A partition adaptive windowing constant false alarm rate millimeter wave radar target detection enhancement method

Abstract

The application discloses a kind of partition adaptive windowing constant false alarm rate millimeter wave radar target detection enhancement methods, belong to vehicle-mounted millimeter wave radar signal processing technical field, by the adaptive distance partition based on echo energy accumulation distribution, and partition boundary is generated by actual echo data drive;Partition independent windowing, high sidelobe suppression window is used to suppress strong target sidelobe in near distance, resolution is preserved using narrow main lobe window in middle distance, and balance is appropriately windowed in far distance;Partition independent CFAR detection, near distance OS-CFAR is resistant to outliers, middle distance CFAR is considered with stability by deleting mean value, far distance CA-CFAR maximizes sensitivity, reference window, protection unit and threshold are adaptively adjusted with target density and dynamic range, to ensure that the false alarm rate of each interval is consistent and the detection rate is optimal;Boundary overlap joint detection and deduplication avoid missing or repetition;Equivalent window replacement is used in frequency domain, local correction range image is used using equivalent convolution kernel, complete FFT is not redone, and real-time calculation load is reduced.

Description

Technical Field

[0001] This invention belongs to the field of vehicle-mounted millimeter-wave radar signal processing technology, specifically relating to a partitioned adaptive windowing constant false alarm rate millimeter-wave radar target detection enhancement method. Background Technology

[0002] The vehicle-mounted FMCW millimeter-wave radar is one of the core sensors in an intelligent driving environment perception system. Its basic signal processing chain typically includes: acquiring target reflected echoes via a receiving antenna; obtaining digitized raw echo data through mixing and analog-to-digital conversion sampling; the data is organized in a two-dimensional matrix, with one dimension corresponding to fast time sampling (range dimension) and the other corresponding to the linear frequency modulated signal index (velocity dimension); windowing and Fast Fourier Transform (FFT) are performed on the raw echo data along the range dimension to obtain a one-dimensional range profile; windowing and FFT are performed along the velocity dimension to obtain a range-Doppler two-dimensional spectrum matrix; constant false alarm rate (CFAR) detection is performed on the two-dimensional spectrum to obtain target traces; finally, post-processing of the traces and target tracking are performed.

[0003] In this processing chain, windowing and CFAR detection are two key steps that directly affect the quality of target detection. The purpose of windowing is to suppress the inherent spectral leakage effect of the Fourier transform: without windowing (equivalent to a rectangular window), the Fourier transform will produce high sidelobes of approximately -13dB, and the sidelobes of strong targets may be misclassified as false targets or obscure nearby weak targets. Commonly used window functions include the Hamming window (sidelobes approximately -43dB), the Hanning window (sidelobes approximately -31dB), the Chebyshev window (the sidelobe suppression level can be controlled by parameters, such as -60dB and above), and the Blackman window (sidelobes approximately -58dB), etc. Different window functions have different main lobe widths and sidelobe suppression characteristics. The Hamming window achieves a good balance between main lobe width and side lobe suppression, with its main lobe width approximately 1.46 times that of a rectangular window. The Hanning window has a main lobe width approximately 1.62 times that of a rectangular window, but its side lobes decay more rapidly. The Chebyshev window minimizes the main lobe width while maintaining a specified side lobe suppression level. The Blackman window achieves side lobe suppression of approximately -58 dB, but its main lobe width is approximately 2.3 times that of a rectangular window. The core dilemma in window function selection lies in the fact that stronger side lobe suppression results in a wider main lobe, but lower range resolution, potentially leading to the inability to separate nearby targets at range.

[0004] Currently, mainstream solutions in the industry employ a unified windowing strategy and unified CFAR detection parameters across the entire distance dimension. This "unified processing across the entire distance" approach presents three core technical contradictions.

[0005] The first contradiction is the conflict between sidelobe masking of strong targets at close range and the sensitivity of detecting weak targets at long range. In typical vehicle-mounted radar operating scenarios, strong reflective targets such as pedestrians and curbs are often present in the close range (e.g., 0 to 30 meters). The high sidelobes of these targets may cause false detections in the long range. At the same time, weak targets (such as small obstacles at a distance) in the long range (e.g., above 80 meters) require high detection sensitivity. If a window function with strong sidelobe suppression (such as the Chebyshev window) is chosen, the broadening of the main lobe will cause distant neighboring targets to stick together in the range dimension and become indistinguishable. If a window function with a narrow main lobe (such as the Hanning window) is chosen, the high sidelobes of strong targets at close range will generate a large number of false alarms throughout the entire range, forming a "strong masking weak" phenomenon.

[0006] The second contradiction lies in the discrepancy between the statistical characteristics of clutter across different range intervals and the unified CFAR parameters. Clutter in the near-range range of vehicle-mounted radar (ground clutter, curb clutter, etc.) exhibits a non-Gaussian heavy-tailed distribution; clutter in the mid-range range shows a transition from a heavy-tailed to a Rayleigh distribution; and clutter in the long-range range tends towards a Rayleigh distribution, primarily consisting of receiver thermal noise. Existing unified CFAR detection strategies use the same detector type and parameters for all range cells, inevitably leading to false alarm rates deviating from design values ​​or a sharp drop in detection rates in at least some ranges. For example, a high threshold factor set with near-range heavy-tailed clutter as a reference will result in severely insufficient detection sensitivity in long-range uniform clutter environments.

[0007] The third contradiction is the information loss caused by uniform windowing across the entire range. Windowing is essentially a trade-off between range resolution and spectral purity, and its effectiveness is highly dependent on the local characteristics of the signal. Applying the same window function to near-range and far-range regions where the dynamic range of the signal and the density of the target differ significantly will inevitably cause performance loss in at least some intervals—either excessive resolution and insufficient sensitivity, or insufficient sidelobe suppression leading to false alarms.

[0008] In existing technologies, several patent documents have addressed different aspects of millimeter-wave radar signal processing. CN115453469A proposes a method for suppressing interference from FMCW millimeter-wave radar slice reconstruction. Its technical approach utilizes the time-frequency domain characteristic differences between the real target echo and external interference for interference identification and filtering. This scheme processes external interference signals and does not involve differentiated processing of the range dimension of the target echo signal itself under interference-free conditions. CN115562620A proposes a real-time super-resolution method for millimeter-wave TDM-MIMO radar based on FPGA. Its innovation lies in the hardware modification of the angle dimension super-resolution algorithm, belonging to the angle estimation stage, which is completely different from windowing and CFAR detection optimization in range dimension signal processing within the radar signal processing chain. CN117634531A proposes a data-cascaded vehicle-mounted 4D millimeter-wave radar signal processing method. Its core lies in the data flow division between master and slave chips and the cross-chip transmission and decoupling of angle dimension information. This is a system architecture-level design and does not involve differentiated strategies for range dimension signal processing within a single chip. CN119291644A proposes a signal processing method for stepped linear frequency modulated continuous wave systems. Its core is cross-frame velocity-dimensional phase compensation, operating on the slow time dimension to reduce radial relative velocity-dimensional main lobe broadening, without involving spatial differentiation processing of the range dimension within the same frame. CN121413222A proposes an improved minimum variance distortion-free response beamforming method based on diagonal loading, belonging to the field of array spatial filtering, dealing with signal merging problems in the spatial domain, which is completely different from the signal domain of range-dimensional spectral processing. CN121165059A proposes a multipath false target detection and identification method for vehicle-mounted millimeter-wave radar. Its core is to determine the nature of the detected points based on their physical attributes. It deals with the post-processing stage after CFAR detection, rather than parameter optimization of CFAR detection itself. Furthermore, its range segmentation is only for setting different energy thresholds to determine multipath, without involving differentiated configurations of window functions and CFAR detector types.

[0009] In summary, existing technologies do not address the technical solutions for partitioned differential windowing and partitioned differential CFAR detection based on spatial differences in the distance dimension within the same frame signal processing. How to simultaneously address near-range sidelobe suppression and far-range weak target detection within the same frame, adapt to the differences in clutter statistical characteristics across different distance intervals, and minimize information loss caused by uniform windowing are pressing technical problems that need to be solved in this field. Summary of the Invention

[0010] To address the shortcomings of the existing technologies, this application provides a method for enhancing target detection in millimeter-wave radar with partitioned adaptive windowing and constant false alarm rate.

