CN116908794A - Distributed fitting Constant False Alarm Rate (CFAR) detection - Google Patents

Abstract

Distribution fitting Constant False Alarm Rate (CFAR) detection is described. Noise data in cells or bins surrounding the target cell are fitted to a noise distribution model, such as a Rayleigh distribution model. With an appropriate noise distribution curve from the distribution model, the CFAR threshold of the cell along the curve can be determined. The quantile function of the noise distribution model of a bin or unit provides the CFAR threshold for that bin or unit use. The distributed fit CFAR enables more accurate CFAR thresholds to be set for each bin or cell and may use significantly less computational resources than the ordered statistical CFAR. The radar detector may better prevent false alarm detection in a plurality of different driving scenarios by adapting to different environments and dynamically changing the noise profile used based on a best fit analysis of the noise profile model of the noise characteristics of neighboring bins or cells.

Description

Distributed fitting Constant False Alarm Rate (CFAR) detection

Background

A perception system of a vehicle (e.g., an advanced safety or autopilot system) may rely on a radar system to detect objects present in a driving scene (e.g., on a road). There may be interference caused by unknown sources in other vehicle systems or environments; if not processed, such noise may be manifested as false detection of an actually non-existent target reported by the radar system, or actual detection of a target not detected by the radar system. There are several ways in which the radar false alarm rate can be reduced. One way includes performing constant false alarm rate (Constant False Alarm Rate, CFAR) processing techniques in which a power threshold applied to radar detection is adjusted to be above an estimated noise level of an operating environment. CFAR allows reporting echoes that may originate from actual targets and suppressing echoes from other sources in the environment. In an ideal case, noise occurs at a known inferred level, however, in reality, ambient noise can appear as a highly achievable and unpredictable signal that is non-uniform in time and space, which makes adjusting CFAR power thresholds for changing driving scenarios a challenge.

Disclosure of Invention

This document describes techniques and systems for distributed fit CFAR detection. This summary is not intended to identify essential features of the claimed subject matter, nor is it intended to be used to determine the scope of the claimed subject matter.

In one example, a system includes a processor configured to obtain a plurality of samples of radar echoes including noise reflected from an environment external to a vehicle. The processor is further configured to: maintaining an array of samples (e.g., a data cube) that includes each of the samples in a different corresponding cell of the array; and for each of the samples, determining a respective Constant False Alarm Rate (CFAR) threshold for filtering noise from a corresponding cell of the array. The processor is configured to determine the CFAR threshold by: determining a set of neighboring cells of the corresponding cell to use as training cells; generating a histogram that organizes the samples of the training unit into columns representing a continuous range of magnitudes of the samples of the training unit; fitting the histogram to a noise distribution function; and determining a respective CFAR threshold for the corresponding cell from the noise distribution function fitted to the histogram of the corresponding cell. The processor is further configured to: filtering samples from the array that do not meet the respective CFAR threshold for the corresponding cell; and in some examples, in response to filtering the samples from the array, outputting the array for use by the vehicle function in detecting objects present in the environment.

In addition to this and other examples, methods for performing distribution fitting CFAR detection are also described. In some cases, a Computer Readable Medium (CRM) includes instructions that, when executed by a processor, configure the processor to perform the methods. The present disclosure also contemplates systems that include means for performing these methods. By implementing these and other examples contemplated by the present disclosure, distributed fit CFAR detection may be implemented to more accurately discern an actual target from noise in a radar signal than using other CFAR detection techniques.

Drawings

Details of the distribution fitting CFAR detection are described in this document with reference to the accompanying drawings, which may use the same numbers to reference like features and components, and hyphenated numbers to designate variations of these like features and components. The attached drawings are organized as follows:

FIG. 1 illustrates a conceptual diagram of an example environment for performing distributed fit CFAR detection in accordance with the described techniques;

FIG. 2 illustrates a conceptual diagram of an example noise estimator of a radar system configured to perform distributed fit CFAR detection in accordance with the described techniques;

FIG. 3 illustrates a flow chart of an example process for performing distributed fit CFAR detection in accordance with the described techniques;

FIG. 4 shows a line graph comparing the performance of a distribution fit CFAR detection with ground truth (ground true);

FIG. 5 shows a line graph comparing the performance of a distribution fit CFAR test with an ordered statistics CFAR test; and

fig. 6 shows a line graph comparing the performance of a distribution fit CFAR test with a unit average CFAR test.

