The present invention is directed to radar systems, and in particular to radar systems for vehicles.
The use of radar to determine range, velocity, and angle (elevation or azimuth) of objects in an environment is important in a number of applications including automotive radar and gesture detection. Radar typically transmits a radio frequency (RF) signal and listens for the reflection of the radio signal from objects in the environment. A radar system estimates the location of objects, also called targets, in the environment by correlating delayed versions of the received radio signal with the transmitted radio signal. A radar system can also estimate the velocity of the target by Doppler processing. A radar system with multiple transmitters and multiple receivers can also determine the angular position of a target.
A radar system consists of transmitters and receivers. The transmitters generate a baseband signal which is up-converted to a radio frequency (RF) signal that propagates according to an antenna pattern. The transmitted signal is reflected off of objects or targets in the environment. The received signal at each receiver is the totality of the reflected signal from all targets in the environment. The receiver down-converts the received signal to a baseband signal and compares the baseband received signal to the baseband signal at one or more transmitters. This is used to determine the range, velocity, and angle of targets in the environment. The resultant data sets produced by the range estimation, Doppler/velocity estimation, and angle estimation are stored in so-called radar data cubes (RDCs) that store complex samples corresponding to different values of range, Doppler/velocity, and angle to complex vectors. Conceptually, the data is stored in a cube with the axes corresponding to range, Doppler/velocity, and angle. The information stored in a radar data cube is stored in a memory for subsequent processing. This may require a very large amount of memory for high resolution radar systems. This may also take a long time to transfer externally to an integrated circuit (off-chip memory). The storage of large blocks of data in memory (such as the storage of RDCs) is also prone to errors. Methods and mechanisms are needed to detect and correct errors.
Embodiments of the present invention provide methods and an apparatus to improve the performance of a radar system. An exemplary radar system is configured to optimize a memory storage and to access of radar data cubes (RDCs) generated by Doppler and angle processing in the radar system. Depending on the throughput of the radar system (the number of newly generated RDCs per second), and the number of bins (the quantization level) used to specify range, velocity, and angle of the targets in the environment, storing the RDCs can have a large memory requirement. Embodiments of the present invention provide methods and techniques for categorizing regions of the RDCs and/or thresholding values to sparsify RDCs, optimizing DRAM access patterns in reads and writes to the first RDC, transforming samples with on-the-fly block floating point in the first RDC to reduce bit-width, pipelining the execution to maximize concurrency across different compute modules, adding error detecting or error correcting bits to samples for functional verification, and storing the sparsified versions of the RDCs to memory at specific locations for subsequent processing.
A radar sensing system for a vehicle in accordance with an embodiment of the present invention includes a transmit pipeline, a receive pipeline, a sparsifier processor, and a memory. The transmit pipeline includes a plurality of transmitters configured for installation and use on a vehicle and configured to transmit radio signals. The receive pipeline includes a plurality of receivers configured for installation and use on the vehicle and configured to receive radio signals that include the transmitted radio signals transmitted by the transmitters and reflected from objects in an environment. The receive pipeline is configured to correlate the received radio signals of the plurality of receivers with a plurality of time-delayed replicas of the transmitted radio signals to produce samples arranged in a first three-dimensional array (RDC) representing ranges of different objects in the environment as measured by different receivers. The receive pipeline is configured to process the first RDC to estimate a Doppler for given ranges and given receivers to produce a second RDC. The receive pipeline is configured to process the second RDC to estimate angles of targets in the environment for given ranges and Doppler to produce a third RDC. The receive pipeline is configured to output the second RDC and the third RDC to the sparsifier processor. The sparsifier processor is configured to analyze the second RDC and the third RDC to compute sparse versions of the second RDC and the third RDC. The sparse versions of the second RDC and the third RDC are stored in the memory for further processing.
A method for optimizing memory performance and access of a three-dimensional array (RDC) in a receive pipeline of a radar system for a vehicle in accordance with an embodiment of the present invention includes transmitting, with a plurality of transmitters, radio signals. A plurality of receivers receive radio signals that include the transmitted radio signals transmitted by the transmitters and reflected from objects in an environment. The received radio signals of the plurality of receivers are correlated with a plurality of time-delayed replicas of the transmitted radio signals to produce samples arranged in a first RDC representing ranges of different objects in the environment as measured by different receivers. The first RDC is processed to estimate a Doppler at a given range and for a given receiver to produce a second RDC. The first RDC is output from a correlation processor to a dynamic random access memory (DRAM) arranged by receivers in an inner loop, followed by range-bin and time. The first RDC is input from the DRAM to a Doppler processor by means of prefetching to a local buffer. The method further includes maximizing DRAM read performance by burst locality by preserving access to data from multiple receivers.
A method for optimizing memory performance and access of a three-dimensional array (RDC) in a receive pipeline of a radar system for a vehicle in accordance with an embodiment of the present invention includes transmitting, with a plurality of transmitters, radio signals. A plurality of receivers receive radio signals that include the transmitted radio signals transmitted by the transmitters and reflected from objects in an environment. The received radio signals of the plurality of receivers are correlated with a plurality of time-delayed replicas of the transmitted radio signals to produce samples arranged in a first RDC representing ranges of different objects in the environment as measured by different receivers. The first RDC is processed to estimate Doppler at a given range and for a given receiver to produce a second RDC. The method further includes computing one or more common block exponents for the first RDC to reduce a bit-width required to represent complex numbers in samples of the first RDC.
In an aspect of the present invention, an exemplary sparsifier processor estimates voxels (three-dimension data) in a radar data cube that corresponds to a range of low absolute velocity, and thereby tags voxels as belonging to a static image, which are to be stored for subsequent processing. For all remaining voxels, the sparsifier processor then performs a threshold comparison against the noise floors estimated for each range bin across all Doppler values and angles. The sparsifier processor may use a pipelined architecture to read radar data cubes from buffers as they are being computed, apply a sparsification process, and output sparse versions of the radar data cubes to a memory for subsequent processing.
These and other objects, advantages, purposes and features of the present invention will become apparent upon review of the following specification in conjunction with the drawings.
The present invention will now be described with reference to the accompanying figures, wherein numbered elements in the following written description correspond to like-numbered elements in the figures. Methods and systems of the present invention may achieve higher throughput of an exemplary radar system and have reduced memory requirements. Radar data cubes store complex numbers generated by processing the correlation between delayed versions of the transmitted signals with received signals. As discussed herein, the radar data cubes, which may be quite large, may be sparsified before storing to an external (e.g., off-chip) memory. The exemplary sparsification is achieved through one or more methods that include identification of regions in the computed radar data cubes of low absolute velocity and thresholding of remaining regions based on software thresholds or thresholds of noise floors estimated with power histograms. Exemplary embodiments may also achieve high throughput by employing a pipelined architecture for the sparsification step.