[0011] In its first aspect, this application proposes a partitioned adaptive windowing constant false alarm rate millimeter-wave radar target detection enhancement method, comprising the following steps: Step S1: Acquire the raw radar echo data and perform initial windowing and range-dimensional Fourier transform to obtain a one-dimensional range profile and echo power distribution curve; Step S2: Based on the energy accumulation distribution characteristics of the echo power distribution curve, the distance dimension is adaptively divided into multiple processing intervals; Step S3: For each processing interval, independently select a window function based on the signal characteristics within the processing interval and perform windowing processing to optimize the one-dimensional distance image of each interval; Step S4: Perform velocity-dimensional Fourier transform on the optimized one-dimensional distance image of each interval to generate a distance-Doppler two-dimensional spectrum; Step S5: Perform constant false alarm rate detection independently for each processing interval on the range-Doppler two-dimensional spectrum, wherein each processing interval uses a constant false alarm rate detector type and adaptive parameters determined by the interval clutter characteristics; Step S6: Perform overlapping transition processing on the boundary regions of adjacent processing intervals, merge the detection results of each processing interval and boundary region, and output the final target point set.

[0012] Furthermore, in step S2, based on the energy accumulation distribution characteristics of the echo power distribution curve, the distance dimension is adaptively divided into multiple processing intervals, including: Smooth the echo power distribution curve; Calculate the normalized cumulative energy curve of the smoothed power distribution curve; Based on a preset energy ratio threshold, the initial partition boundary is determined on the normalized cumulative energy curve. Within the local search range near the initial partition boundary, find the location with the most gradual power change, and fine-tune the location with the most gradual power change as the final partition boundary.

[0013] Furthermore, the signal characteristics on which the independent selection of the window function in step S3 is based include the signal dynamic range and target density within the processing interval; Among them, the signal dynamic range refers to the difference between the maximum and minimum echo power within the processing interval, and the target density refers to the proportion of the number of distance units exceeding the initial detection threshold within the processing interval to the total number of distance units in the processing interval. When the signal dynamic range of a processing interval is higher than the first preset value and the target density is lower than the second preset value, a high sidelobe suppression window is selected. When the signal dynamic range of a processing interval is lower than the first preset value and the target density is higher than the second preset value, a narrow main lobe window is selected.

[0014] Furthermore, the high sidelobe suppression window is a Chebyshev window, and the narrow main lobe window is a Hanning window.

[0015] Furthermore, in step S5, when performing constant false alarm rate detection independently for each processing interval, an ordered statistic constant false alarm rate detector is used for the near-distance processing interval; a truncated mean constant false alarm rate detector is used for the medium-distance processing interval; and a unit average constant false alarm rate detector is used for the far-distance processing interval.

[0016] Furthermore, the adaptive parameters for the constant false alarm rate (CFAR) detection include the reference window size, the number of protection units, and the threshold factor. The adaptive parameters are adaptively adjusted according to the target density and signal dynamic range within the processing interval.

[0017] Furthermore, step S6 involves overlapping transition processing of the boundary regions of adjacent processing intervals, including: A preset number of distance units are extended on both sides of the boundary of adjacent processing intervals to form an overlapping area; Within the overlapping area, the full set of constant false alarm rate detection parameters for the two adjacent processing intervals are used for independent detection, and the two sets of detection results are logically ORed and merged. The merged set of points is deduplicated based on distance tolerance and velocity tolerance. If the distance difference and velocity difference between two points are both less than the preset tolerance value, the two points are determined to correspond to the same target, and the one with the higher tolerance is retained.

[0018] Furthermore, the windowing process in step S3 is implemented through a frequency domain equivalent method, specifically: using the frequency domain equivalent convolution kernel between the target window function and the initial window function in step S1, local convolution correction is performed on the data in the corresponding processing interval of the one-dimensional distance image obtained in step S1. The window function used in the initial windowing in step S1 is a Hamming window.

[0019] Furthermore, the echo power distribution curve in step S1 is: the one-dimensional power distribution vector obtained by averaging the echo energy of each range cell in the one-dimensional range image on all linear frequency modulated signals.

[0020] Secondly, this application proposes a partitioned adaptive windowing constant false alarm rate millimeter-wave radar target detection enhancement system, comprising: The data preprocessing module is used to acquire raw radar echo data, perform initial windowing and range-dimensional Fourier transform on the raw echo data, and generate a one-dimensional range profile and echo power distribution curve. The interval division module, connected to the data preprocessing module, is used to adaptively divide the distance dimension into multiple processing intervals based on the energy accumulation distribution characteristics of the echo power distribution curve. The partitioned windowing optimization module, connected to the interval division module, is used to independently select a window function and perform windowing processing on each processing interval based on the signal characteristics within the processing interval, so as to optimize the one-dimensional distance image of each interval. The two-dimensional spectrum generation module, connected to the partition windowing optimization module, is used to perform a velocity-dimensional Fourier transform on the optimized one-dimensional range image of each interval to generate a range-Doppler two-dimensional spectrum. The partitioned constant false alarm rate (CFAR) detection module is connected to the two-dimensional spectrum generation module and the interval division module. It is used to independently perform CFAR detection on each processing interval on the distance-Doppler two-dimensional spectrum, wherein each processing interval uses a CFAR detector type and adaptive parameters determined by the clutter characteristics of the interval. The boundary fusion and output module is connected to the partition constant false alarm detection module and the interval division module. It is used to perform overlapping transition processing on the boundary regions of adjacent processing intervals, merge the detection results of each processing interval and boundary region, and output the final target point set.

[0021] Thirdly, this application proposes an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0022] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described above.

[0023] The beneficial effects of this invention are: By employing an adaptive distance partitioning strategy based on the cumulative echo energy distribution, this approach breaks away from the traditional fixed pattern of uniform processing across all distances. The partition boundaries are generated driven by the actual echo energy distribution data of the current frame, rather than a preset fixed distance threshold. This automatically adapts to significant changes in echo energy distribution under different driving scenarios, ensuring that the partitioning results always match the echo characteristics of the current scenario. By employing partition-based adaptive windowing, a balance is achieved between strong sidelobe suppression in the near-range region and high range resolution in the far-range region within the same frame. A high sidelobe suppression window is selected for the near-range high dynamic range, low-density region to effectively suppress sidelobe leakage from strong near-range targets; a narrow main lobe window is selected for the mid-range, high-density region to ensure range resolution in multi-target scenarios; and a suitable window function is retained for the far-range region to achieve a balance between sidelobe suppression and sensitivity. By selecting and configuring independent adaptive CFAR detectors for each zone, the detection performance of each distance interval is made close to the theoretical optimum. In the near-range interval, an ordered statistic constant false alarm rate (CFAR) detector is used to resist outlier interference from heavy-tail clutter; in the mid-range interval, a pruned mean CFAR detector is used to balance stability and anti-interference capability; and in the far-range interval, a cell-average CFAR detector is used to maximize detection sensitivity against a uniform noise background. The reference window size, the number of guard cells, and the threshold factor are adaptively adjusted according to the target density and signal dynamic range of each interval to ensure consistent false alarm rates and optimal detection rates across all intervals. By using a joint detection and deduplication mechanism for overlapping boundary regions, the problem of missed or duplicate detection of targets near the boundary of the interval that may be caused by partitioned processing is effectively solved, ensuring the consistency and continuity of detection results across the entire range. By replacing the window function in the frequency domain, the high computational cost of re-performing the full range-dimensional FFT for each window function change is avoided. The obtained range image is locally corrected using equivalent convolution kernels in the frequency domain between different window functions. The computational cost depends only on the length of the interval where the window function needs to be replaced and the length of the convolution kernel, significantly reducing the real-time computational load for engineering implementation. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the overall processing flow of the method of the present invention.

[0025] Figure 2 This is a sub-flowchart for the adaptive distance partitioning step S2.

[0026] Figure 3 The structural block diagram of step S3 for adaptive windowing of partitions.

[0027] Figure 4 This is a schematic diagram of a joint detection and deduplication mechanism for overlapping boundary regions.

[0028] Figure 5 This is an architecture diagram of the system of the present invention. Detailed Implementation

[0029] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein; rather, these embodiments are provided so that a more thorough understanding of the invention can be achieved and that the full scope of the invention can be conveyed to those skilled in the art.