Detailed Description

SUMMARY

The vehicle may include a multiple-input multiple-output (MIMO) radar system to generate a three-dimensional data cube of radar results obtained by processing multiple samples of individual radar chirp transmitted over multiple channels. For each viewing period, the new data cube may be stored in memory as a one-dimensional array; each cell is individually addressable using a unique combination of parameters indicating distance (range) bin, chirp identification, and channel. The distance processing stage fills the data cube with continuously chirped distance Fast Fourier Transform (FFT) results by addressing rows of cells. By addressing the columns of cells, the Doppler process retrieves the range FFT results of successive ranges.

In radar systems, effective detection in a cell or bin of a data cube may be surrounded by ambient noise. The ambient noise in the environment is constantly changing. False alarms or false detections may be reported when the ambient noise (e.g., its power) is at a level that masks valid detections. In order to improve the accuracy of target detection and the functionality of the vehicle system that relies on radar data, the radar is set to operate at a low false alarm rate. The threshold may be used to separate noise from valid detection. The threshold may be adaptively set using a set of techniques to maintain Constant False Alarm Rates (CFARs) for various scenarios. CFAR processing techniques may be applied before or after range processing or doppler processing occurs. Due to their simplicity and robustness, two of the most common CFAR processing techniques include unit average CFAR and ordered statistics CFAR. In both unit average CFAR and ordered statistics CFAR, the radar system checks for the presence of an actual target based on noise power estimates obtained from neighboring (e.g., preceding and lagging) units or range bins.

The cell average CFAR sets the detection threshold as an average (e.g., average) of the power estimates (or factors thereof) of the neighboring cells. If the unit exceeds the average power detection threshold, the unit records target detection. When two conditions are satisfied, the unit average CFAR may have accurate and stable performance. The first condition is that each real object is isolated from the other real objects (e.g., each object is an independent object), otherwise the unit average CFAR would provide inaccurate results. Two targets are not independent if they are consecutively aligned with each other in the range or doppler domain. Second, it is assumed that neighboring cells sampled for estimating noise power are independent and identically distributed. When adjacent samples are affected by noise, such effects may cause inaccurate deviations in the average calculation.

Ordered statistics CFAR the unit average CFAR is improved to address common multi-objective scenarios. The ordered statistics CFAR also analyzes noise power from neighboring bins. However, rather than taking into account the average value of the noise power, the power magnitudes at each neighboring cell or bin are ordered in sequence. The cell with the highest power is selected and the noise threshold is set accordingly. Unlike the unit average CFAR, the ordered statistics CFAR is not conditioned on the homogeneous clutter of each independent target; however, an excessively high false alarm rate may occur at the edges of the clutter. Adjusting which cells are included among the neighboring cells may have varying degrees of impact on performance. The number of adjacent cells selected for ordering may depend to a large extent on driving conditions and noise sources in the environment. The environment in vehicle applications is constantly changing; adjusting the adjacent sample size of the ordered statistics CFAR to accommodate fluctuating conditions can be challenging. The increased accuracy of the ordered statistics CFAR compared to the unit average CFAR may be offset by an increase in its processing complexity (e.g. ranking function) which requires faster and often more expensive computational resources, which limits the use of ordered statistics CFAR to more expensive radar applications.

The main drawback of widely used CFAR techniques is that they cannot adapt to estimate the noise threshold of different measurement environments. For example, in automotive radar applications, a host vehicle may be traveling on an open street for a while and then quickly switch to traveling on a highway with guardrails on one or both sides. The noise or clutter models used to estimate noise in these two different scenarios should behave differently to account for the presence or absence of guardrails. However, neither the ordered statistical CFAR nor the unit average CFAR provides a noise model that can obtain correct results for both cases; instead, one is better than the other.

In contrast to existing CFAR techniques, this document describes a distributed fitting CFAR technique for radar detection. Noise data in cells or bins surrounding the target cell are fitted to a noise distribution model, such as a Rayleigh distribution model. With the appropriate noise distribution curve obtained from the distribution model, the CFAR threshold of the cell along the curve can be determined. The quantile function of the noise distribution model of a bin or unit provides the CFAR threshold for that bin or unit use. The distributed fit CFAR enables more accurate CFAR thresholds to be set for each bin or cell and may use significantly less computational resources than the ordered statistical CFAR. The radar detector may better prevent false alarm detection in a plurality of different driving scenarios by adapting to different environments and dynamically changing the noise profile used based on a best fit analysis of the noise profile model of the noise characteristics of neighboring bins or cells.

Example Environment

FIG. 1 illustrates a conceptual diagram of an example environment 100 for performing distributed fit CFAR detection in accordance with the described techniques. The environment 100 includes a vehicle 102, the vehicle 102 including a radar system 104. The radar system 104 enables other systems of the vehicle 102 (not shown in the figures for simplicity) to detect the object 108, which object 108 may influence how or whether the vehicle 102 may continue to travel.

