates each sample set up to a large power of 2 by padding with zeros. In our exemplar this is 256, 512 or 1024 points. This now constitutes the start of the Azimuth/Elevation processing cycle and this continues for every sample set stored during the Doppler measurement cycle. The Azimuth/Elevation cycle applies samples in TDM format through modules 403, 404, 40,109, 119, where unused modules are bypassed. In this mode a direction-of-arrival (DOA) peak detector (424) is used to identify the peaks in a highly interpolated FFT spectrum, which requires a different algorithm to the CFAR. Only samples corresponding to bin indices identified by the DOA peak detector are stored together with their Azimuth/Elevation FFT index and SNR estimate. Since the DOA peak detector operates on streaming data there is no need to buffer-up or store data other than detected samples, which provides a significant improvement on the prior art.

In one aspect of the system, at the end of the Azimuth/Elevation processing cycle we have a stored list of measurements, their SNR, and interpolated Range bin, interpolated Doppler bin, interpolated Azimuth bin and interpolated Elevation bin. These measurement are applied to a transform unit (413) module which converts from interpolated bin indices to real-world units in single precision floating point format, so Range bin becomes meters, Doppler bin becomes meters per second, Azimuth and Elevation bins become radians. Furthermore the SNR is converted to standard deviation for each respective real-world measurement by reference to the Cramer-Rao lower bound for the estimation method. Furthermore the conversion to real-world coordinates takes account of the mode in operation; PR, SRR, MRR or LRR so that an object detected in more than one mode is correctly mapped to the same natural real-world coordinates allowing the subsequent processing to take advantage of fusing measurements from different operational modes. The real world

04 18 measurements plus the associated standard deviation of those measurements are further transformed from a polar/spherical coordinate system to a Cartesian coordinate system. In so doing it derives the Cartesian measurement vector and the Cartesian measurement error covariance matrix, which are both stored in memory. The polar/spherical coordinates and their diagonal measurement error covariance matrix are useful for operating on by an Extended Kalman Filter, whereas the Cartesian measurement vector and associated measurement error covariance matrix are useful for operating on by a Linear Kalman Filter.

In one aspect of the system it has a track association unit (414). This reads in each active track state prediction that the Kalman Filter is maintaining and calculates a cost function relating how close the measurement is to the track prediction, whilst taking the measurement error covariance matrix and track state prediction error covariance matrix into consideration to ensure all distances are normalised facilitating a direct comparison. Only those distances less than a programmed threshold are retained and the distance is written to a cost function matrix, indexed by track and measurement. The next step in the measurement association unit, once distances are calculated from all combinations of track and measurement pairs, is to perform a neighbour assignment algorithm for example Nearest Neighbour or Global Nearest Neighbour to find a possible association of tracks and measurements in some optimum sense.

In one aspect of the system it has a track management unit within the track association unit (414). This is responsible for initialising new Candidate tracks with un-associated measurements, deleting Candidate and Established tracks that have not had an associated measurement for a predetermined number of measurement cycles and updating existing Candidate and Established tracks with associated measurements. The track management is based on standard M-out-of-N logic for testing the track quality based on the number of successful track updates in the last N attempts. The M-out-of-N logic can be applied to both the criterion for promoting a track from Candidate to Established, and for deleting a Candidate track.

In one aspect of the system it has a Kalman filter unit (415). The Kalman filter supports multiple state models including at least one of Constant Velocity, Constant Acceleration and Coordinated Turn. It supports state update based on Linear or Extended Kalman filter steps.

In one aspect of the system the Established tracks are sent over a digital bus, for example FlexRay or Ethernet to an Electronic Control Unit (ECU) that can use the object list and confidence metrics to perform higher level processing necessary for vehicle control or safety.

Claims (32)

1. A digital signal processing system for radar comprising an Application Specific Integrated Circuit (ASIC) or Field Programmable Gate Array (FPGA), wherein the digital signal processing system receives samples from a plurality of channels where the channels are simultaneously sampled at a uniform sampling rate and the digital signal processing system assembles the plurality of channel samples during each sample period and interleaves them onto one or more time division multiplex (TDM) sequences containing said channel samples and applies these TDM sequences to a digital signal processing chain.

2. The digital signal processing system according to claim 1, applied in the area of automotive radar

3. The digital signal processing system according to claim 1 or 2 where the digital signal processing system is integrated into a mixed signal ASIC with one or more analogue to digital converters.

4. The digital signal processing system according to claim 1, 2 or 3 where the digital signal processing system is integrated into a mixed signal ASIC with one or more RF transceivers.

5. The digital signal processing system according to claim 1, 2, 3 or 4 where the digital signal processing system is integrated into a mixed signal ASIC includes at least one power supply regulator circuit.

6. A digital signal processing system according to claim 1, 2, 3, 4 or 5 that operates on samples derived from an RF modulation scheme consisting of a sequence of fast chirp RF modulated pulses.

7. A digital signal processing system according to claim 6 that alternates or sequences measurement cycle processing of radar signals generated from parking radar, short range radar, medium range radar or long range radar.