[0030] Firstly, this application proposes a partitioned adaptive windowing constant false alarm rate millimeter-wave radar target detection enhancement method, such as... Figure 1 As shown, it includes the following steps: Step S1: Acquire the raw radar echo data and perform initial windowing and range-dimensional Fourier transform to obtain a one-dimensional range profile and echo power distribution curve; Furthermore, the echo power distribution curve in step S1 is: the one-dimensional power distribution vector obtained by averaging the echo energy of each range cell in the one-dimensional range image on all linear frequency modulated signals.

[0031] The receiving antenna receives the reflected echo from the target, and after mixing and ADC sampling, the raw echo data is obtained in digitized form. The data is organized in the form of a two-dimensional matrix, with the row direction corresponding to the distance dimension (fast time sampling within each chirp) and the column direction corresponding to the velocity dimension (chirp sequence within the same frame).

[0032] First, an initial window function is used to uniformly window all the raw data of chirp, with the Hamming window being preferred because it provides a good trade-off between the main lobe width and side lobe suppression, providing sufficient information for subsequent processing to perform interval feature analysis, while avoiding signal contamination at interval boundaries due to excessively high side lobes.

[0033] After windowing, a fast Fourier transform is performed along the distance dimension to obtain a one-dimensional distance image: each row of the matrix corresponds to a distance cell, each column corresponds to a chirp, and the amplitude of the matrix elements reflects the echo intensity of the distance cell in the chirp.

[0034] Subsequently, for each range cell, the average echo energy across all chirps is calculated, resulting in a one-dimensional power distribution curve that varies along the range dimension. The fluctuations of this curve reflect the changes in echo energy intensity at different range locations: strong reflective targets correspond to peak values, while noisy regions correspond to valleys. This power distribution curve is the core input data for subsequent adaptive range partitioning.

[0035] The reason for choosing a Hamming window instead of a rectangular window as the initial window is that a rectangular window produces higher sidelobes, causing the sidelobe energy of strong near-range targets to contaminate the power distribution estimation of far-range regions. This could lead to subsequent steps misclassifying "false sidelobe energy" as "real far-range weak target signals," thus incorrectly affecting the location of the partition boundaries. The -43dB sidelobes of the Hamming window are sufficient to avoid such misclassifications in most scenarios.

[0036] Step S2: Based on the energy accumulation distribution characteristics of the echo power distribution curve, the distance dimension is adaptively divided into multiple processing intervals; Furthermore, such as Figure 2 As shown, in step S2, based on the energy accumulation distribution characteristics of the echo power distribution curve, the distance dimension is adaptively divided into multiple processing intervals, including: Smooth the echo power distribution curve; Calculate the normalized cumulative energy curve of the smoothed power distribution curve; Based on a preset energy ratio threshold, the initial partition boundary is determined on the normalized cumulative energy curve. Within the local search range near the initial partition boundary, find the location with the most gradual power change, and fine-tune the location with the most gradual power change as the final partition boundary.

[0037] This step utilizes the range-dimensional power distribution curve obtained in step S1. By analyzing the cumulative distribution pattern and local variation characteristics of signal energy along the range, the entire range detection range is automatically divided into several processing intervals. The number of intervals and their boundary positions are driven by the data itself, rather than preset fixed values. Specifically, the power distribution curve is first smoothed using a moving average. The principle is that in multi-target scenarios, the original power values ​​of each range cell have a large number of random fluctuations. These fluctuations may originate from noise fluctuations between chirps, target micro-movements, and other factors. If used directly for partitioning without smoothing, it is easy to generate incorrect partition boundaries at local power fluctuations. The moving average eliminates high-frequency random jitter by taking the arithmetic mean of the power values ​​of each range cell and several adjacent cells, while retaining the main trend of power variation with distance. After obtaining the smoothed power distribution curve, the normalized cumulative energy curve along the distance dimension is calculated. This curve starts from zero, and the smoothed power value of each distance cell is accumulated until it reaches 100% of the total energy. The slope of the cumulative energy curve reflects the energy density of each distance segment: steeply rising sections indicate a large concentration of echo energy within that range, while gently rising sections indicate sparse echo energy within that range. The logic for determining the partition boundaries based on the cumulative energy curve is as follows: Two energy ratio thresholds are set, such as 30% and 70%. The distance cell positions corresponding to the first reaching of these two thresholds on the cumulative energy curve are found as the initial candidate boundaries for partitioning. These two positions divide the distance dimension into three intervals: The first interval, from the closest distance to the first boundary, contains the first 30% of the accumulated echo energy; The second interval, from the first boundary to the second boundary, contains the middle 40% of the accumulated echo energy; The third interval, from the second boundary to the farthest distance, contains the last 30% of the accumulated echo energy.

[0038] The rationale behind this energy-based zoning strategy lies in the fact that, in typical road scenarios, the echo energy in the near-range region is usually much higher than that in the far-range region. This is because near-range targets not only have shorter reflection paths and lower attenuation, but also that the near-range reflective surfaces in the road environment, such as vehicles, guardrails, and buildings, are far more abundant than those at a distance. Therefore, areas where energy accumulates rapidly correspond precisely to the near-range strong target areas that require strong sidelobe suppression, while areas where energy accumulates slowly correspond precisely to the far-range weak target areas that require high detection sensitivity. In other words, the spatial distribution of echo energy itself contains the natural basis for zoning.

[0039] Near the candidate boundary location determined based on the accumulated energy, further fine-tuning is needed to avoid the partition boundary cutting exactly onto the main lobe of a strong target's echo. If the boundary passes through the target's main lobe, the target's main lobe energy will be split into two adjacent intervals, which may lead to missed detections due to energy dispersion in the independent detection of the two intervals.

[0040] The fine-tuning method utilizes a fundamental fact in signal processing: between the main lobes of the echoes of two adjacent strong targets, there exists a local minimum in the power curve. Setting the partition boundary at this point minimizes the impact on target segmentation.

[0041] Specifically, within a local search range near the candidate boundary, the location with the most gradual power change is identified as the final, fine-tuned boundary. A gradual power change means that the location is in the transition region between two energy peaks, rather than on a rising or falling peak slope.

[0042] If the power variation remains large within the local search range, it indicates the presence of a dense and continuous group of targets in that area. In this case, the position with the lowest power value within the search range is selected as the boundary. A relatively weak position is chosen to enter the area where the signal is strongest, which is equivalent to finding a relatively sparse entry point in a continuously distributed group of targets.

[0043] After the above processing, several range intervals are obtained. Each interval covers a continuous range of range cells, with no gaps or overlaps between adjacent intervals. Subsequent steps will add additional overlap processing bands at the boundaries. In a preferred scheme, there are three partitions, roughly corresponding to the near-range zone (approximately 0-30m), the medium-range zone (approximately 30-80m), and the far-range zone (approximately 80m and above). However, the specific boundaries are automatically determined by the data-driven method in this step, which can adapt to different road scenarios and different radar hardware parameters.

[0044] Step S3: For each processing interval, independently select a window function based on the signal characteristics within the processing interval and perform windowing processing to optimize the one-dimensional distance image of each interval; Furthermore, such as Figure 3As shown, the signal characteristics on which the independent selection of the window function in step S3 is based include the signal dynamic range and target density within the processing interval; Among them, the signal dynamic range refers to the difference between the maximum and minimum echo power within the processing interval, and the target density refers to the proportion of the number of distance units exceeding the initial detection threshold within the processing interval to the total number of distance units in the processing interval. When the signal dynamic range of a processing interval is higher than the first preset value and the target density is lower than the second preset value, a high sidelobe suppression window is selected. When the signal dynamic range of a processing interval is lower than the first preset value and the target density is higher than the second preset value, a narrow main lobe window is selected.

[0045] Furthermore, the high sidelobe suppression window is a Chebyshev window, and the narrow main lobe window is a Hanning window.

[0046] Furthermore, the windowing process in step S3 is implemented through a frequency domain equivalent method, specifically: using the frequency domain equivalent convolution kernel between the target window function and the initial window function in step S1, local convolution correction is performed on the data in the corresponding processing interval of the one-dimensional distance image obtained in step S1. The window function used in the initial windowing in step S1 is a Hamming window.