The depicted environment 100 includes a vehicle 102 traveling on a roadway. Although shown as a passenger vehicle, the vehicle 102 may represent other types of motor vehicles (e.g., automobiles, motorcycles, buses, tractors, semi-trailers), non-motor vehicles (e.g., bicycles), rail vehicles (e.g., trains), watercraft (e.g., ships), aircraft (e.g., airplanes), spacecraft (e.g., satellites), and the like.

The radar system 104 has a region of interest associated with the radar system 104 that at least partially surrounds the vehicle 102. This region of interest is referred to as the field of view 106 (also referred to as the instrumented field of view). The radar system 104 may transmit radar signals 110-1 into the field of view 106 and process radar returns 110-2 reflected from the environment 100 to determine a position, angle, rate of change of distance, or other characteristic of the position and orientation of the object 108 relative to the vehicle 102. Careful selection and/or positioning of the components of the radar system 104 may cause the field of view 106 to have a particular shape or size. The components of radar system 104 may be mounted on any portion of vehicle 102, mounted to any portion of vehicle 102, or integrated with any portion of vehicle 102, such as a front, rear, top, bottom, or side of vehicle 102, a bumper, a side view mirror, a portion of a head light and/or a tail light, or at any other internal or external location of vehicle 102.

As previously described, the vehicle 102 includes other vehicle systems that are operatively and/or communicatively coupled to the radar system 104 using wired and/or wireless links that serve as interconnections, paths, or buses for communication between vehicle components. These other vehicle systems use the output from radar system 104 to perform vehicle-based functions, which may include functions for vehicle control, among other functions. Any conceivable device, apparatus, component, module, part, subsystem, routine, circuit, processor, controller, etc. may be configured as a vehicle system that operates on behalf of the vehicle 102 using radar data. As some non-limiting examples, other vehicle systems may include systems for autonomous control, systems for security, systems for positioning, systems for vehicle-to-vehicle communication, systems for interfacing with occupants, and systems for interfacing with radar or multi-sensor trackers.

The radar system 104 includes a Monolithic Microwave Integrated Circuit (MMIC) 112, a processor 114, and a Computer Readable Medium (CRM) 116. The processor 114 is operatively coupled to an interface of a multiple-input multiple-output (MIMO) array (not shown) through the MMIC 112. The MMIC 112, processor 114, and/or CRM 116 may be operatively and/or communicatively coupled via a wired or wireless link (not shown) and may be part of a radar chip (which may be referred to as a system-on-a-chip). Other devices, antennas, and other radar components may be used by radar system 104. The radar system 104 includes an antenna array, such as a multiple-input multiple-output (MIMO) array, capable of transmitting multiple chirps across a frequency range on multiple channels.

MMIC 112 accumulates radar data from the MIMO array on behalf of processor 114. The radar data includes information about the position and movement of objects in the field of view 106, such as the location and range rate of radar detection reflected from the objects 108. MMIC 112 receives instructions from processor 114 to indicate characteristics (e.g., timing, phase, frequency range, channel) of radar signal 110-1 and its corresponding reflection (i.e., radar echo 110-2). MMIC 112 causes radar signal 110-1 to be transmitted via the MIMO array and into environment 100, and then causes radar echo 110-2 to be detected and received.

The processor 114 processes the radar data generated by the MMIC 112 and outputs the processed radar data as a data structure (e.g., a one-dimensional array; a multi-dimensional array) available to other vehicle systems of the vehicle 102. The data cube 118 is an example of processed radar data generated by the processor 114 from radar data obtained by the MMIC 112. In accordance with the techniques of this disclosure, data cube 118 is generated by performing a distribution fit CFAR detection. Any single or multi-dimensional data structure may be used; data cube 118 is just one example of a suitable format for communicating information about radar returns 110-2 for use in achieving distributed fit CFAR detection. The processor 114 may include a hardware accelerator, a controller, a control circuit, a microprocessor, its own chip, its own system-on-chip, device, processing unit, digital signal processing unit, graphics processing unit, or central processing unit. Processor 114 may include multiple processors or cores, embedded memory storing executable software or firmware, internal/private/secure caches, or any other computer element that enables processor 114 to execute machine-readable instructions for generating radar output.