A target at a given range and angle, and with a given velocity, will produce a complex number which may be stored in memory at an associated location in the RDC. While the information is described as stored in a radar data cube, the information contained in the radar cube may be stored in a variety of manners including storing the information in a linear fashion. That is, storing the information for a first time and a first virtual receiver for all ranges in an array of memory elements, followed by the information for the second time and first virtual receiver and so on. These RDCs need to be stored to memory for subsequent processing. Information stored in an RDC is related to the strength of the signal (the magnitude of the complex number) corresponding to a particular range, Doppler, or virtual receiver. The magnitude of these numbers generally will exhibit locality, meaning large values will be concentrated around a small set of values of range, Doppler, or virtual receiver. For real-time radar systems, such as automotive radar systems, memory considerations may prevent the storage of entire RDCs. Thus, RDCs need to be sparsified, i.e. compressed to retain only the values corresponding to certain regions of interest without sacrificing accuracy of subsequent processing or retrieval.
Write operations to a “first RDC” will store samples in the first RDC in a particular order. Read operations from the first RDC will access samples in the first RDC in a different order. Write operations to the first RDC will exhibit locality in virtual receivers. Write operations first take place (inner loop) with respect to an index corresponding to the virtual receiver. That is, the first memory location to be written in will correspond to the first virtual receiver for the first range bin. The second memory location to be written in will be the second virtual receiver for the first range bin. Once data for all memory locations for the first range bin have been written in, data corresponding to the second range bin will be written to memory. Meanwhile, read operations from the first RDC for Doppler/velocity processing expect locality in the time axis. This mismatch can greatly reduce the performance of memory access by not utilizing locality. Thus, specific memory access patterns and prefetching to intermediate buffers need to be enabled to maximize memory throughput.
The computed samples in the first RDC, as well as in a “second RDC,” are complex numbers. Computing and storing floating point numbers requires the storage of exponents, which reduces accuracy in exchange for a larger range of values with a fixed number of bits. Meanwhile, fixed-point notation will limit dynamic range but will have better accuracy. As a trade-off, block floating-point notation may be used, which stores common exponents for a block of data, along with the bit-width optimized for range and accuracy. However, the common exponent per block needs to be computed on-the-fly and appropriately stored during writes and correctly decoded during reads.
There are several compute modules in a receive pipeline of a radar system that need to run concurrently for high performance. If the modules are run concurrently, the modules may consequently read and write smaller blocks of the radar cubes to intermediate buffer storage and thereby avoid large memory requirements. Thus, as described herein, to optimize the performance of an exemplary radar system, a pipelined architecture with different compute modules interfaced with appropriately sized buffers is needed.
However, as discussed herein, the storage of large blocks of data in memory is prone to errors. Especially, for automotive radars, where the system must be able to self-diagnose any errors and preferably correct them. Therefore, mechanisms are needed that will append suitable parity and check bits to detect and correct errors, before subsequent processing.
An exemplary radar system operates by transmitting one or more signals from one or more transmitters and then listening for reflections of those signals from objects in the environment by one or more receivers. By comparing the transmitted signals and the received signals, estimates of the range, velocity, and angle (azimuth and/or elevation) of the objects can be estimated.
There are several different types of signals that transmitters in radar systems employ. A radar system may transmit a continuous signal or a pulsed signal. In a pulsed radar system, the signal is transmitted for a short time and then no signal is transmitted. This is repeated over and over. When the signal is not being transmitted, the receiver listens for echoes or reflections from objects in the environment. Often, a single antenna is used for both the transmitter and receiver and the radar transmits on the antenna and then listens to the received signal on the same antenna. This process is then repeated. In a continuous wave radar system, the signal is continuously transmitted. There may be an antenna for transmitting and a separate antenna for receiving. One type of continuous wave radar signal is known as a frequency modulated continuous wave (FMCW) radar signal. In FMCW, the transmitted signal is a sinusoidal signal with a varying frequency. By measuring a time difference between when a certain frequency was transmitted and when the received signal contained that frequency, the range to an object can be determined.
A second type of continuous wave signal used in radar systems is a phase modulated continuous wave (PMCW) signal. In this type of radar system, the transmitted signal is a sinusoidal signal in which the phase of the sinusoidal signal varies. Typically, the phase during a given time period (called a chip period or chip duration) is one of a finite number of possible phases. A spreading code consisting of a sequence of chips, (e.g., +1, +1, −1, +1, −1 . . . ) is mapped (e.g., +1→0, −1→π) into a sequence of phases (e.g., 0, 0, π, 0, π . . . ) that is used to modulate a carrier to generate the radio frequency (RF) signal. The spreading code could be a periodic sequence or could be a pseudo-random sequence with a very large period so it appears to be a nearly random sequence. The spreading code could be a binary code (e.g., +1 or −1). The resulting signal has a bandwidth that is proportional to the rate at which the phases change, called the chip rate RC, which is the inverse of the chip duration TC=1/RC. By comparing the return signal to the transmitted signal, the receiver can determine the range and the velocity of reflected objects.
There are several ways to implement a radar system. One way, illustrated in
A radar system with multiple antennas, transmitters and receivers is illustrated in
The radar system 300 may be connected to a network via an Ethernet connection or other types of network connections 314, such as, for example, CAN-FD and FlexRay. The radar system will have memory (310, 312) to store software and data used for processing the radio signals in order to determine range, velocity, and location of objects. Memory 310, 312 may also be used to store information about targets in the environment. There may also be processing capability contained in the ASIC 208 apart from the transmitters 203 and receivers 204.
The description herein includes an exemplary radar system in which there are NT transmitters and NR receivers for NT×NR virtual radars, one for each transmitter-receiver pair. For example, a radar system with eight transmitters and eight receivers will have 64 pairs or 64 virtual radars (with 64 virtual receivers). When three transmitters (Tx1, Tx2, Tx3) generate signals that are being received by four receivers (Rx1, Rx2, Rx3, Rx4), each of the receivers is receiving the transmission from each of the transmitters reflected by objects in the environment. Each receiver can attempt to determine the range and Doppler shift (velocity) of objects by correlating with delayed replicas of the signal from each of the transmitters. Each physical receiver may then be “divided” into three separate virtual receivers, each virtual receiver correlating with delay replicas of one of the transmitted signals to yield a total of 4×3=12 virtual receivers.
The receivers of the radar system may be considered part of a receive pipeline that outputs complex valued digitized samples (i.e., values comprising a mathematical real component and a mathematical imaginary component) to a processor. The complex valued digitized samples may be produced based on correlations between the reflected signals and time-delayed replicas of the transmitted signals. Such correlation values may be determined through the use of a matched filter, although this is not the only way to produce these correlation values. These resulting samples are arranged in two-dimensional arrays known as time slices. The samples are placed into respective range bins of the two-dimensional array (as used herein, a range bin refers to a distance range corresponding to a particular time delay corresponding to the round-trip time of the radar signal from a transmitter, to the target/object, and back to the receiver). The different virtual receivers of the radar system define an x-axis of the two-dimensional time slice and the range bins define a y-axis of the two-dimensional time slice. A new time slice comprising complex valued samples may be generated every 2-30 microseconds; that is, every time a new correlation is produced for a given range bin. Over a longer time interval, herein referred to as a “scan” (typically, in a duration of 1-60 milliseconds or longer), multiple time slices are accumulated to form a three-dimensional radar data cube. An x-axis of the three-dimensional radar data cube is defined by time (with each respective time slice requiring 2-30 microseconds), while the receivers (or virtual receivers) define a z-axis of the three-dimensional radar data cube, and the range bins and their corresponding time delays (ranges) define a y-axis of the three-dimensional radar data cube 1 (RDC1). A radar data cube may have a preselected or a dynamically defined quantity of time slices. For example, a RDC may include 200 time slices or more preferably 3000 time slices or more.