[0047] For each distance interval determined in step S2, this step independently selects the window function most suitable for the signal characteristics and clutter environment of that interval, and uses the selected window function to re-window the echo data of that interval, replacing the initial window function used uniformly in step S1. Specifically: for each distance interval, first analyze the signal characteristics of that interval and extract two key decision criteria: Dynamic range of signal within the interval: This evaluates the difference between the strongest signal and the noise floor within the interval. The strongest signal is taken as the maximum value of the power distribution curve within the interval; the noise floor is taken as the median of the power distribution curves on several distance cells outside the interval that are not affected by the target signal within the interval. The cells outside the interval are selected as the noise reference to avoid the strong target signal within the interval from contaminating the noise estimation.

[0048] The dynamic range of a signal reflects the prominence of the strongest target relative to the background noise within that range. A larger dynamic range (e.g., exceeding 30 dB) indicates the presence of one or more strong targets that are very prominent relative to the noise within that range, with correspondingly higher sidelobe energy, and thus more severe potential interference to adjacent range cells.

[0049] Target density within the interval: Count the number of energy peaks in the interval that exceed the local noise threshold (the noise floor multiplied by a moderate multiple, such as 8 to 12 times, corresponding to a detection threshold of about 9 to 11 dB), and divide by the total number of distance cells contained in the interval to obtain the normalized target density index.

[0050] Target density reflects the target distribution pattern within the interval: low density indicates that there are only a few isolated strong targets within the interval, while high density indicates that there are a large number of target reflection points that are close to each other within the interval.

[0051] Based on the above two criteria, different window function selection logic is triggered for each distance interval: Logic 1: High dynamic range and low density. This range contains a small number of very strong isolated targets, and the distance between these targets is ample. In this case, preventing the sidelobes of strong targets from contaminating other range cells is the top priority. A high sidelobe suppression window is selected, with its sidelobe level parameter set according to the signal dynamic range within the range. The sidelobe level is suppressed to a level that is one safety margin lower than the strongest signal in the range (e.g., 10 dB lower plus the dynamic range value), ensuring that even at the strongest signal, the sidelobe energy is sufficiently suppressed to not pose a threat to the detection of any range cell.

[0052] Logic 2: Low dynamic range and high density. This range contains a large number of targets, but their energy is generally low, and the targets are closely spaced. In this case, maintaining range resolution and avoiding adjacent targets clustering together in the spectrum are the top priorities. A narrow main lobe window is chosen, sacrificing some sidelobe suppression capability in exchange for optimal range resolution, ensuring that each individual target in the dense target group can be clearly distinguished.

[0053] Logic 3: High dynamic range and high density. This range contains both strong reflectors and dense target groups, typical of complex urban road environments. In this case, a balance needs to be found between sidelobe suppression and range resolution. An adjustable balanced window is selected, whose parameters are dynamically adjusted according to the dynamic range: the larger the dynamic range, the higher the parameter values, resulting in stronger sidelobe suppression, but also greater mainlobe broadening.

[0054] Logic 4: Low dynamic range and low density. Targets in this range are sparse and have weak energy, typical of suburban areas or open roads. A moderate sidelobe suppression window is selected to provide basic sidelobe suppression while maintaining a good signal-to-noise ratio, without sacrificing the detection sensitivity of weak targets due to excessive sidelobe suppression.

[0055] Based on the above selection results, each interval is re-windowed and Fourier transformed using its corresponding window function, replacing the initial window function used uniformly in step S1. After the above partitioned windowing process, the spectrum in each distance interval of the optimized one-dimensional range image is obtained by processing the window function that best matches the interval. The near-range region prioritizes sidelobe suppression, the mid-range region balances resolution and sidelobe suppression, and the far-range region prioritizes detection sensitivity.

[0056] Step S4: Perform velocity-dimensional Fourier transform on the optimized one-dimensional distance image of each interval to generate a distance-Doppler two-dimensional spectrum; In step S3, the optimized one-dimensional range image is windowed and subjected to Fast Fourier Transform along the velocity dimension to generate a range-Doppler two-dimensional spectral matrix. Hamming windows are typically used for windowing in the velocity dimension to maintain reasonable velocity resolution while suppressing Doppler sidelobes.

[0057] In the two-dimensional spectrum generated in this step, each interval of the range dimension has used a window function optimally selected for that interval, thereby achieving the following: the sidelobes of strong targets in the near-range interval are strongly suppressed, and no spurious responses spanning multiple range cells and Doppler cells are generated in the two-dimensional spectrum; the signals of weak targets in the far-range interval are preserved with the best range resolution, which is beneficial for distinguishing them from noise in subsequent CFAR detection.

[0058] Step S5: Perform constant false alarm rate detection independently for each processing interval on the range-Doppler two-dimensional spectrum, wherein each processing interval uses a constant false alarm rate detector type and adaptive parameters determined by the interval clutter characteristics; Furthermore, in step S5, when performing constant false alarm rate detection independently for each processing interval, an ordered statistic constant false alarm rate detector is used for the near-distance processing interval; a truncated mean constant false alarm rate detector is used for the medium-distance processing interval; and a unit average constant false alarm rate detector is used for the far-distance processing interval.

[0059] Furthermore, the adaptive parameters for the constant false alarm rate (CFAR) detection include the reference window size, the number of protection units, and the threshold factor. The adaptive parameters are adaptively adjusted according to the target density and signal dynamic range within the processing interval.

[0060] In this step, for each range interval determined in step S2, the CFAR detector type is independently selected and its parameters are configured on a range-Doppler two-dimensional spectrum, and target detection is performed interval by interval. Specifically: Different CFAR detectors have their own advantages and disadvantages in different clutter environments; no single detector is optimal in all regions. This step selects the appropriate detector based on prior knowledge of the statistical characteristics of clutter in each region: In the near-range region, the clutter amplitude distribution exhibits a heavy-tailed characteristic, with a higher probability of large-amplitude clutter than predicted by the classic Rayleigh distribution. Under this non-uniform clutter background, the performance of CA-CFAR degrades significantly. When a high-energy clutter point falls within the reference window, the mean is raised by this outlier, causing the threshold to be excessively raised, thus obscuring the echoes of truly weak targets near that cell. Therefore, OS-CFAR is preferred in the near-range region. This detector does not use the mean; instead, it sorts the reference cells by energy from low to high and selects a relatively high-order value (e.g., the 75th percentile, approximately 3 / 4 of the way down) as the noise estimate. Since high-energy outliers are pushed to the end of the sequence after sorting, selecting a representative value from the beginning of the sequence as the noise estimate effectively masks the interference of outliers. The specific value of the ordinal position determines the degree of masking. A higher ordinal position (closer to the maximum value within the reference window) is closer to the mean effect of CA-CFAR (good robustness, but sensitive to multi-target scenarios). A lower ordinal position is more robust to outliers and interfering targets, but the statistical stability of noise estimation decreases. Selecting an ordinal position of approximately 3 / 4 is an empirically optimal trade-off between stability and robustness.

[0061] Mid-range region: The clutter statistical characteristics in this region fall between a heavy-tailed distribution and a uniform distribution, and the target density is typically high because mid-range is the densest sensing area in urban traffic environments. In this scenario, CMLD-CFAR is selected. This detector first sorts the reference cells by energy, then removes the highest-energy cells, and finally takes the average of the remaining cells as the noise estimate. The average value after removing high-energy outliers and interfering targets has both statistical stability and is not contaminated by a few anomalous cells. The removal ratio is dynamically adjusted according to the target density in this region: the removal ratio is increased when the density is high and decreased when the density is low.

[0062] Long-range range: In this range, clutter tends to have a uniform Rayleigh distribution, and targets are sparse. Under these ideal conditions, CA-CFAR has theoretically optimal detection performance. In uniform Rayleigh clutter, CA-CFAR can achieve the highest detection probability while maintaining a constant false alarm rate because the averaging of a large number of reference cells is the most effective estimate of the Rayleigh noise variance. Therefore, CA-CFAR is preferred in the long-range range to maximize the detection sensitivity for distant, weak targets.

[0063] In addition to the detector type, the three key parameters of CFAR are also adaptively set according to the interval characteristics: Reference window size determines the number of reference cells used for noise estimation. A larger reference window results in more stable noise estimation but is less sensitive to changes in local clutter characteristics and requires more computation. The adaptive principle for reference window size is: use a smaller reference window in densely populated areas to reduce the probability of multiple targets falling into the reference window, and use a larger reference window in sparsely populated areas to fully utilize the statistical advantages under uniform clutter conditions. The adaptive adjustment of the reference window size is not a simple choice between a large and a small window, but rather a continuous interpolation based on the target density index; the higher the density, the smaller the window, but the window will not be smaller than a preset lower limit nor larger than a preset upper limit.