In some examples, at least CRM 116 and processor 114 are a single component, such as an embedded system or a system on a chip. At least a portion of the CRM 116 is configured as a dedicated storage for the processor 114. CRM 116 may include a memory portion (e.g., memory) reserved by processor 114 to maintain data cube 118 before or after performing the distribution fit CFAR. Access to CRM 116 may be shared by other components of radar system 104. CRM 116 may also store machine readable instructions for performing radar operations. As two examples, CRM 116 stores instructions for performing radar functions performed by measurement estimator 120 and noise estimator 122.

The measurement estimator 120 is configured to estimate and store values for detection, including range, doppler and/or angle. It should be appreciated that the distributed fit CFAR technique may be applied to the radar system 104 regardless of whether the measurement estimator 120 may estimate only one of range, doppler, and angle, whether the measurement estimator may determine two of range, doppler, and angle, or whether all three of range, doppler, and angle may be estimated. With these measurements, the data cube 118 and information derived therefrom (e.g., radar tracking of objects) may be used to implement advanced safety or autonomous driving functions that avoid obstacles at a position and velocity inferred from the data cube 118. Information including data cube 118 may be communicated within radar system 104 to implement other functions (e.g., object classification, object tracking) of other radar systems (other radar systems not shown in fig. 1 for simplicity of the drawing). It is also possible that other systems of the vehicle 102 and/or other vehicles and external systems (e.g., using a vehicle-to-everything communication network) receive information (including the data cube 118) from the radar system 104 to enable these other vehicles to also drive safely.

Noise estimator 122 is configured to filter noise from data cube 118 to improve its accuracy and eliminate false alarm detection. Noise estimator 122 may be performed on behalf of measurement estimator 120 or as a preprocessing or post-processing step. Noise associated with data cube 118 may be filtered out of data cube 118 at different times in the processing pipeline of radar system 104. The noise estimator 122 may be before or after the execution of the measurement estimator 120. Noise estimator 122 may estimate noise at different stages of measurement estimator 120; this includes determining noise before, after, or simultaneously with the measurement estimator 120 performing range processing, doppler processing, and/or angle estimation. However, executing the noise estimator 122 earlier may improve the throughput of the radar system 104 because the measurement estimator 120 may avoid wasting processing resources (e.g., cycle time of the processor 114, memory capacity of the CRM 116) to estimate range, doppler, or angle that may be attributed to detection of noise.

Example noise estimator

Fig. 2 shows a conceptual diagram of an example noise estimator 200 of a radar system configured to perform distributed fit CFAR detection in accordance with the described techniques. Noise estimator 200 is an example of noise estimator 122 and is described in the context of environment 100 as part of radar system 104 of vehicle 102.

Noise estimator 200 is configured to apply a distribution fit CFAR to each cell or bin of data cube 118. Data cube 118 (or a memory location of data cube 118) is received as an input to noise estimator 200. The CFAR threshold for each cell is output for filtering noise from the data cube 118. Units of the data cube 118 that meet their respective CFAR thresholds may be processed by the measurement estimator 120, and units of the data cube 118 that do not meet their respective CFAR thresholds may be ignored. In the example shown in fig. 2, the noise estimator 200 divides the distribution fit CFAR check into four phases: calculate noise power, erase potential targets, fit noise data into a distribution model, and determine CFAR thresholds.

The adjacent selection component 202 determines the current unit under test (denoted as x _ij ) And its neighbors or bins (referred to as training unit 210) to perform the first phase. In this example, the training unit210 are drawn as a two-dimensional matrix, however, the training units may be organized as a single-dimensional array. The training unit is selected by adjacent selecting unit 202 to include the unit x having the function of the test unit _ij Noise data of similar statistics.

The data erasure component 204 performs the next phase of the distribution fitting CFAR check. An erasure threshold is calculated to remove a first group of training units 210 that have easily identifiable noise. For example, training unit 210 is subdivided into (e.g., six) equal subgroups, each subgroup containing four different training units. Calculate the average value of each subgroup and can sum the coarse noise power sigma _r Set equal to the minimum average of all subgroups. Then, the unit under test x may be determined by solving S using equation 1 _ij Is not limited by the erase threshold:

S＝βσ _r equation 1.

In equation 1, β is a scaling factor that may be empirically selected by the data erasure component 204 (e.g., it may vary over time to adjust for a particular application). During the erasure phase of the distribution fit CFAR check, training cells 210 with magnitudes (e.g., power) below S are saved for further processing, while training cells 210 that do not have magnitudes below S are immediately considered noise bin N _Storehouse And may be discarded. Training element 212 is shown to include saved white elements and ignored or discarded shadow elements.