The radar sensing system of the present invention may utilize aspects of the radar systems described in U.S. Pat. Nos. 9,846,228; 9,806,914; 9,791,564; 9,791,551; 9,772,397; 9,753,121; 9,599,702; 9,575,160 and/or 9,689,967, and/or U.S. Publication Nos. US-2017-0309997; US-2017-0307728 and/or US-2017-0310758, and/or U.S. patent application Ser. No. 15/496,038, filed Apr. 25, 2017, Ser. No. 15/689,273, filed Aug. 29, 2017, and/or Ser. No. 15/705,627, filed Sep. 15, 2017, and/or U.S. provisional applications, Ser. No. 62/486,732, filed Apr. 18, 2017, Ser. No. 62/528,789, filed Jul. 5, 2017, Ser. No. 62/573,880, filed Oct. 18, 2017, Ser. No. 62/598,563, filed Dec. 14, 2017, and/or Ser. No. 62/623,092, filed Jan. 29, 2018, which are all hereby incorporated by reference herein in their entireties.
As illustrated in
As illustrated in
A radar system can determine the velocity of an object by measuring the frequency offset between the transmitted signal and the received signal. The signal reflecting from a moving object will have a frequency shift relative to the transmitted signal. This is called the Doppler Effect and the frequency shift is called the Doppler shift. The Doppler shift is directly related to the radial velocity of the object relative to the radar system. A particular Doppler shift or frequency value may be referred herein as a Doppler value or just as a “Doppler.” Doppler is most noticeable with regards to trains passing through a road crossing where the frequency of the sound will go down once the train has passed the road crossing. The signal reflected off an object moving toward the radar sensor will have a higher frequency than the transmitted signal, and an object moving away from the radar sensor will have a lower frequency. If the radar sensor is also moving, then it is the relative velocity between the radar sensor and the object that will determine the amount of frequency shift. Thus, measuring the frequency shift will determine the relative velocity of an object in the environment. The time duration of the transmitted signal will determine the accuracy or resolution of the Doppler value.
A “slice” of a cube refers to a two-dimensional subset of the data where one of the variables (either range, time, or receiver) is fixed. A “skewer” of a cube refers to a one-dimensional subset of the data where two of the variables are fixed. To estimate the velocity of an object, a particular “skewer” of the radar data cube RDC1 is processed as shown in
As illustrated in
The Doppler resolution of a radar system is dependent on the number of time slices in a scan: the longer the scan the greater the Doppler resolution. For example, a scan of 500 time slices may provide for a maximum Doppler/velocity resolution of 2.0 m/s, while a scan of 2,000 time slices may provide for a maximum Doppler/velocity resolution of 0.5 m/s, a 4× increase. However, longer scans, while requiring larger memory, may also result in longer latency between successive radar images or objects. Nevertheless, a Doppler resolution of 0.5 m/s or lower is preferable. The information stored in radar cube 2 (RDC2) will then contain Doppler estimates for each range bin and virtual receiver. For example, a single scan stored in the RDC2 may preferable contain data from 256 virtual receivers for 512 range bins and for over 2000 Doppler values with resolution 0.5 m/s. This requires storage of about 260 million complex numbers, each requiring 4 bytes of storage, which is about 1,000 MB of storage, for the RDC2 for the single scan. As in the case of the RDC1, on-chip memory limitations and the communications limitation of off-chip bandwidth, make it impractical to store or transmit the entire RDC2 for each scan.
A radar system can estimate the angle of targets in the environment by beamforming in different directions (azimuthal and/or elevation), i.e., by using steering vectors to weight the different receiver outputs. This computation is performed in exemplary angle processor 602, illustrated in
Histogram Computation for Noise Floor Estimation:
Using a large number of receivers in an exemplary radar system, with corresponding antennas arranged in an array, provides higher angular accuracy and resolution, such as for the detection of vulnerable road users (VRU) in an automotive radar application. Like in the case of radar data cubes RDC1 and RDC2, higher resolution of the angle processor 602 can result in the size of the radar data cube RDC3 being impractical to store on-chip or to be transferred off-chip. For example, for 64 resolved angles, the size of a single scan of RDC3 will involve preferably 512 range bins, 200 Doppler bins, and 64 angle bins, each with a complex number taking up 4 bytes, totally a memory requirement of more than 25 MB per scan. This is impractical both for on-chip storage and off-chip communication.
Thus, an exemplary radar system should preferably choose only a small set of regions of interest (RoI) in the three radar data cubes which are to be stored in memory or sent off-chip and used for subsequent processing. Preferably, the samples within these regions have similar Doppler, range and/or angle (azimuth and/or elevation). In a preferred embodiment, these regions of interest could correspond to a subset of low absolute velocity, such as due to ground clutter. Alternatively, these regions could correspond to regions with signal power larger than certain thresholds, wherein these thresholds can be either specified in software or computed from estimating noise floors in previous time slices or controlled in an adaptive manner. In a preferred embodiment, these sparsification operations are done by an exemplary sparsifier processor 603 (part of the receive pipeline), illustrated in
In one embodiment, the output of the angle processor 602, which is stored in RDC3, is analyzed to compute a histogram of the magnitudes of the complex samples stored to estimate noise floors and subsequently threshold regions of interest in the radar data cubes. This computation is performed in an exemplary histogram engine 710, which is part of the sparsifier processor 603, illustrated in
The histogram of the power level is computed for each range bin computed across Doppler frequencies and angles and is used to estimate the noise floor.
Static Image Extraction:
In an exemplary radar system, such as in automotive radar, a large fraction of reflections received by the receivers may correspond to objects that are truly stationary, i.e., objects with near zero absolute velocity. In an automotive radar, the ground and objects along a road, such as lamp-posts and street signs, will all have zero absolute velocity. Any subsequent processing of the radar data cubes must store and analyze such objects.
In one embodiment, illustrated in
For an automotive radar, if the equipped automobile is stationary, then the static image corresponds to all objects which have an identified Doppler-based velocity of 0 m/s, i.e., voxels in all range and angle bins corresponding to the slice of RDC3 where the Doppler axis is 0. In the case of a stationary automobile, all such voxels for Doppler bin 0 and for all identified range bins and angles can be output as the static image. If instead, the radar is mounted on a moving automobile moving at a speed vautomobile, then the Doppler values vDoppler as a function of the angle of objects (relative to bore sight or the direction of the automobile) in the static image are given by
vDoppler=vautomobile×cos(angle)
In the above, the cosine of the angle of the object may be used to translate the relative velocity of the automobile in the direction of the object. For the case of a stationary automobile, the value of vDoppler for each angle is 0.