8. A digital signal processing system according to claim 6 or 7 that operates during a period prior to the current measurement cycle, referred to hereinafter as a pre-scan, in order to determine a subset of Range FFT bins to store during the measurement cycle, where the subset has an upper limit that is less than the longest FFT.

9. A digital signal processing system according to claims 6, 7 or 8 where the FFT operates on TDM streaming data at its input and produces TMD streaming data at its output.

10. A digital signal processing system according to claim 9 including a TDM streaming bitreversed index to TDM streaming natural order operation that provides output samples in order 0...N/2 for range processing cycle and -N/2...N/2-1 for Doppler, Azimuth/Elevation processing cycles.

11. A digital signal processing system according to claim 10 including a TDM streaming windowing operation that is performed by convolution in the frequency domain.

12. A digital signal processing system according to claim 11 including multiple beamforming operations to coherently combine a subset of the TDM channels into orthogonal spatial beams and to combine the orthogonal spatial beams according to a programmed operation.

13. A digital signal processing system according to claim 12 where a first subset corresponds to TDM steam time slots containing samples from antennas arranged for reception in Azimuth and an optional second subset corresponding to TMD stream time slots containing samples from antennas arranged for reception in Elevation.

14. A digital signal processing system according to claim 13 where the subset beamformer outputs of all pulses in a measurement cycle are incoherently combined to produce a subset power spectral density whose frequency domain bins correspond to Range bins for that subset beamformer

15. A digital signal processing system according to claim 14 that includes a detection operation on one or more of the subset power spectral densities based on at least one CFAR processor, however if two CFAR processors are used then one is optimised for operation in the presence of interference and the other optimised for operation in white noise and a method is provided to combine the detection result from each to output a single detection flag for detected samples higher than a programmable threshold above the noise-plus-interference estimate and also to return an estimate of the signal-to noise-plus-interference ratio of each detected sample.

16. A digital signal processing system according to claim 15 where the CFAR programmable threshold can be a function of Range bin.

17. A digital signal processing system according to claim 15 or 16 or that includes a peak extraction operation that removes neighbouring detection flags next to the largest detected sample in a programmable length window of samples.

18. A digital signal processing system according to claim 17 that includes a peak interpolation operation that performs digital interpolation to return the peak position and amplitude to a higher resolution by assuming the samples in the programmable window are derived from an underlying bandlimited signal.

19. A digital signal processing system according to claim 18 where the only Range FFT bin samples stored are those identified by the peak extraction unit from a combination of either or both of the pre-scan or previous measurement cycle.

20. A digital signal processing system according to claim 19 where the stored samples from the same Range bin index, starting with the earliest, for all pulses in a measurement cycle are fed back around into the TDM unit which sequences them in antenna order to start a Doppler processing cycle, and repeats this for all the stored Range bins.

21. A digital signal processing system according to claim 20 where Doppler processing consists of reusing the processing units used during the Range processing cycle, in particular the time division multiplex FFT, time division multiplex natural order, time division multiplex frequency domain windowing, time division multiplex beamforming, CFAR detection and peak extraction.

22. A digital signal processing system according to claim 21 where the only Doppler FFT output samples stored for Azimuth/Elevation processing are ones detected in the FFT by the peak extraction unit during the Doppler processing cycle for that particular FFT.

23. A digital signal processing system according to claim 22 where Azimuth/Elevation processing consists of reusing the processing units used during the Range and Doppler processing cycles, in particular the time division multiplex FFT and time division multiplex natural order units.

24. A digital signal processing system according to claim 23 where the antenna samples are zero-padded to a high power of 2 during application to the FFT to form an interpolated spatial spectrum in Azimuth/Elevation.

25. A digital signal processing system according to claim 24 where the only Azimuth/Elevation FFT output samples stored are ones detected by a direction-of-arrival peak detector working on the interpolated spatial spectrum.

26. A digital signal processing system according to claim 25 where the measurement list of interpolated Range bin, interpolated Doppler bin, interpolated Azimuth/Elevation bin and SNR are provided to a measurement transform unit to derive single precision floating point real-world coordinates in metres, meters per second, radians and the standard deviation of those measurements based on the SNR and estimator.

27. A digital signal processing system according to claim 26 where the positional and speed measurement list in real-world coordinates are optionally mapped to a Cartesian measurement vector and Cartesian measurement covariance matrix.

28. A digital signal processing system according to claim T1 where the coordinate transform unit is followed by a track association unit that calculates a cost function for each pairing of measurement vector and track state vector representing a confidence in the measurement being a measurement of the track.

29. A digital signal processing system according to claim 28 where the most likely measurement and track pairings are retained.

30. A digital signal processing system according to claim 29 where a track management function determines the initialisation of a new Candidate track from an un-paired measurement, update each track with an associated measurements, deletes a track according to M-out-ofN logic and promotes a track from Candidate to Established based according to M-out-of-N logic.

31. A digital signal processing system according to claim 30 where the tracks are updated using a Linear or Extended Kalman filter operating in accordance with an applicable state model.

32. A digital signal processing system according to claim 31 where the Established tracks and confidence metrics are sent over a digital data bus.