[0064] Number of guard elements: Guard elements are elements around the target element that are excluded from the reference window and do not participate in noise estimation. Their function is to prevent the target's own energy from spreading through sidelobes and contaminating the noise estimation. The number of guard elements is set according to the size characteristics of the expected target within the interval: the target size varies greatly in the near-range interval, requiring a more flexible configuration of guard elements; the target in the far-range interval is close to the point target model, but considering the physical distance resolution corresponding to the bandwidth, moderate protection is also required.

[0065] Threshold factor: The threshold factor determines the ratio of the detection threshold to the noise estimate. A larger threshold factor results in a higher detection threshold, a lower false alarm rate, but also a lower detection rate; conversely, a smaller threshold factor has the opposite effect. The adaptive principle of the threshold factor is: use a higher threshold factor in regions with a large signal dynamic range to strongly suppress false detections caused by strong target sidelobes; moderately increase the threshold factor in regions with high target density to control the overall false alarm rate in dense environments; and use a lower threshold factor in regions with distant weak targets to improve detection sensitivity.

[0066] The specific value of the threshold factor is obtained by comprehensively considering the two inputs of dynamic range and density, by table lookup or interpolation, and this mapping relationship can be calibrated for a specific radar platform through offline simulation.

[0067] For each distance interval, CFAR detection is performed independently on the distance rows covered by that interval using the detector type selected for that interval and all configured parameters. The detection process for each interval is independent of each other; OS-CFAR in near-distance intervals does not affect the CA-CFAR threshold in far-distance intervals, and vice versa. Each interval outputs a set of detected points within that interval, with each point containing its distance index, velocity index, and signal amplitude.

[0068] Step S6: Perform overlapping transition processing on the boundary regions of adjacent processing intervals, merge the detection results of each processing interval and boundary region, and output the final target point set.

[0069] Furthermore, such as Figure 4As shown, step S6 involves overlapping transition processing of the boundary regions of adjacent processing intervals, including: A preset number of distance units are extended on both sides of the boundary of adjacent processing intervals to form an overlapping area; Within the overlapping area, the full set of constant false alarm rate detection parameters for the two adjacent processing intervals are used for independent detection, and the two sets of detection results are logically ORed and merged. The merged set of points is deduplicated based on distance tolerance and velocity tolerance. If the distance difference and velocity difference between two points are both less than the preset tolerance value, the two points are determined to correspond to the same target, and the one with the higher tolerance is retained.

[0070] Because steps S3 to S5 use different window functions and CFAR parameters for different distance intervals, two types of problems may occur at the partition boundaries: Boundary false negatives: The main lobe of a real target may span two adjacent intervals. Because different window functions are used in the two intervals, the shape and energy distribution of the main lobe on both sides of the boundary are inconsistent, which may cause the target to fail to reach its respective threshold in independent CFAR detection in both intervals; Boundary duplication detection: When the main lobe of the spectrum of the same target crosses the boundary, it may be detected by two intervals at the same time, generating duplicate traces representing the same target. If no processing is done, the subsequent tracking process will misclassify the target as two independent targets.

[0071] This step addresses the above issues by introducing a transition handling mechanism for overlapping boundary regions: For each pair of adjacent intervals, a certain range is extended to both sides of their partition boundary. The width of this range is set according to the distance resolution, usually taking 2 to 4 times the number of distance units corresponding to the distance resolution, forming an overlapping region. The physical meaning of this overlapping region is to cover the maximum range of possible morphological changes in the main lobe of the target at the adjacent boundary due to different window function processing, ensuring that any target affected by boundary effects within the overlapping region can be double-detected and covered. Within the overlapping region, instead of using the detection parameters of a single interval alone, a joint strategy is employed: The first step is to simultaneously perform detection using the complete detection parameters of the left interval and the complete detection parameters of the right interval within the overlapping area, resulting in two independent sets of detection results.

[0072] The second step is to combine the two sets of results using a logical "OR". Any point detected under either set of parameters is retained. In the boundary region, due to the variation of signal processing parameters, no single set of parameters can guarantee optimal detection of targets crossing the boundary. Therefore, dual detection is adopted and combined using "OR" to strategically prevent missed detections at the boundary.

[0073] The third step is to perform spatial deduplication on the merged set of points. If the distance position difference and velocity position difference between two points are both within a very small tolerance range (the distance tolerance is usually set to 2 distance units, and the velocity tolerance is usually set to 1 Doppler unit), then it is determined that they were generated by the same target being repeatedly detected under two different sets of parameters. The point with the higher signal amplitude is retained, and the other is deleted.

[0074] The detection results of the non-overlapping parts of each interval are combined with the results of the overlapping regions after deduplication and merging to generate the final set of points for the current frame. This set fully covers the entire detection range, and the detection quality within each interval has been specifically optimized through adaptive windowing and adaptive CFAR. The detection results at the interval boundaries have also undergone joint processing and deduplication for seamless integration.

[0075] The final set of detected points can be directly input into subsequent point clustering and multi-frame target tracking modules for further processing.

[0076] The technical solution of the present invention will be described in detail below with reference to an embodiment based on a typical 77GHz automotive forward millimeter-wave radar platform.

[0077] The radar system operating parameters of this embodiment are as follows: center frequency 77 GHz, operating bandwidth 1 GHz, each frame contains 256 linear frequency modulated (LFM) signals (i.e., 256 velocity dimension sampling points), and each LFM signal contains 1024 fast-time sampling points (i.e., 1024 range dimension sampling points). The corresponding theoretical range resolution is approximately 0.15 meters, and the maximum unambiguous detection range is approximately 153.6 meters. The following steps, S1 to S6, are described in detail, in conjunction with the processing of the same frame of measured data.

[0078] Step S1: Distance-dimensional FFT preprocessing and power distribution extraction; Step S1 acquires the raw echo data of the current frame, stored as a 1024-row × 256-column two-dimensional complex matrix. First, a 1024-point Hamming window is used as the initial window function. Windowing and 1024-point FFT processing are then applied to each of the 1024 sampling points of all 256 linear frequency modulated (LFM) signals. This step yields a 1024 × 256 one-dimensional range image matrix, where each row corresponds to a range cell, and each column corresponds to the spectral value of an LFM signal within that range cell. The Hamming window is chosen as the initial window function because it provides a relatively balanced trade-off between sidelobe suppression and main lobe width, providing an acceptable initial one-dimensional range image for all range intervals before subsequent partitioned windowing processing. This balanced characteristic of the Hamming window also keeps the required convolution correction within a small range when the window function is subsequently changed using frequency domain equivalence.

[0079] Subsequently, to extract the distribution characteristics of echo power along the range dimension, the power average value was calculated row-wise for the one-dimensional range image matrix. Specifically, the power of the corresponding 256 linear frequency modulated signal spectral values ​​was calculated for each range cell, and then the arithmetic mean of these 256 power values ​​was taken to obtain a one-dimensional echo power distribution vector of length 1024. This power distribution vector exhibits significant spatial non-uniformity in typical scenarios: the initial segment of the vector shows a significant power peak, reflecting the presence of a nearby strong reflective target; the power gradually decreases in the middle segment of the vector; and the latter segment of the vector is dominated by a noise floor, with the power value tending to flatten.

[0080] Step S2: Adaptive distance partitioning based on echo energy accumulation distribution; S2-1: Power Distribution Smoothing. A moving average smoothing process is applied to the echo power distribution vector to eliminate high-frequency jitter caused by random noise fluctuations and occasional power spikes in individual range cells. The moving average operation is performed in the time domain: a fixed-length sliding window moves point-by-point along the range dimension, and the smoothed value at each range cell is the arithmetic mean of the original power values ​​of all range cells within its window. For the boundary regions at both ends of the vector, a symmetrical extension method is used to complete the window data, avoiding boundary effects caused by insufficient data. The smoothed power distribution curve retains the macroscopic distribution trend of the echo energy while filtering out high-frequency random fluctuations, providing stable input data for subsequent calculations of the normalized cumulative energy curve.