The static image extraction unit 711 identifies the voxels corresponding to stationary objects as the static image. In a preferred embodiment, a configurable parameter halfWindow is specified by a control processor, such that Doppler bins less than an absolute Doppler of halfWindow are chosen to belong to the static image. Specifically, a band of size fullWindow=2*halfWindow+1 is extracted centered around the computed vDoppler as discussed herein. Thereby, the static image extraction includes objects of ‘near’ zero absolute velocity where the threshold on the velocity is as specified in the halfWindow parameter.
In one embodiment, to optimize the runtime processing, the static image extraction unit processing can be split into two modules. In a first module, a first static voxel satisfying the static image criterion for each angle slice can be computed using the velocity formula, along with the window size to be used. This may be stored in a small memory index table. Then in a second module, the memory index table can be read along with the window size to identify and tag voxels which belong to the static image.
Thresholding Engine:
In a preferred embodiment, a thresholding engine 712, as illustrated in
Those voxels which satisfy the thresholding condition, i.e., are above the software specified threshold or above the noise floor computed in the previous radar scan, are stored in the memory. This applies to both the RDC2 and the RDC3 radar data cubes, generating their sparse versions denoted as sRDC2 and sRDC3 respectively. In a preferred embodiment, for RDC2, the thresholding engine 712 applies the thresholding function after a channelizer operation in the Doppler processor 601. The channelizer in 601 performs multiple FFTs on overlapping windows of a time series while the FFT in 601 performs an FFT over the whole time series. The channelizer output can be preferably used for super-resolution accelerated post-processing. Thus, sRDC2 (720) stores the sparse channelizer complex output on the axes of range, Doppler, and virtual receiver, while the sRDC3 (722) stores the sparse beamforming magnitude output with axes of range, Doppler, and angle.
DRAM Throughput Optimization in Storage of RDC1:
The output of the correlation processor 410 is stored as RDC1, with the three axes of time, range, and virtual receiver, with each value storing a complex number of preferably 32 bits, i.e., 16 bits each for the I and Q components. This output is generated per time sample for each range bin and virtual receiver. On the other hand, the Doppler processing unit 601 reads the information stored in RDC1 in a different order. For each range bin and virtual receiver, the Doppler processing unit 601 reads all the time samples, and then performs FFT operations across the time axis to generate velocity estimates.
Dynamic Random Access Memory (DRAM) is a commonly used technology to implement the storage of large blocks of data, such as the radar data cube RDC1. In an exemplary DRAM, each bit is stored in separate tiny capacitors with an integrated circuit, which are all arranged in a rectangular grid, called a bank, of rows and columns. To access (read or write) data, a DRAM bank is first charged, a row is activated, and a specific column or block of data is accessed. Given this access pattern, it is significantly faster to access all bits stored in a row than accessing bits across multiple rows. This performance aspect of the DRAM is combined with locality in memory access patterns to distribute data to maximize reading or writing bits from a row simultaneously.
If RDC1, as computed by the correlation processor 410, is stored in a DRAM and then read by a Doppler processing unit 601, there will be a mismatch in the reading and writing axes order. This can significantly reduce memory performance due to random access, i.e., by not utilizing the row access locality of DRAM. Specifically, the storage of logical addresses to physical locations in the memory can be optimized to benefit from locality in only one of the operations, either read or write. As an example, if an address assignment is such that writing by the correlation processor 410 leads to local accesses, then RDC1 will be laid out in memory by virtual receiver for a given range bin and time. Thus, writing to such a memory from the correlation processor 410 will have high locality and thus high performance. However, the reading operation by the Doppler processor 601 would read samples per time sample for a given range-bin and virtual receiver. Thus, reading operations would have random accesses with no locality, thereby resulting in very low DRAM performance (for the read operations).
In one embodiment, DRAM optimization may be achieved by ordering a writing of the RDC1 output to the DRAM and its subsequent processing by the Doppler processing unit 601 to maximize DRAM throughput. Specifically, RDC1 is written to memory with continuous virtual receivers as the inner loop for each range bin and time sample. That is, for a given first range bin and first time the writing is first done for various virtual receivers. The writing (of RDC1) will include high “burst locality” and thus high DRAM performance. Here burst locality refers to the property of the memory access pattern of reading sequentially—in a burst—multiple data items which are stored in consecutive locations in the DRAM. Such access patterns have high DRAM throughput because the DRAM module can read/write multiple consecutive locations simultaneously into its “row buffer.” On the other hand, while reading data, the Doppler/velocity processing unit needs to read samples per time sample for a given range-bin and virtual receiver. This memory access pattern does not exhibit burst locality, as the data has been written to the DRAM to exhibit locality along the time axis. To modify this memory access pattern, RDC1 may be read by the Doppler processing unit 601 by prefetching data into a local buffer. Specifically, data from multiple virtual receivers are simultaneously pre-fetched to a local buffer. The number of virtual receivers to pre-fetch is such that a resultant burst length is long enough for high DRAM throughput, i.e., it reads data from all elements of the row buffer of the DRAM. With this modification to the memory access pattern, the underlying memory read/write requests by the Doppler/velocity processing unit to the DRAM, high throughput can be maintained both for write and read operations.
In one embodiment, the size of the buffer to be used at the input of the Doppler processing unit 601 is chosen based on the DRAM module's row size. As an example, in a case of 64 virtual receivers, the output of RDC1 will have complex values for 64 virtual receivers for each range bin and time. Of these, the samples for 8 virtual receivers can be pre-fetched in one memory read. Given that each sample in RDC1 has 32 bits of information, a total of 8×32=256 bits are to be read at a time, thereby achieving high DRAM throughput. The pre-fetching buffer should then allow the storage of multiple such pre-fetches for a given range-bin and for all possible time samples. Subsequently, the Doppler/velocity processing would apply FFT operations along the time axis for a given range-bin and virtual receiver.
In a further embodiment, the buffer of the Doppler processing unit 601 can be implemented in a shared memory. In such a setup, the amount of memory available for implementing this buffer can be increased or decreased. Correspondingly, the number of virtual receivers which are pre-fetched in one memory read can be dynamically adjusted based on the available memory.
On-the-Fly Block-Floating-Point Operation:
A block-floating-point (BFP) operation may emulate a floating point while using fixed point hardware. In such an exemplary operation, an entire block of data is assigned a common exponent. Thereby, the necessity of saving individual floating points values is avoided. This creates the potential to optimize the bit-width of different parts of a compute pipeline, e.g. a FFT processing pipeline. Thus, BFP operation is a trade-off between fixed-point and full floating-point architecture.
For a given block of data, a common exponent is found as a data item with a largest amplitude in the block, i.e., the maximum number of leading bits across all data items of the block. In other words, the number of left shifts needed for the data to be normalized to the dynamic range of the hardware is used as the common exponent.
Consider an exemplary FFT operation that is performed with BFP operations. Unlike a true floating-point operation, a BFP FFT block does not provide exponents for individual data items. A complex value integer pair may be represented with one scale factor that is shared through the block among other complex value integer pairs. After every FFT stage, the largest output value is detected and the result is scaled to improve the precision. The number of left or right shifts to be performed in the scaling is recorded in the common exponent.