The histogram builder component 206 generates a histogram 214 of the training unit 212 that remains after erasure. Training elements 212 are grouped into columns N of similar magnitude that increase along the x-axis _Storehouse Is a kind of medium. The y-axis of the histogram 214 is the estimated Probability Density Function (PDF) of the group. In some cases, column N _Storehouse Possibly with a training unit 212 of lower amplitude than S. However, due to their limited number, they are in column N _Storehouse May not have a significant impact on the results of CFAR threshold calculation.

In the fourth stage, the noise estimator relies on fitting model component 208, wherein histogram 214 is fitted to noise model 216. For example, when the amplitude of noise is calculated from complex measurements, a Rayleigh distribution can be observed, which is an uncorrelated normal distribution with equal variance and zero mean.

As a practical example, to effectively determine the CFAR threshold, the average value of the rayleigh random variable may be determined from equation 2:

once each column N is calculated _Storehouse From equation 3, the average rayleigh distribution scaling parameter σ can be determined _μ ：

Offset a may be used to account for non-zero average noise as given by equation 4.

The offset a is applied to the noise samples to account for known sources of noise, such as DC offset or phase noise, which enables real world measurements to be made by taking into account any shift in the PDF of the rayleigh distribution.

Another way to take this shift into account is to use the peak σ of the rayleigh distribution scaling parameter _pk The peak value of the rayleigh distribution scaling parameter may be calculated by determining the peak value of the histogram 214, the peak value of the histogram 214 being the maximum PDF (f _{Maximum value} ) As provided by equation 5:

in equation 5, f _{Maximum value} Is the peak of histogram 214. Then, a second scaling parameter (peak Rayleigh distribution scaling parameter σ _pk ) Can be calculated from equation 6:

average Rayleigh distribution scaling parameter sigma _μ And peak Rayleigh distribution scaling parameter sigma _pk The difference between them is given by equation 7:

the difference may be used to represent an unknown bias level applied to the noise samples. The p-th quantile function is defined by equation 8, where p is the false alarm rate:

finally, fitting model component 208 may generate current unit under test x using equation 9 _ij CFAR threshold T of (2) _ij ：

In equation 9, α is a scaling factor called a threshold factor. The threshold factor α may be used to account for constant offsets from other system parameters, such as accounting for correlation between antenna channels of the radar system 104. Using the appropriate alpha value, a CFAR threshold may be output for filtering noise from that particular cell of data cube 118.

In the above example, it is assumed that the ambient noise follows the rayleigh distribution. The distribution fitting CFAR may be applied to the case where noise is subject to other distributions in addition to only the rayleigh distribution. In the case where another distribution better simulates variable noise (such as a normal distribution), a distribution fit CFAR may also be used. The peak of the histogram 214 and its average value can be used to find all the parameters of the distribution quantile function. For example, a normal distributed quantile function is provided by equation 10:

in equation 10, the average μmay be calculated directly from the noise samples, and the standard deviation σ may be calculated based on equation 11:

in this way, the gaussian distribution-based CFAR can be calculated from equation 12:

T _n ＝α·Q(1-P _fa the method comprises the steps of carrying out a first treatment on the surface of the μ, σ) equation 12.

Example procedure for distribution fitting CFAR detection

Fig. 3 shows a flowchart of an example process for performing distributed fit CFAR detection in accordance with the described techniques. For ease of description, the process 300 is described primarily in the context of being performed by the radar system 104 using the processor 114 to access the CRM 116. For example, noise estimators 122 and 200 may write to portions of CRM 116 to write to data cube 118. Noise estimators 122 and 200 may read from portions of CRM 116 to read from data cube 118. The operations (also referred to as steps) of process 300 are numbered sequentially. However, this numbering does not necessarily imply a particular order of operation. The steps of process 300 may be rearranged, skipped, repeated, or performed in a different manner than the particular manner shown in the flow chart of fig. 3.

At step 302, a plurality of samples of radar returns are obtained, the radar returns including noise reflected from the environment external to the vehicle. For example, radar system 104 obtains radar echo 110-2 reflected from object 108 and other features in environment 100. Radar echo 110-2 includes noise in the actual target detection.

At step 304, an array of samples (e.g., a data cube) is maintained that includes each of the samples in a different corresponding cell of the array. For example, the data cube 118 is written by the processor 114 to the CRM 116.

At step 306, for each of the samples, a respective CFAR threshold for filtering noise from the corresponding cell of the data cube is determined by performing a distributed fit CFAR technique (e.g., shown as steps 306-1 through 306-5). For example, noise estimator 122 or noise estimator 200 reads each cell of data cube 118 and applies a distribution fit CFAR technique at each cell to generate a CFAR threshold for that cell.