[0081] S2-2: Calculate the normalized cumulative energy curve. The smoothed power distribution curve is accumulated point-by-point along the distance dimension to obtain the cumulative energy curve. Then, each point on the cumulative energy curve is divided by the total cumulative energy value of the entire curve to obtain the normalized cumulative energy curve, which ranges from zero to one. The normalized cumulative energy curve visually reflects the cumulative distribution characteristics of echo energy along the distance direction: in areas with concentrated strong echoes at close range, the curve rises rapidly; in areas with a distant noise floor, the curve tends to rise more gently.

[0082] S2-3: Determining Initial Candidate Boundaries for Partitioning Based on Energy Ratio Thresholds. Two energy ratio thresholds are set: a first energy ratio threshold and a second energy ratio threshold. In this embodiment, the first energy ratio threshold is 30%, and the second energy ratio threshold is 70%. The range cell indices corresponding to the first time the cumulative energy value reaches or exceeds these two thresholds on the normalized cumulative energy curve are found and used as the initial candidate boundaries for partitioning. In a typical scenario of this embodiment, the first energy ratio threshold corresponds to approximately the 195th range cell, and the second energy ratio threshold corresponds to approximately the 510th range cell. These two candidate boundaries initially divide the entire range dimension into three intervals: Interval 1 (near range), Interval 2 (medium range), and Interval 3 (far range). It should be noted that the specific values ​​of the above two energy ratio thresholds can be flexibly adjusted according to the actual application scenario and performance requirements of the radar. The 30% and 70% selected in this embodiment are only exemplary values.

[0083] S2-4: Local Fine-tuning of Boundaries. To avoid the partition boundary cutting exactly into the main lobe region of the target echo, which would cause the target energy to be split into two intervals and affect the accuracy of subsequent CFAR detection, local fine-tuning is performed on each initial candidate partition boundary. Specifically, a local search range is taken before and after the candidate boundary. Within this local search range, the absolute value of the first-order difference of the smoothed power curve is calculated, and the position with the smallest absolute value is taken as the final partition boundary. The principle is that the position where the power curve changes most gently indicates that the echo power at that position is in a stable low level, close to the noise floor. Partitioning at this position has the lowest probability of being cut into the target main lobe. In this embodiment, the local search range is preferably 8 distance units before and after the candidate boundary.

[0084] After the boundary fine-tuning in S2-4 above, the final partitioning result of this embodiment is as follows: Interval 1 (near range) corresponds to the 1st to the 203rd distance unit, covering a distance range of approximately 0 to 30.5 meters; Interval 2 (medium range) corresponds to the 204th to the 498th distance unit, covering a distance range of approximately 30.5 to 74.7 meters; Interval 3 (far range) corresponds to the 499th to the 1024th distance unit, covering a distance range of approximately 74.7 to 153.6 meters. The above partitioning results are generated driven by the echo energy distribution data of the current frame. The partitioning boundaries of different frames may vary with scene changes, reflecting the adaptive characteristics of this invention.

[0085] Step S3: Independent adaptive windowing processing for each distance partition; S3-1: Interval signal feature extraction. For each processing interval, based on the initial one-dimensional range profile obtained in step S1, two key feature parameters are calculated: signal dynamic range and target density.

[0086] The signal dynamic range is defined as the difference between the maximum and minimum echo power of all range cells within a processing interval, expressed in decibels (dB). It reflects the energy span between the strongest and weakest echo targets within that interval. A larger dynamic range indicates a more significant difference in the strength of target echoes within that interval, and a higher requirement for sidelobe suppression. In this embodiment, the signal dynamic range for interval one is 42 dB, for interval two it is 26 dB, and for interval three it is 15 dB.

[0087] Target density is defined as the proportion of the number of distance cells exceeding the initial detection threshold within a processing interval to the total number of distance cells in that interval. The initial detection threshold is determined as follows: the median of the power distribution vector obtained in step S1 is taken, and a preset offset is added to this median as the detection threshold. Distance cells exceeding this threshold are considered to potentially contain targets. The higher the target density, the denser the distribution of potential targets within the interval, and the higher the requirement for distance resolution. In this embodiment, the target density of interval one is 0.03, the target density of interval two is 0.08, and the target density of interval three is 0.01.

[0088] S3-2: Adaptive selection of window function. The window function is independently selected based on the characteristic parameters of each interval, according to the following decision logic: When the dynamic range of a signal in a processing interval is higher than a first preset value and the target density is lower than a second preset value (in this embodiment, interval one meets this condition: dynamic range 42dB > 30dB, density 0.03 < 0.05), it indicates that there are strong isolated targets in this interval, but the target distribution is sparse. At this time, a high sidelobe suppression window is selected; in this embodiment, a Chebyshev window is specifically used, with its sidelobe suppression level set to approximately -52dB relative to the strongest signal. The characteristic of the Chebyshev window is that, while ensuring a specified sidelobe suppression level, it minimizes the main lobe width. That is, under the same sidelobe suppression requirements, the main lobe of the Chebyshev window is the narrowest, which is crucial for suppressing the leakage of sidelobes of strong near-range targets into the far-range interval.

[0089] When the dynamic range of a signal in a processing interval is lower than a first preset value and the target density is higher than a second preset value (in this embodiment, interval two meets this condition: dynamic range 26dB < 30dB, density 0.08 > 0.05), it indicates that the target distribution in this interval is dense but the intensity difference is not significant. In this case, a narrow main lobe window is selected; in this embodiment, a Hanning window is specifically chosen. The characteristics of the Hanning window are: its main lobe width is approximately 1.62 times that of a rectangular window, and its side lobes attenuate relatively quickly (side lobes attenuate at a rate of -18dB per octave), which is beneficial for separating nearby targets with small spacing and is suitable for scenarios with densely distributed multiple targets in the mid-range region.

[0090] When the signal dynamic range of a processing interval is lower than a first preset value and the target density is lower than a second preset value (in this embodiment, interval three meets this condition: dynamic range 15dB < 30dB, density 0.01 < 0.05), the environment in this interval is relatively stable and does not require extreme sidelobe suppression or an extremely narrow main lobe. At this time, the Hamming window used in step S1 can be retained, and there is no need to perform a window function replacement operation.

[0091] It should be noted that the specific values ​​of the first preset value (30dB in this embodiment) and the second preset value (0.05 in this embodiment) can be adjusted according to the radar system parameters and application requirements. The values ​​listed in this embodiment are exemplary preferred parameters that have been verified by engineering. In addition, when a processing unit simultaneously satisfies high dynamic range and high target density, an adjustable balanced window function such as the Kaiser window can be selected, whose parameters can adjust the trade-off between main lobe width and side lobe suppression.

[0092] S3-3: Frequency Domain Equivalent Window Function Replacement. For the intervals where the window function needs to be replaced (interval one and interval two in this embodiment), this invention uses a frequency domain equivalent method to replace the window function, avoiding returning to the time domain to re-execute the complete distance-dimensional FFT.

[0093] The principle of frequency domain equivalent replacement is as follows. Let the time-domain representation of the initial window function (Hamming window) be wi (where the subscript i represents the i-th sampling point), and the time-domain representation of the target window function (such as a Chebyshev window or Hanning window) be wt. Directly replacing the window function requires multiplying the original echo data point by point by the target window function wt in the time domain and then re-performing the FFT. The computational complexity is proportional to the number of sampling points in the interval multiplied by the computational complexity of the FFT. The frequency domain equivalent method utilizes a fundamental property of the Fourier transform: time-domain windowing is equivalent to frequency-domain convolution. Specifically, multiplying the original echo data by the target window function wt in the time domain is mathematically equivalent to convolving the FFT result windowed with the initial window function wi in the frequency domain with an equivalent convolution kernel. This equivalent convolution kernel is determined by the ratio of the frequency domain response of the target window function wt to the frequency domain response of the initial window function wi. Since both the target window function and the initial window function are known and fixed sequences of window functions, the equivalent convolution kernel can be pre-calculated and stored. In actual frame-by-frame processing, it is only necessary to read the equivalent convolution kernel from the storage and perform local frequency domain convolution correction on each distance unit in the interval where the window function needs to be changed.