Specifically, for the receive pipeline, as illustrated in
As discussed herein, BFP operations require identifying a common exponent for all data. This operation is also prohibitive as it requires storing and analyzing the entire block. To avoid this, one embodiment may employ an on-the-fly BFP operation, wherein the conversion to the block exponent happens on-the-fly while reading each sample of RDC1 computed by the correlation operation. In other words, as each sample is read from the RDC1, it is transformed into BFP notation using a local computation. However, different samples may require different adjustments, i.e. left or right shifts, with respect to the common exponent. Such adjustments are written to a table which records which samples must be adjusted and by how many bits. While reading these samples, this table is referred to in order to adjust each sample by the right number of bits.
In one embodiment, the real component of the 512 time samples, computed after correlating the received and delayed transmitted signals, are stored in a skewer of RDC1, with each data point provided as a 24-bit value. Similarly for the imaginary component. These samples are to be stored as 16-bit values in the memory with a block exponent. As discussed herein, the block exponent cannot be computed by first reading all values, but can only be computed on-the-fly. The samples are read in a time sequence. A first sample is read, and a number of leading zeros in the number is identified as being one less than a number of repeating sign bits. For example, if 0 repeats 5 times in this first sample, it indicates 4 leading zeros in the number. This number, i.e. 4, is recorded as the first block exponent value. This recording is done in a separate table called the exponent normalization table. The first sample is then adjusted as per the block exponent value, i.e., the value is left-shifted by 4 bits and optimized to be sent as a 16-bit value, by dropping all less significant bits that do not fit. Then the next sample value in sequence is read. If the absolute value is less than that of the first sample value, i.e. if it has equal or more leading zeros, then the same block exponent is used and the number is left-shifted by the same number of bits, which is 4 bits in the given example. Similarly, as long as sample values have fewer leading zeros, the same block exponent is used and the number is left-shifted by the same number of bits, which is 4 bits in the given example.
However, any sample value which has fewer leading zeros, i.e., whose absolute value is higher than that of the previous value at which the block exponent was calculated, two things are done. First, the number of leading zeros in the current number is calculated, and then an entry is made in the exponent normalization table. For example, if the 102nd sample has only 3 leading zeros, then the exponent normalization table notes that the block exponent is changed to 3. This process is continued through all samples, which is 512 in this example. In a further example, if the 255th value has only 2 leading zeros, then the exponent normalization table notes the index 255 and the corresponding change in the block exponent. Let all subsequent samples have 2 or more leading zeros. Then, the generated exponent normalization table looks like:
Applying the above described on-the-fly block floating operation, the output of the RDC1 is realized as 16-bit values of samples stored in the memory, and in addition an exponent normalization table like the one illustrated above. Because the bit-width is reducing from 24 to 16 bits, there can be at most 8 adjustments made in the exponent normalization table, which will record the corresponding index values where the adjustments happen. As illustrated above, the exponent normalization table also includes an entry for the total number of adjustments.
The Doppler processing unit 601 reads the stored 16-bit values along with the exponent table. These values are to be parsed with the exponent normalization table by making appropriate adjustments. First, the exponent normalization table is read to identify the maximum exponent. In the example above, the maximum exponent is −2, as the sample at index 255 had that as the block exponent. This value is set as the common exponent for this entire block. However, all values which were left-shifted by a greater number of bits than the maximum exponent will need to be right-shifted by the appropriate number of bits. This may be done by sequentially traversing the exponent normalization table and adjusting the number of bit shifts to be performed. In the example considered, the following operations are performed:
Samples 0-101: shift incoming values by 2 bits to the right.
Samples 102-254: shift incoming values by 1 bit to the right.
Samples 255-511: no changes required.
After the above processing, the Doppler processing unit 601 will receive a stream of numbers with a single common exponent. This on-the-fly computation of block exponent and subsequent adjustment can be done in a pipelined manner with intermediate buffer storage.
In one embodiment, different parts of the RDC1 are treated as separate blocks with multiple exponent-adjustment tables marking the transitions in the adjustments to be made. RDC1 data may be read on a per range-bin or virtual receiver basis to convert time-samples into Doppler samples. In one embodiment, a different block exponent could be assigned to each range-bin, and thus there are as many adjustment tables as there are range-bins.
Pipelined Implementation:
In one exemplary embodiment, to meet high throughput design constraints while limiting memory and transfer constraints, different modules of the receive pipeline may be configured to execute concurrently interfaced with buffers which can be read from and written to simultaneously. Specifically, all the following operations can be pipelined: (a) estimating range by correlating received signals with delayed copies of transmitted signals, (b) estimating velocity with Doppler FFT operations, (c) estimating angles with rough beamforming algorithms, (d) channelizing to trade-off Doppler resolution for multiple looks used in super-resolution angle estimation, (e) extracting a static image slice from the output of the angle processing, and (f) sparsifying the RDC2 and RDC3 radar data cubes based on the static image and noise floor estimates. Between the different modules that execute these operations, intermediate buffers may be used to store partial results to enable concurrent processing.
Specifically, the radar data cubes RDC2 and RDC3 can be stored in first-in, first-out (FIFO) buffers where all the information associated with a particular scan (e.g., range, Doppler, and virtual receiver) are not necessarily stored simultaneously but are being written into the FIFO while other information is being read out. Thus, while the Doppler and angle processors (601, 602) write parts of the RDC2 and RDC3 to the buffers, the histogram engine 710 and thresholding engine 712 can read other parts of the cubes. As the information passes through the FIFO buffers, the static image extraction unit 711 tags voxels which belong to the static image. For each particular untagged voxel sample in the RDC3, the sparsifier (603) compares the magnitude of the sample to one or more thresholds configured by the control processor. Depending on the outcome of the comparison, a set of samples, including the sample, are stored in one or more distinct memory buffers. Preferably, the set of samples comprise samples with a same range or Doppler/velocity as a particular voxel sample. The outputs—including the static image and the sparsified cubes sRDC2 and sRDC3—can also be written to FIFO buffers.
With buffers between computing modules, the execution of these modules may be pipelined, thereby increasing the throughput. Further, as data is processed on-the-fly, there are reduced storage requirements. However, optimization is necessary to size the buffers depending on the throughput of the individual modules and the sizes of the outputs from the respective processing modules. Dataflow analysis may be performed to estimate the sizes of each of the buffers to avoid stalling any of the modules. In a preferred embodiment, the buffers could be stored in a shared memory where their sizes can be dynamically changed at runtime by passing through appropriate pointers across the pipeline.