At step 306-1, a set of neighboring cells to the corresponding cell is determined to be used as a training cell. For example, the noise estimator 200 relies on adjacent selection components 202 to analyze cells located in nearby distance, chirp, or channel domains to identify cells containing cells x to be measured with correspondence _ij Training unit 210 of noise data for similar statistics. The training units of the corresponding units may include units associated with the same chirp as the corresponding units and associated with the same distance or the same channel as the corresponding units. The training units may be associated with the same channel as the corresponding units. In other cases, the training units may be associated with the same distance as the corresponding units.

At step 306-2, training cells with high noise levels may be erased (scrub) prior to the next step. For example, the data erasure component 204 can enable the noise estimator 200 to reduce the number of training units 210 identified by the adjacent selection component 202 to remove units with high noise levels that are unlikely to represent an actual target.

At step 306-3, a histogram is generated that organizes the samples of the training unit into columns representing a continuous range of magnitudes of the samples of the training unit. For example, the noise estimator 200 uses a histogram builder component 206, the histogram builder component 206 generating a histogram 214 of training units 212 that remain after erasure of training units 210 originally selected by the adjacent selection component 202. The histogram builder component 206 may organize the samples of the training unit 212 into columns representing a continuous range of amplitudes based on the respective amplitudes of the samples of the training unit 212. The magnitudes of the columns in the histogram 214 correspond to the PDF of the samples of the training unit 212, which are associated with the range of magnitudes of the columns.

At step 306-4, the histogram is fitted to the noise distribution function. For example, the noise distribution function may be a Rayleigh (Rayleigh) distribution. The noise distribution function may be a normal distribution or other distribution functions. The noise estimator 200 executes a fitting model component 208, the fitting model component 208 fitting the histogram 214 to a noise model 216 applicable to the environment 100.

Noise in the environment 100 often varies. The noise model 216 used by the noise estimator 200 may also vary from one unit under test to the next in order to accommodate varying noise in the environment 100 (e.g., different noise models may be used for different units within the data cube 118). For example, a Rayleigh distribution function may be selected for a first cell of the data cube 118 to fit noise, and a normal distribution function may be selected for a second cell of the data cube 118. In other words, noise estimator 200 may fit a first histogram of a first cell of data cube 118 to a first noise distribution function (e.g., normal) and a second histogram of a second cell of data cube 118 to a second noise distribution function (e.g., rayleigh, other) that is different from the first noise distribution function.

At step 306-5, a respective CFAR threshold for the corresponding cell is determined from the noise distribution function fitted to the histogram of the corresponding cell. For example, where the noise model 216 is fitted to the histogram 214, the CFAR threshold for the unit under test may be determined by the noise estimator 200 by correlating with values on a curve defined by the noise model 216 fitted to the unit under test.

At step 308, samples from the data cube that do not meet the respective CFAR threshold for the corresponding cell are filtered. For example, each of the cells of the data cube 118 is filtered by the processor 114 using its respective CFAR threshold to eliminate false detections that occur due to varying (e.g., unpredictable) noise in the environment 100.

At step 310, in response to filtering the samples from the data cube, the data cube is output for use by the vehicle function in detecting objects present in the environment. For example, the processor 114 outputs an indication of the data cube 118 for enabling other systems of the vehicle 102 or systems external to the vehicle 102 to track objects in the field of view 106 of the radar system 104.

Performance of a distribution fitting CFAR detection

Fig. 4 shows a line graph 400 comparing the performance of the distribution fit CFAR detection with ground truth. Line graph 400 shows the results collected from a test chamber in which a single target was located one hundred forty away from the bin. Because the power of the return signal is so high, the phase noise creates a hump in graph 400 between one hundred and one hundred forty distance indices. During the measurement, hundreds of frames are recorded for each distance index, and each frame contains hundreds of samples. A total of hundreds of thousands of samples are collected in each index from which the actual CFAR threshold level for a given distance index can be estimated. Unlike the true CFAR threshold, the distribution fit CFAR is calculated based on data from a single frame for each distance index, which results in a cloud-like appearance in graph 400. As can be observed from graph 400, the center of the distribution fit CFAR point cloud is aligned with the true CFAR with only a small amount of deviation.