[0094] Step S4: Perform FFT processing based on the velocity dimension of the optimized range image; This step corresponds to step S4 in independent claim 1. The optimized one-dimensional range image obtained in step S3 is processed using FFT along the velocity dimension. Specifically, for each range cell in the optimized one-dimensional range image matrix corresponding to 256 linear frequency modulated signal data points, windowing (in this embodiment, a Hamming window is uniformly used along the velocity dimension) and 256-point FFT processing are performed to generate a 1024×256 range-Doppler two-dimensional spectrum matrix. The row indices of this two-dimensional spectrum matrix correspond to range cells, and the column indices correspond to Doppler cells. This completes the full conversion from the one-dimensional range dimension to the two-dimensional range-Doppler domain, preparing for subsequent partitioned CFAR detection. It should be noted that windowing in the velocity dimension is also to suppress spectral leakage in the Doppler dimension. In this embodiment, a Hamming window is uniformly used in the velocity dimension, but other window functions can be selected according to actual needs.

[0095] Step S5: Independent adaptive CFAR detection for each distance partition; S5-1: Detector type allocation. Different CFAR detectors are assigned based on the location of each interval in the distance dimension and the corresponding clutter characteristics.

[0096] Interval 1 (near range, corresponding to rows 1 to 203 of the two-dimensional spectrum in this embodiment) is assigned an ordered statistics constant false alarm rate detector (OS-CFAR). The OS-CFAR works by arranging the energy values ​​of all reference cells within the reference window in ascending order, selecting the energy value at a preset sorting position (i.e., the Kth ordered statistic) as an estimate of the current background noise level, and then multiplying this noise estimate by a threshold factor to obtain the detection threshold. OS-CFAR has a natural resistance to occasional strong outlier interference in heavy-tailed clutter (such as isolated strong echoes caused by roadside metal reflectors commonly found in near range areas). Even if a few reference cells are contaminated by outliers, as long as the number of outliers does not exceed the number of cells beyond the sorting position in the total number of reference cells, the noise estimation will not be substantially affected by the outliers. In this embodiment, the reference window of Interval 1 contains 16 cells in the distance dimension and 12 cells in the velocity dimension (neither excluding guard cells), and the selected sorting position is approximately 75th percentile (i.e., the Kth ordered statistic is located at approximately the 0.75th percentile of the total number of reference cells).

[0097] Interval two (the mid-range region, corresponding to rows 204 to 498 of the two-dimensional spectrum in this embodiment) is assigned a reduced mean constant false alarm rate detector (CMLD-CFAR). The CMLD-CFAR works by first sorting the energy values ​​of all reference cells within the reference window, then removing a certain percentage (25% in this embodiment) of the reference cells with the highest energy values. Finally, the arithmetic mean of the energy values ​​of the remaining reference cells is used as the noise estimate. In this embodiment, the reference window in interval two contains 24 cells in the distance dimension and 16 cells in the velocity dimension.

[0098] Interval 3 (the far-range region, corresponding to rows 499 to 1024 of the two-dimensional spectrum in this embodiment) is allocated with a constant false alarm rate (CA-CFAR) detector. CA-CFAR directly takes the arithmetic mean of the energy values ​​of all reference cells within the reference window as the noise estimate. Clutter in the far-range region is mainly receiver thermal noise, following a uniform Rayleigh distribution. In this case, taking the mean of all reference cells yields the minimum variance unbiased estimate of the noise power, making CA-CFAR theoretically the optimal detector. In this embodiment, the reference window in Interval 3 contains 32 cells in the range dimension and 20 cells in the velocity dimension. The reference window size in the far-range region is larger than that in the near-range and mid-range regions because the clutter environment in the far-range region is more uniform. A larger reference window is beneficial for improving the noise estimation accuracy, thereby maximizing the detection sensitivity for weak targets at long distances.

[0099] S5-2: Adaptive Threshold Factor Configuration. The threshold factor determines the multiple of the detection threshold relative to the noise estimation, directly affecting the false alarm rate and detection rate. In this invention, the threshold factor is adaptively adjusted according to the target density of each interval. The basic adjustment principle is as follows: for intervals with higher target density, the threshold factor is appropriately increased to control the false alarm rate and prevent cascading false alarms caused by the reference window being contaminated by target energy and thus increasing the noise estimation in dense target scenarios; for intervals with lower target density, the threshold factor is appropriately decreased to improve the detection sensitivity for weak targets. In this embodiment, the threshold factor for interval 1 (target density 0.03) is approximately 14dB, the threshold factor for interval 2 (target density 0.08) is approximately 12dB, and the threshold factor for interval 3 (target density 0.01) is approximately 9dB. The specific values ​​of the above threshold factors are exemplary preferred parameters obtained through system calibration, and can be adjusted according to the noise figure of the radar system and the working environment in actual deployment.

[0100] S5-3: Adaptive Configuration of Reference Window Size and Protection Units. The size of the reference window in both the range and velocity dimensions is adaptively configured based on the clutter characteristics and target distribution features of each interval. Generally, in intervals with more uniform clutter (e.g., long-range intervals), the reference window size is increased to fully capture clutter statistical information and improve noise estimation accuracy; in intervals with denser targets (e.g., mid-range intervals), the reference window size is moderate to avoid too many nearby targets entering the reference window; in intervals with non-uniform clutter and sparse targets (e.g., short-range intervals), the reference window size is appropriately reduced to maintain the locality of noise estimation. The protection unit prevents target main lobe energy leakage into the reference window, thus improving noise estimation. In this embodiment, the number of protection units in each interval is set to approximately one-quarter of the size of its respective reference window (in the range dimension), and protection units are set in both the range and velocity dimensions.

[0101] S5-4: Independent Detection Execution. Following the detector type and parameters independently determined for each interval in S5-1 to S5-3, CFAR detection is performed on the corresponding range-Doppler two-dimensional spectral range for each interval. The detection result for each interval is a list of points within that interval, with each point containing three attributes: distance index, velocity index, and amplitude value.

[0102] Step S6: Process the overlapping transition of partition boundaries and merge the full-distance detection results; S6-1: Define the overlapping region. A preset number of distance units are extended on both sides of the boundary of adjacent intervals to form an overlapping region. In this embodiment, the extension number is 10 distance units (approximately 1.5 meters), therefore the width of each overlapping region is 20 distance units. Taking the first boundary between interval one and interval two (distance unit 203) as an example, the overlapping region covers distance units 194 to 213. The setting of this preset extension number needs to comprehensively consider the following factors: if the extension number is too small, it may not be able to fully cover the target detection differences on both sides of the boundary caused by windowing and CFAR parameter differences; if the extension number is too large, it will introduce excessive redundant calculations. The 10 distance units selected in this embodiment are an exemplary preferred value under typical vehicle radar parameters.

[0103] S6-2: Joint Detection of Overlapping Regions. Within overlapping regions, independent detection is performed using the complete set of CFAR detection parameters for the two adjacent intervals, including detector type, reference window size, number of guard cells, and threshold factor, resulting in two sets of detection results. These two sets of detection results are then logically ORed and merged; that is, if either set of detection parameters determines that a certain range-Doppler cell is a target, it is included in the candidate point set.

[0104] S6-3: Deduplication and Merging. The candidate point set after logical OR merging within the overlapping area may contain duplicate points where the same target is detected simultaneously by two sets of parameters, resulting in similarly located points. Therefore, spatial deduplication is required for the candidate point set. The deduplication rule is: if the distance index difference between two points is less than or equal to a preset distance tolerance, and the velocity index difference is less than or equal to a preset velocity tolerance, then the two points are determined to correspond to the same target. For multiple points determined to be the same target, the one with the highest amplitude is retained as the representative point, and the rest are discarded. In this embodiment, the distance tolerance is preferably set to 2 distance units (approximately 0.3 meters), and the velocity tolerance is preferably set to 1 velocity unit. The deduplication tolerance setting must match the radar system parameters: the distance tolerance value should be greater than the distance resolution (approximately 0.15 meters) to avoid misjudging the same target as two different targets due to slight positional jitter caused by different detection parameters; at the same time, the tolerance value should be less than the minimum distance between typical targets to avoid incorrectly merging the points of two truly adjacent targets.

[0105] S6-4: Summary of Full-Range Detection Results. The following components are summarized to form the final target point set output: the detection results of interval 1 after removing overlapping areas; the detection results of the first overlapping area (the boundary between interval 1 and interval 2) after deduplication and merging; the detection results of interval 2 after removing the overlapping areas on both sides; the detection results of the second overlapping area (the boundary between interval 2 and interval 3) after deduplication and merging; and the detection results of interval 3 after removing the overlapping area. The final output point set covers all detected targets within the full range, and the detection at the boundaries is continuous, without repetition, and without omission.