Functional Correctness:
Functional safety is an important aspect of automotive devices. It would be advantageous if an exemplary radar device as part of an automotive system was able to self-diagnose any potential issues, however, rare. This applies to different parts of the radar system, including the receive pipeline (401) illustrated in
In one embodiment, a CRC engine 713, illustrated in
In one embodiment, the CRC engine 713 could be upgraded to an exemplary error correction code (ECC) engine that can be used to compute parity checks for an error correcting code, also called a forward error correction code. The parity checks can be used to correct errors. More parity checks will allow for the correction of more bits. ECC codes can also do a combination of error correction and error detection. Codes capable of single error correcting and double error detecting are known as SEC-DED codes. Error-correcting codes append check bits to every data word to increase the minimum Hamming distance between different valid words. Here, the Hamming distance between any two binary numbers is the number of bits at which the two numbers differ. Adding a single parity check bit, which causes the minimum Hamming distance of 2 between valid codewords, may be used to detect single bit errors. Multiple check bits can be added to increase the Hamming distance to 3 to correct single bit errors. For a Hamming distance of 4, a single-bit error can be corrected, and a two-bit error can be detected, though not corrected. ECC can be implemented using Hamming codes which is a family of perfect linear block codes, and other codes including BCH codes and Reed-Solomon codes. Like the CRC engine (713), an exemplary ECC could be applied to the RDC1, RDC2 cubes and the static image slices, either before or after the sparsification. When these memory blocks are read for subsequent processing, when using an SEC-DED code, single bit errors can be corrected and double errors detected. Further, if two bit errors are detected, then an interrupt can be generated which can be used to trigger a software-based corrective action.
Memory Interface:
An exemplary memory interface or control unit 730, illustrated in
Optionally, the memory interface unit 730 can store other data structures to the memory, as configured by parameters set by a control processor. These additional data structures can include the histogram used to estimate the noise floor and the bit map of angles where the signal amplitude was above the configured noise floor threshold.
Changes and modifications in the specifically described embodiments can be carried out without departing from the principles of the present invention which is intended to be limited only by the scope of the appended claims, as interpreted according to the principles of patent law including the doctrine of equivalents.
The present application claims the filing benefits of U.S. provisional application, Ser. No. 62/457,394, filed Feb. 10, 2017, which is hereby incorporated by reference herein in its entirety.
Number | Name | Date | Kind |
---|---|---|---|
1882128 | Fearing | Oct 1932 | A |
3374478 | Blau | Mar 1968 | A |
3735398 | Ross | May 1973 | A |
3750169 | Strenglein | Jul 1973 | A |
3896434 | Sirven | Jul 1975 | A |
4078234 | Fishbein et al. | Mar 1978 | A |
4176351 | De Vita et al. | Nov 1979 | A |
4566010 | Collins | Jan 1986 | A |
4882668 | Schmid et al. | Nov 1989 | A |
4910464 | Trett et al. | Mar 1990 | A |
4939685 | Feintuch | Jul 1990 | A |
5001486 | Bächtiger | Mar 1991 | A |
5034906 | Chang | Jul 1991 | A |
5087918 | May et al. | Feb 1992 | A |
5151702 | Urkowitz | Sep 1992 | A |
5175710 | Hutson | Dec 1992 | A |
5218619 | Dent | Jun 1993 | A |
5272663 | Jones et al. | Dec 1993 | A |
5280288 | Sherry et al. | Jan 1994 | A |
5302956 | Asbury et al. | Apr 1994 | A |
5341141 | Frazier et al. | Aug 1994 | A |
5345470 | Alexander | Sep 1994 | A |
5379322 | Kosaka et al. | Jan 1995 | A |
5497162 | Kaiser | Mar 1996 | A |
5508706 | Tsou et al. | Apr 1996 | A |
5581464 | Woll et al. | Dec 1996 | A |
5657021 | Ehsani-Nategh et al. | Aug 1997 | A |
5657023 | Lewis et al. | Aug 1997 | A |
5691724 | Aker et al. | Nov 1997 | A |
5712640 | Andou | Jan 1998 | A |
5724041 | Inoue et al. | Mar 1998 | A |
5892477 | Wehling | Apr 1999 | A |
5917430 | Greneker, III et al. | Jun 1999 | A |
5920285 | Benjamin | Jul 1999 | A |
5931893 | Dent et al. | Aug 1999 | A |
5959571 | Aoyagi et al. | Sep 1999 | A |
5970400 | Dwyer | Oct 1999 | A |
6067314 | Azuma | May 2000 | A |
6069581 | Bell et al. | May 2000 | A |
6121872 | Weishaupt | Sep 2000 | A |
6121918 | Tullsson | Sep 2000 | A |
6151366 | Yip | Nov 2000 | A |
6163252 | Nishiwaki | Dec 2000 | A |
6184829 | Stilp | Feb 2001 | B1 |
6191726 | Tullsson | Feb 2001 | B1 |
6288672 | Asano et al. | Sep 2001 | B1 |
6307622 | Lewis | Oct 2001 | B1 |
6347264 | Nicosia et al. | Feb 2002 | B2 |
6400308 | Bell et al. | Jun 2002 | B1 |
6411250 | Oswald et al. | Jun 2002 | B1 |
6417796 | Bowlds | Jul 2002 | B1 |
6424289 | Fukae et al. | Jul 2002 | B2 |
6583753 | Reed | Jun 2003 | B1 |
6614387 | Deadman | Sep 2003 | B1 |
6624784 | Yamaguchi | Sep 2003 | B1 |
6674908 | Aronov | Jan 2004 | B1 |
6714956 | Liu et al. | Mar 2004 | B1 |
6747595 | Hirabe | Jun 2004 | B2 |
6768391 | Dent et al. | Jul 2004 | B1 |
6865218 | Sourour | Mar 2005 | B1 |
6975246 | Trudeau | Dec 2005 | B1 |