Fig. 5 shows a line graph 500 comparing the performance of a distribution fit CFAR test with an ordered statistics CFAR test. Fig. 6 shows a line graph 600 comparing the performance of a distribution fit CFAR test with a unit average CFAR test. Unlike other CFAR algorithms (such as unit average CFAR and ordered statistical CFAR), a distribution fit CFAR can provide accurate CFAR thresholds even when the noise distribution changes. As observed from graphs 500 and 600, in the pure noise region (distance index between forty and sixty), all CFAR algorithms can provide good CFAR threshold estimates. As phase noise becomes more dominant in the distance bins, the noise distribution gradually changes and the ordered statistical CFAR and the unit average CFAR no longer provide the correct CFAR threshold estimate. Shifts between true CFARs and ordered statistics CFARs and unit average CFARs indicate that these techniques are not adaptable to environmental changes. In contrast, a distribution fit CFAR can maintain a good CFAR threshold estimate because noise samples are analyzed and fit into the appropriate distribution model for the unit cell or bin under test. This procedure ensures that an accurate CFAR threshold estimate can be determined even if the noise distribution changes.

Further example

Some further examples in view of the above techniques include:

example 1: a method, the method comprising: obtaining, by a processor of the system, a plurality of samples of radar echoes, the radar echoes including noise reflected from an environment external to the vehicle; maintaining, by the processor, an array of samples, the array including each of the samples in a different corresponding cell of the array; for each of the samples, a respective Constant False Alarm Rate (CFAR) threshold for filtering noise from a corresponding cell of the array is determined by the processor by: determining a set of neighboring cells of the corresponding cell to use as training cells; generating a histogram that organizes the samples of the training unit into columns representing a continuous range of magnitudes of the samples of the training unit; fitting the histogram to a noise distribution function; and determining a respective CFAR threshold for the corresponding cell according to a noise distribution function fitted to the histogram of the corresponding cell; and filtering, by the processor, samples from the array that do not meet the respective CFAR threshold for the corresponding cell; and outputting, by the processor, the array for use by the vehicle function in detecting objects present in the environment in response to filtering the samples from the array.

Example 2: the method of example 1, wherein determining the respective CFAR threshold for filtering noise from the corresponding cells of the array comprises erasing training cells having a high noise level prior to generating the histogram.

Example 3: the method of any of the above examples, wherein the training unit of the corresponding unit comprises units associated with the same chirp as the corresponding unit and associated with the same distance or the same channel as the corresponding unit.

Example 4: the method of any of the above examples, wherein the training unit of the corresponding unit comprises units associated with the same channel as the corresponding unit.

Example 5: the method of any of the above examples, wherein the training unit of the corresponding unit comprises units associated with the same distance as the corresponding unit.

Example 6: the method of any of the above examples, wherein generating the histogram comprises: the samples of the training unit are organized into columns representing a continuous range of amplitudes based on respective amplitudes of the samples of the training unit, the amplitudes of the columns corresponding to a probability density function of the samples of the training unit associated with the range of amplitudes of the columns.

Example 7: the method of any of the examples above, wherein the noise distribution function comprises a rayleigh distribution.

Example 8: the method of any of the above examples, wherein the noise distribution function comprises a normal distribution.

Example 9: the method of any of the above examples, wherein fitting the histogram to the noise distribution function comprises: fitting a first histogram of the first unit to a first noise distribution function; and fitting a second histogram of the second unit to a second noise distribution function different from the first noise distribution function.

Example 10: the method of any of the above examples, wherein the array comprises a data cube.

Example 11: a system comprising means for performing the method of any of the examples above.

Example 12: a system comprising a processor configured to perform the method of any of the examples above.

Example 13: a computer readable medium comprising instructions that, when executed, cause a processor to perform the method of any of the examples above.

Idioms of the knot

While various embodiments of the present disclosure have been described in the foregoing description and shown in the accompanying drawings, it is to be understood that the disclosure is not so limited, but may be practiced in various ways within the scope of the following claims. From the foregoing description, it will be apparent that various modifications may be made without departing from the scope of the disclosure as defined by the following claims. In addition to radar systems, problems associated with CFAR threshold setting may also occur in other systems (e.g., lidar systems) that handle sensor point cloud detection in noisy environments, including driving situations. Thus, while described as improving radar detection, the techniques described previously may also be adapted and applied to other problems to effectively detect objects in a scene using other types of sensors.

The use of "or" and grammatical-related terms, unless the context clearly dictates otherwise, represents a non-exclusive alternative. As used herein, a phrase referring to "at least one of a list of items refers to any combination of such items, including individual members. As an example, "at least one of a, b, or c" is intended to encompass: a. b, c, a-b, a-c, b-c, and a-b-c, as well as any combination of a plurality of identical elements (e.g., a-a-a, a-a-b, a-a-c, a-b-b, a-c-c, b-b-b, b-b-c, c-c, and c-c-c, or any other ordering of a, b, and c).