[0106] The present invention has been described in detail above with reference to specific embodiments. It should be noted that the specific values ​​listed in the above embodiments (such as tolerances for 30dB, 0.05, 10 range cells, 2 range cells, and 1 velocity cell), window function types, and detector types are all exemplary preferred parameters and do not constitute a limitation on the scope of protection of the present invention. For example, in addition to the Hamming window, other window functions with good balance characteristics such as the Blackman window or the Kaiser window can also be selected for the initial window function; the energy ratio threshold can be flexibly adjusted according to the application scenario; the number of partitions is not limited to three, and can be divided into two or four or more intervals according to the complexity of the echo energy distribution; the selection of CFAR detectors for different intervals can also be based on the actual clutter characteristics, and other types of constant false alarm rate detectors can be selected.

[0107] Secondly, this application proposes a partitioned adaptive windowing constant false alarm rate millimeter-wave radar target detection enhancement system, such as... Figure 5 As shown, it includes: The data preprocessing module is used to acquire raw radar echo data, perform initial windowing and range-dimensional Fourier transform on the raw echo data, and generate a one-dimensional range profile and echo power distribution curve. The interval division module, connected to the data preprocessing module, is used to adaptively divide the distance dimension into multiple processing intervals based on the energy accumulation distribution characteristics of the echo power distribution curve. The partitioned windowing optimization module, connected to the interval division module, is used to independently select a window function and perform windowing processing on each processing interval based on the signal characteristics within the processing interval, so as to optimize the one-dimensional distance image of each interval. The two-dimensional spectrum generation module, connected to the partition windowing optimization module, is used to perform a velocity-dimensional Fourier transform on the optimized one-dimensional range image of each interval to generate a range-Doppler two-dimensional spectrum. The partitioned constant false alarm rate (CFAR) detection module is connected to the two-dimensional spectrum generation module and the interval division module. It is used to independently perform CFAR detection on each processing interval on the distance-Doppler two-dimensional spectrum, wherein each processing interval uses a CFAR detector type and adaptive parameters determined by the clutter characteristics of the interval. The boundary fusion and output module is connected to the partition constant false alarm detection module and the interval division module. It is used to perform overlapping transition processing on the boundary regions of adjacent processing intervals, merge the detection results of each processing interval and boundary region, and output the final target point set.

[0108] Thirdly, this application proposes an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0109] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described above.

[0110] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0111] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0112] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0113] In the embodiments provided in this disclosure, it should be understood that the disclosed apparatus / computer devices and methods can be implemented in other ways. For example, the apparatus / computer device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. Multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0114] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0115] Furthermore, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0116] If an integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. A computer-readable medium may include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in a computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0117] The above are merely preferred embodiments of the present invention. It should be noted that any modifications and improvements made by those skilled in the art without departing from the present technical solution should also be considered to fall within the scope of protection claimed by the present solution.

Claims

1. A method for enhancing target detection in millimeter-wave radar with partitioned adaptive windowing and constant false alarm rate, characterized in that: Includes the following steps: Step S1: Acquire the raw radar echo data and perform initial windowing and range-dimensional Fourier transform to obtain a one-dimensional range profile and echo power distribution curve; Step S2: Based on the energy accumulation distribution characteristics of the echo power distribution curve, the distance dimension is adaptively divided into multiple processing intervals; Step S3: For each processing interval, independently select a window function based on the signal characteristics within the processing interval and perform windowing processing to optimize the one-dimensional distance image of each interval; Step S4: Perform velocity-dimensional Fourier transform on the optimized one-dimensional distance image of each interval to generate a distance-Doppler two-dimensional spectrum; Step S5: Perform constant false alarm rate detection independently for each processing interval on the range-Doppler two-dimensional spectrum, wherein each processing interval uses a constant false alarm rate detector type and adaptive parameters determined by the interval clutter characteristics; Step S6: Perform overlapping transition processing on the boundary regions of adjacent processing intervals, merge the detection results of each processing interval and boundary region, and output the final target point set.

2. The method according to claim 1, characterized in that, In step S2, based on the energy accumulation distribution characteristics of the echo power distribution curve, the distance dimension is adaptively divided into multiple processing intervals, including: Smooth the echo power distribution curve; Calculate the normalized cumulative energy curve of the smoothed power distribution curve; Based on a preset energy ratio threshold, the initial partition boundary is determined on the normalized cumulative energy curve. Within the local search range near the initial partition boundary, find the location with the most gradual power change, and fine-tune the location with the most gradual power change as the final partition boundary.

3. The method according to claim 2, characterized in that, The signal characteristics on which the independent selection of the window function in step S3 is based include the signal dynamic range and target density within the processing interval. Among them, the signal dynamic range refers to the difference between the maximum and minimum echo power within the processing interval, and the target density refers to the proportion of the number of distance units exceeding the initial detection threshold within the processing interval to the total number of distance units in the processing interval. When the signal dynamic range of a processing interval is higher than the first preset value and the target density is lower than the second preset value, a high sidelobe suppression window is selected. When the signal dynamic range of a processing interval is lower than the first preset value and the target density is higher than the second preset value, a narrow main lobe window is selected.

4. The method according to claim 3, characterized in that, The high sidelobe suppression window is a Chebyshev window, and the narrow main lobe window is a Hanning window.

5. The method according to claim 4, characterized in that, In step S5, when performing constant false alarm rate detection independently for each processing interval, an ordered statistic constant false alarm rate detector is used for the near-distance processing interval; a pruned mean constant false alarm rate detector is used for the medium-distance processing interval; and a cell average constant false alarm rate detector is used for the far-distance processing interval.

6. The method according to claim 5, characterized in that, The adaptive parameters for constant false alarm rate detection include reference window size, number of protection units, and threshold factor. The adaptive parameters are adaptively adjusted according to the target density and signal dynamic range of the processing interval.

7. The method according to claim 6, characterized in that, Step S6 involves overlapping transition processing of the boundary regions of adjacent processing intervals, including: A preset number of distance units are extended on both sides of the boundary of adjacent processing intervals to form an overlapping area; Within the overlapping area, the full set of constant false alarm rate detection parameters for the two adjacent processing intervals are used for independent detection, and the two sets of detection results are logically ORed and merged. The merged set of points is deduplicated based on distance tolerance and velocity tolerance. If the distance difference and velocity difference between two points are both less than the preset tolerance value, the two points are determined to correspond to the same target, and the one with the higher tolerance is retained.

8. The method according to claim 7, characterized in that, The windowing process in step S3 is implemented through a frequency domain equivalent method. Specifically, it uses the frequency domain equivalent convolution kernel between the target window function and the initial window function in step S1 to perform local convolution correction on the data in the corresponding processing interval of the one-dimensional distance image obtained in step S1. The window function used in the initial windowing in step S1 is a Hamming window.

9. The method according to claim 8, characterized in that, The echo power distribution curve in step S1 is obtained by averaging the echo energy of each range cell in the one-dimensional range image on all linear frequency modulated signals to obtain the one-dimensional power distribution vector.

10. A partitioned adaptive windowing constant false alarm rate millimeter-wave radar target detection enhancement system, characterized in that, include: The data preprocessing module is used to acquire raw radar echo data, perform initial windowing and range-dimensional Fourier transform on the raw echo data, and generate a one-dimensional range profile and echo power distribution curve. The interval division module, connected to the data preprocessing module, is used to adaptively divide the distance dimension into multiple processing intervals based on the energy accumulation distribution characteristics of the echo power distribution curve. The partitioned windowing optimization module, connected to the interval division module, is used to independently select a window function and perform windowing processing on each processing interval based on the signal characteristics within the processing interval, so as to optimize the one-dimensional distance image of each interval. The two-dimensional spectrum generation module, connected to the partition windowing optimization module, is used to perform a velocity-dimensional Fourier transform on the optimized one-dimensional range image of each interval to generate a range-Doppler two-dimensional spectrum. The partitioned constant false alarm rate (CFAR) detection module is connected to the two-dimensional spectrum generation module and the interval division module. It is used to independently perform CFAR detection on each processing interval on the distance-Doppler two-dimensional spectrum, wherein each processing interval uses a CFAR detector type and adaptive parameters determined by the clutter characteristics of the interval. The boundary fusion and output module is connected to the partition constant false alarm detection module and the interval division module. It is used to perform overlapping transition processing on the boundary regions of adjacent processing intervals, merge the detection results of each processing interval and boundary region, and output the final target point set.

Images