7119739 | Struckman | Oct 2006 | B1 |
7289058 | Shima | Oct 2007 | B2 |
7299251 | Skidmore et al. | Nov 2007 | B2 |
7338450 | Kristofferson et al. | Mar 2008 | B2 |
7395084 | Anttila | Jul 2008 | B2 |
7460055 | Nishijima et al. | Dec 2008 | B2 |
7474258 | Arikan et al. | Jan 2009 | B1 |
7545310 | Matsuoka | Jun 2009 | B2 |
7545321 | Kawasaki | Jun 2009 | B2 |
7564400 | Fukuda | Jul 2009 | B2 |
7567204 | Sakamoto | Jul 2009 | B2 |
7609198 | Chang | Oct 2009 | B2 |
7642952 | Fukuda | Jan 2010 | B2 |
7663533 | Toennesen | Feb 2010 | B2 |
7728762 | Sakamoto | Jun 2010 | B2 |
7791528 | Klotzbuecher | Sep 2010 | B2 |
7847731 | Wiesbeck et al. | Dec 2010 | B2 |
7855677 | Negoro et al. | Dec 2010 | B2 |
7859450 | Shirakawa et al. | Dec 2010 | B2 |
8019352 | Rappaport et al. | Sep 2011 | B2 |
8049663 | Frank et al. | Nov 2011 | B2 |
8059026 | Nunez | Nov 2011 | B1 |
8102306 | Smith, Jr. et al. | Jan 2012 | B2 |
8154436 | Szajnowski | Apr 2012 | B2 |
8330650 | Goldman | Dec 2012 | B2 |
8390507 | Wintermantel | Mar 2013 | B2 |
8471760 | Szajnowski | Jun 2013 | B2 |
8532159 | Kagawa et al. | Sep 2013 | B2 |
8547988 | Hadani et al. | Oct 2013 | B2 |
8686894 | Fukuda et al. | Apr 2014 | B2 |
8694306 | Short et al. | Apr 2014 | B1 |
9121943 | Stirlin-Gallacher et al. | Sep 2015 | B2 |
9239378 | Kishigami et al. | Jan 2016 | B2 |
9239379 | Burgio et al. | Jan 2016 | B2 |
9282945 | Smith et al. | Mar 2016 | B2 |
9335402 | Maeno et al. | May 2016 | B2 |
9541639 | Searcy et al. | Jan 2017 | B2 |
9568600 | Alland | Feb 2017 | B2 |
9575160 | Davis et al. | Feb 2017 | B1 |
9599702 | Bordes et al. | Mar 2017 | B1 |
9689967 | Stark et al. | Jun 2017 | B1 |
9720073 | Davis et al. | Aug 2017 | B1 |
9753121 | Davis | Sep 2017 | B1 |
9753132 | Bordes et al. | Sep 2017 | B1 |
9772397 | Bordes et al. | Sep 2017 | B1 |
9791551 | Eshraghi et al. | Oct 2017 | B1 |
9791564 | Harris et al. | Oct 2017 | B1 |
9806914 | Bordes et al. | Oct 2017 | B1 |
9829567 | Davis et al. | Nov 2017 | B1 |
9846228 | Davis et al. | Dec 2017 | B2 |
9869762 | Alland et al. | Jan 2018 | B1 |
20010002919 | Sourour et al. | Jun 2001 | A1 |
20020004692 | Nicosia et al. | Jan 2002 | A1 |
20020044082 | Woodington et al. | Apr 2002 | A1 |
20020075178 | Woodington et al. | Jun 2002 | A1 |
20020118522 | Ho et al. | Aug 2002 | A1 |
20020130811 | Voigtaender | Sep 2002 | A1 |
20020147534 | Delcheccolo et al. | Oct 2002 | A1 |
20020155811 | Prismantas | Oct 2002 | A1 |
20030001772 | Woodington et al. | Jan 2003 | A1 |
20030011519 | Breglia et al. | Jan 2003 | A1 |
20030058166 | Hirabe | Mar 2003 | A1 |
20030080713 | Kirmuss | May 2003 | A1 |
20030102997 | Levin et al. | Jun 2003 | A1 |
20030235244 | Pessoa et al. | Dec 2003 | A1 |
20040012516 | Schiffmann | Jan 2004 | A1 |
20040015529 | Tanrikulu et al. | Jan 2004 | A1 |
20040066323 | Richter | Apr 2004 | A1 |
20040138802 | Kuragaki et al. | Jul 2004 | A1 |
20050069162 | Haykin | Mar 2005 | A1 |
20050156780 | Bonthron et al. | Jul 2005 | A1 |
20050201457 | Allred et al. | Sep 2005 | A1 |
20050225476 | Hoetzel et al. | Oct 2005 | A1 |
20060012511 | Dooi et al. | Jan 2006 | A1 |
20060036353 | Wintermantel | Feb 2006 | A1 |
20060050707 | Sterin | Mar 2006 | A1 |
20060093078 | Lewis et al. | May 2006 | A1 |
20060109170 | Voigtlaender et al. | May 2006 | A1 |
20060109931 | Asai | May 2006 | A1 |
20060114324 | Farmer et al. | Jun 2006 | A1 |
20060140249 | Kohno | Jun 2006 | A1 |
20060181448 | Natsume et al. | Aug 2006 | A1 |
20060244653 | Szajnowski | Nov 2006 | A1 |
20060262007 | Bonthron | Nov 2006 | A1 |
20060262009 | Watanabe | Nov 2006 | A1 |
20070018884 | Adams | Jan 2007 | A1 |
20070018886 | Watanabe et al. | Jan 2007 | A1 |
20070109175 | Fukuda | May 2007 | A1 |
20070120731 | Kelly, Jr. et al. | May 2007 | A1 |
20070132633 | Uchino | Jun 2007 | A1 |
20070152870 | Woodington et al. | Jul 2007 | A1 |
20070152871 | Puglia | Jul 2007 | A1 |
20070152872 | Woodington | Jul 2007 | A1 |
20070164896 | Suzuki et al. | Jul 2007 | A1 |
20070171122 | Nakano | Jul 2007 | A1 |
20070182619 | Honda et al. | Aug 2007 | A1 |
20070182623 | Zeng | Aug 2007 | A1 |
20070188373 | Shirakawa et al. | Aug 2007 | A1 |
20070200747 | Okai | Aug 2007 | A1 |
20070279303 | Schoebel | Dec 2007 | A1 |
20080208472 | Morcom | Aug 2008 | A1 |
20080272955 | Yonak et al. | Nov 2008 | A1 |
20090015459 | Mahler et al. | Jan 2009 | A1 |
20090015464 | Fukuda | Jan 2009 | A1 |
20090051581 | Hatono | Feb 2009 | A1 |
20090072957 | Wu et al. | Mar 2009 | A1 |
20090073025 | Inoue et al. | Mar 2009 | A1 |
20090074031 | Fukuda | Mar 2009 | A1 |
20090079617 | Shirakawa et al. | Mar 2009 | A1 |
20090085827 | Orime et al. | Apr 2009 | A1 |
20090103593 | Bergamo | Apr 2009 | A1 |
20090121918 | Shirai et al. | May 2009 | A1 |
20090212998 | Szajnowski | Aug 2009 | A1 |
20090237293 | Sakuma | Sep 2009 | A1 |
20090267822 | Shinoda et al. | Oct 2009 | A1 |
20090289831 | Akita | Nov 2009 | A1 |
20090295623 | Falk | Dec 2009 | A1 |
20100019950 | Yamano et al. | Jan 2010 | A1 |
20100116365 | McCarty | May 2010 | A1 |
20100156690 | Kim et al. | Jun 2010 | A1 |
20100198513 | Zeng et al. | Aug 2010 | A1 |
20100277359 | Ando | Nov 2010 | A1 |
20100289692 | Winkler | Nov 2010 | A1 |
20110006944 | Goldman | Jan 2011 | A1 |
20110032138 | Krapf | Feb 2011 | A1 |
20110074620 | Wintermantel | Mar 2011 | A1 |
20110187600 | Landt | Aug 2011 | A1 |
20110196568 | Nickolaou | Aug 2011 | A1 |