Claims (20)

1. A system, the system comprising:

a processor configured to:

obtaining a plurality of samples of radar echoes, the radar echoes including noise reflected from an environment external to the vehicle;

maintaining an array of the samples, the array including each of the samples in a different corresponding cell of the array;

for each of the samples, a respective Constant False Alarm Rate (CFAR) threshold for filtering the noise from the corresponding cell of the array is determined by:

determining a set of neighboring cells of the corresponding cell to use as training cells;

generating a histogram that organizes the samples of the training unit into columns representing a continuous range of amplitudes of the samples of the training unit;

fitting the histogram to a noise distribution function; and

determining the respective CFAR threshold for the corresponding cell from the noise distribution function fitted to the histogram of the corresponding cell; and

filtering samples from the array that do not meet the respective CFAR threshold for the corresponding cell; and

in response to filtering samples from the array, the array is output for use by a vehicle function in detecting objects present in the environment.

2. The system of claim 1, wherein the processor is configured to determine the respective CFAR threshold for filtering the noise from the corresponding cell of the data cube by erasing training cells having a high noise level prior to generating the histogram.

3. The system of claim 1, wherein the training unit of the corresponding unit comprises units associated with the same chirp as the corresponding unit and associated with the same distance or the same channel as the corresponding unit.

4. The system of claim 1, wherein the training unit of the corresponding unit comprises a unit associated with the same channel as the corresponding unit.

5. The system of claim 1, wherein the training unit of the corresponding unit comprises units associated with the same distance as the corresponding unit.

6. The system of claim 1, wherein the processor is configured to generate the histogram by:

based on the respective amplitudes of the samples of the training unit, the samples of the training unit are organized into columns representing a continuous range of amplitudes,

the magnitude of the column corresponds to a probability density function of the samples of the training unit associated with a range of magnitudes of the column.

7. The system of claim 1, wherein the noise distribution function comprises a rayleigh distribution.

8. The system of claim 1, wherein the noise distribution function comprises a normal distribution.

9. The system of claim 1, wherein the processor is further configured for fitting the histogram to the noise distribution function by:

fitting a first histogram of the first unit to a first noise distribution function; and

a second histogram of a second unit is fitted to a second noise distribution function different from the first noise distribution function.

10. The system of claim 1, wherein the system comprises a radar system of the vehicle and the array comprises a data cube.

11. A method, the method comprising:

obtaining, by a processor of a system, a plurality of samples of radar echoes, the radar echoes including noise reflected from an environment external to the vehicle;

maintaining, by the processor, a data cube of the samples, the data cube including each of the samples in a different corresponding cell of the data cube;

for each of the samples, determining, by the processor, a respective Constant False Alarm Rate (CFAR) threshold for filtering the noise from the corresponding cell of the data cube by:

determining a set of neighboring cells of the corresponding cell to use as training cells;

generating a histogram that organizes the samples of the training unit into columns representing a continuous range of amplitudes of the samples of the training unit;

fitting the histogram to a noise distribution function; and

determining the respective CFAR threshold for the corresponding cell from the noise distribution function fitted to the histogram of the corresponding cell; and

filtering, by the processor, samples from the data cube that do not meet the respective CFAR threshold for the corresponding cell; and

in response to filtering samples from the data cube, the array is output by the processor for use by a vehicle function in detecting objects present in the environment.

12. The method of claim 11, wherein determining the respective CFAR threshold for filtering the noise from the corresponding cells of the array comprises erasing training cells having a high noise level prior to generating the histogram.

13. The method of claim 11, wherein the training unit of the corresponding unit comprises units associated with the same chirp as the corresponding unit and associated with the same distance or the same channel as the corresponding unit.

14. The method of claim 11, wherein the training unit of the corresponding unit comprises units associated with the same channel as the corresponding unit.

15. The method of claim 11, wherein the training unit of the corresponding unit comprises units associated with the same distance as the corresponding unit.

16. The method of claim 11, wherein generating the histogram comprises:

based on the respective amplitudes of the samples of the training unit, the samples of the training unit are organized into columns representing a continuous range of amplitudes,

the magnitude of the column corresponds to a probability density function of the samples of the training unit associated with a range of magnitudes of the column.

17. The method of claim 11, wherein the noise distribution function comprises a rayleigh distribution.

18. The method of claim 11, wherein the noise distribution function comprises a normal distribution.

19. The method of claim 11, wherein fitting the histogram to the noise distribution function comprises:

fitting a first histogram of the first unit to a first noise distribution function; and

a second histogram of a second unit is fitted to a second noise distribution function different from the first noise distribution function.

20. A computer readable medium comprising instructions that when executed cause a processor to perform the method of any of claims 11-19.