20110248796 | Pozgay | Oct 2011 | A1 |
20110279303 | Smith, Jr. et al. | Nov 2011 | A1 |
20110279307 | Song | Nov 2011 | A1 |
20110285576 | Lynam | Nov 2011 | A1 |
20110291874 | De Mersseman | Dec 2011 | A1 |
20110291875 | Szajnowski | Dec 2011 | A1 |
20110292971 | Hadani et al. | Dec 2011 | A1 |
20110298653 | Mizutani | Dec 2011 | A1 |
20120001791 | Wintermantel | Jan 2012 | A1 |
20120050093 | Heilmann et al. | Mar 2012 | A1 |
20120105268 | Smits et al. | May 2012 | A1 |
20120112957 | Nguyen et al. | May 2012 | A1 |
20120133547 | MacDonald et al. | May 2012 | A1 |
20120173246 | Choi et al. | Jul 2012 | A1 |
20120195349 | Lakkis | Aug 2012 | A1 |
20120257643 | Wu et al. | Oct 2012 | A1 |
20120319900 | Johansson et al. | Dec 2012 | A1 |
20130016761 | Nentwig | Jan 2013 | A1 |
20130021196 | Himmelstoss | Jan 2013 | A1 |
20130027240 | Chowdhury | Jan 2013 | A1 |
20130069818 | Shirakawa et al. | Mar 2013 | A1 |
20130102254 | Cyzs | Apr 2013 | A1 |
20130113653 | Kishigami et al. | May 2013 | A1 |
20130135140 | Kishigami | May 2013 | A1 |
20130169485 | Lynch | Jul 2013 | A1 |
20130176154 | Bonaccio et al. | Jul 2013 | A1 |
20130214961 | Lee et al. | Aug 2013 | A1 |
20130229301 | Kanamoto | Sep 2013 | A1 |
20130244710 | Nguyen et al. | Sep 2013 | A1 |
20130314271 | Braswell et al. | Nov 2013 | A1 |
20130321196 | Binzer et al. | Dec 2013 | A1 |
20140022108 | Alberth, Jr. et al. | Jan 2014 | A1 |
20140028491 | Ferguson | Jan 2014 | A1 |
20140035774 | Khlifi | Feb 2014 | A1 |
20140070985 | Vacanti | Mar 2014 | A1 |
20140085128 | Kishigami et al. | Mar 2014 | A1 |
20140111372 | Wu | Apr 2014 | A1 |
20140139322 | Wang et al. | May 2014 | A1 |
20140159948 | Ishimori et al. | Jun 2014 | A1 |
20140220903 | Schulz et al. | Aug 2014 | A1 |
20140253345 | Breed | Sep 2014 | A1 |
20140285373 | Kuwahara et al. | Sep 2014 | A1 |
20140327566 | Burgio et al. | Nov 2014 | A1 |
20140348253 | Mobasher et al. | Nov 2014 | A1 |
20150002329 | Murad et al. | Jan 2015 | A1 |
20150002357 | Sanford et al. | Jan 2015 | A1 |
20150035662 | Bowers et al. | Feb 2015 | A1 |
20150061922 | Kishigami | Mar 2015 | A1 |
20150103745 | Negus et al. | Apr 2015 | A1 |
20150198709 | Inoue | Jul 2015 | A1 |
20150204966 | Kishigami | Jul 2015 | A1 |
20150204971 | Kuehnle | Jul 2015 | A1 |
20150226848 | Park | Aug 2015 | A1 |
20150234045 | Rosenblum | Aug 2015 | A1 |
20150247924 | Kishigami | Sep 2015 | A1 |
20150255867 | Inoue | Sep 2015 | A1 |
20150301172 | Ossowska | Oct 2015 | A1 |
20150323660 | Hampikian | Nov 2015 | A1 |
20150331090 | Jeong et al. | Nov 2015 | A1 |
20160003939 | Stainvas Olshansky et al. | Jan 2016 | A1 |
20160018511 | Nayyar et al. | Jan 2016 | A1 |
20160033631 | Searcy et al. | Feb 2016 | A1 |
20160033632 | Searcy et al. | Feb 2016 | A1 |
20160041260 | Cao et al. | Feb 2016 | A1 |
20160061935 | McCloskey et al. | Mar 2016 | A1 |
20160084941 | Arage | Mar 2016 | A1 |
20160084943 | Arage | Mar 2016 | A1 |
20160091595 | Alcalde | Mar 2016 | A1 |
20160124086 | Jansen et al. | May 2016 | A1 |
20160139254 | Wittenberg | May 2016 | A1 |
20160146931 | Rao et al. | May 2016 | A1 |
20160154103 | Moriuchi | Jun 2016 | A1 |
20160238694 | Kishigami | Aug 2016 | A1 |
20170010361 | Tanaka | Jan 2017 | A1 |
20170023661 | Richert | Jan 2017 | A1 |
20170023663 | Subburaj et al. | Jan 2017 | A1 |
20170074980 | Adib | Mar 2017 | A1 |
20170153316 | Wintermantel | Jun 2017 | A1 |
20170219689 | Hung et al. | Aug 2017 | A1 |
20170234968 | Roger et al. | Aug 2017 | A1 |
20170293025 | Davis et al. | Oct 2017 | A1 |
20170293027 | Stark et al. | Oct 2017 | A1 |
20170307728 | Eshraghi et al. | Oct 2017 | A1 |
20170309997 | Alland et al. | Oct 2017 | A1 |
20170310758 | Davis et al. | Oct 2017 | A1 |
20170336495 | Davis et al. | Nov 2017 | A1 |
Number | Date | Country |
---|---|---|
0725480 | Nov 2011 | EP |
2374217 | Apr 2013 | EP |
2821808 | Jul 2015 | EP |
2751086 | Jan 1998 | FR |
WO2015175078 | Nov 2015 | WO |
WO2015185058 | Dec 2015 | WO |
WO2016011407 | Jan 2016 | WO |
WO2016030656 | Mar 2016 | WO |
WO2017187330 | Nov 2017 | WO |
Entry |
---|
International Search Report of the International Searching Authority from corresponding Patent Cooperation Treaty (PCT) Application No. PCT/IB18/50819, indicated completed on May 23, 2018. |
Written Opinion of the International Searching Authority from corresponding Patent Cooperation Treaty (PCT) Application No. PCT/IB18/50819, indicated completed on May 23, 2018. |
Chambers et al., “An article entitled Real-Time Vehicle Mounted Multistatic Ground Penetrating Radar Imaging System for Buried Object Detection,” Lawrence Livermore National Laboratory Reports (LLNL-TR-615452), Feb. 4, 2013; Retrieved from the Internet from https://e-reports-ext.llnl.gov/pdf/711892.pdf. |
Fraser, “Design and simulation of a coded sequence ground penetrating radar,” In: Diss. University of British Columbia, Dec. 3, 2015. |
Zhou et al., “Linear extractors for extracting randomness from noisy sources,” In: Information Theory Proceedings (ISIT), 2011 IEEE International Symposium on Oct. 3, 2011. |
V. Giannini et al., “A 79 GHz Phase-Modulated 4 GHz-BW CW Radar Transmitter in 28 nm CMOS, ”in IEEE Journal of Solid-State Circuits, vol. 49, No. 12, pp. 2925-2937, Dec. 2014. (Year: 2014). |
Number | Date | Country | |
---|---|---|---|
20180231636 A1 | Aug 2018 | US |
Number | Date | Country | |
---|---|---|---|
62457394 | Feb 2017 | US |