Traffic radar-based high-capture-rate target detection method, device and storage medium

Abstract

The application discloses a high-capture-rate target detection method and device based on a traffic radar and a storage medium, and relates to the technical field of radar data processing.The method comprises the following steps: dividing a detection matrix into a first region and a second region; performing beam synthesis and detection on the second region, performing detection on the first region, fusing the detection results, and outputting a preliminary detection point cloud; generating a speed hypothesis based on the preliminary detection point cloud, and generating multiple groups of candidate signals; performing angle estimation processing on each group of candidate signals based on the principle of a differential virtual array; calculating the peak amplitude of each angle spectrum, taking the angle value corresponding to the angle spectrum with the maximum peak amplitude as a corrected angle; and performing cross-frame matching and speed deblurring processing on fast and slow frame point clouds of different beam types generated alternately by the radar in a specific time sequence.The application solves the problems of insufficient far-field angle resolution, inaccurate information caused by speed ambiguity, and difficult point cloud matching and fusion in traffic radar detection.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of radar data processing, and particularly relates to a high-capture-rate target detection method, device and storage medium based on a traffic radar. BACKGROUND

[0002] With the rapid development of intelligent transportation systems, accurate, real-time and reliable detection of road traffic has become a key technical support for intelligent traffic control, road network planning and optimization decision-making. Millimeter wave radar, especially high-frequency millimeter wave radar with a working frequency near 90GHz, has become one of the core sensors for fixed-position traffic flow detection due to its high resolution, small weather influence and all-weather working advantages. Its core task is to continuously and stably output three-dimensional target point clouds containing accurate distance, speed and angle information, providing a high-quality data source for subsequent target tracking, trajectory analysis and traffic state research.

[0003] In the prior art, to improve the detection performance of traffic radar, especially the capture ability (i.e., "high capture rate") of long-distance and weak-reflection-intensity targets, two technical paths are mainly developed:

[0004] The first type of scheme focuses on enhancement at the signal detection level. For example, algorithm optimization is performed at the constant false alarm rate (CFAR) detection link, or digital beam forming (DBF) technology is used to superimpose the gain of the received channel signals to realize non-coherent accumulation of weak target energy, thereby improving the detection probability in a single frame of data.

[0005] For example, the patent document with publication number CN119414339A discloses a weak target DBF enhanced detection method for a millimeter wave traffic radar, which acquires more targets by setting an enhanced area and performing multiple DBF scans. However, this type of method mainly focuses on acquiring more original points or preliminary point clouds in the radar detection matrix, and the output target information (especially angle and speed) often has limited accuracy or is ambiguous, and cannot be directly used for subsequent accurate three-dimensional parameters. These preliminary point clouds as intermediate results are difficult to directly ensure the output of continuous, stable and accurate final detection point clouds in complex traffic scenarios.

[0006] The second type of solution expands the performance boundary by improving the hardware waveform and data fusion strategy of the radar. A common practice is to use a wide beam and a narrow beam alternately in the mode of transmission, aiming to balance the large field of view coverage and the high-precision detection at a long distance. However, such a solution faces significant challenges in actual deployment: on the one hand, due to the huge differences in beam shape, resolution and effective detection distance between the wide beam and the narrow beam, the spatio-temporal alignment and accurate fusion algorithm of the point cloud data is extremely complex. If the fusion algorithm is time-consuming, it will not meet the stringent real-time requirements of the traffic radar for high refresh rate (usually tens of milliseconds level output), causing system performance bottleneck. On the other hand, the detection logic of the existing radar system is not well adapted to the road scene. Many systems have fixed and idealized lane detection models (such as only supporting symmetric rectangular regions), which are difficult to adapt to complex road geometries such as curved roads and channelized intersections, resulting in inaccurate detection region settings and a large number of invalid operations or missed detections.

[0007] In summary, the existing technology has the following systematic deficiencies that need to be solved:

[0008] (1) Disconnection between parameter accuracy and integrity: Existing weak target enhancement schemes mostly output "intermediate results" at the detection stage, and fail to systematically solve the core problems of velocity ambiguity correction and high-precision angle estimation in the final output point cloud, resulting in insufficient accuracy and integrity of the three-dimensional information (distance, velocity, angle) of the point cloud.

[0009] (2) Conflict between fusion performance and real-time performance: Multi-source data fusion strategies such as wide and narrow beams are limited by algorithm complexity, making it difficult to meet the high real-time output requirements of traffic radar while ensuring fusion accuracy.

[0010] (3) Lack of scene adaptability and self-adaptation capability: The detection logic lacks deep integration with dynamic and configurable real road models, and cannot be adaptively adjusted according to actual lane parameters, resulting in low intelligence and limiting the optimal detection performance in different road conditions.

[0011] Therefore, there is an urgent need for an innovative, system-level detection method that not only effectively improves the detection sensitivity of weak targets, but also ensures that the final output point cloud has accurate and non-ambiguous three-dimensional parameters, and can adapt to complex road scenes, while ensuring high refresh rate of data output through an efficient fusion mechanism, thereby truly realizing high capture rate and high reliability detection of traffic radar in complex environments. SUMMARY

[0012] In view of the above defects in the prior art, the purpose of the present application is to provide a traffic radar-based high-capture-rate target detection method, device and storage medium, to systematically solve the problem of how to adaptively improve weak target detection capability without reducing radar data refresh rate in complex road scenes, and finally output a three-dimensional target point cloud with accurate, complete and continuous parameters.

[0013] The present application solves the above technical problems by the following technical solutions: a traffic radar-based high-capture-rate target detection method, comprising:

[0014] According to the road geometric parameters, the detection matrix generated based on the radar echo signal is divided into a first region and a second region;

[0015] The data in the second region is subjected to beam synthesis and detection based on road geometric parameters, while the signals in the first region are subjected to detection;

[0016] The detection results of the second region and the first region are fused, and a preliminary detection point cloud containing target distance and speed information is output;

[0017] Based on the speed information of each target in the preliminary detection point cloud and the maximum unambiguous speed of the radar, at least one speed hypothesis is generated, and based on the speed hypothesis, a plurality of groups of candidate signals are generated through Doppler phase compensation;

[0018] Each group of candidate signals is subjected to high-resolution angle estimation processing based on the principle of differential virtual array, to obtain a plurality of angle spectra;

[0019] The peak amplitudes of each angle spectrum are calculated, and the angle value corresponding to the angle spectrum with the largest peak amplitude is taken as the corrected angle of the target, thereby obtaining a detection point cloud containing target distance, speed information and corrected angle;

[0020] Based on the target distance and the corrected angle, the fast and slow frame point clouds of different beam types alternately generated by the radar at a specific time sequence are subjected to cross-frame matching and speed deblurring processing, and the final target point cloud with accurate parameters is output after fusion.

[0021] The application introduces road geometry parameters as core inputs, so that the radar detection logic is changed from a fixed program to a dynamically configurable intelligent strategy. The radar can accurately divide the far distance area (i.e. the second area) that needs to be enhanced and the near distance area (i.e. the first area) that needs to be processed regularly in the detection matrix according to the actual lane width, direction and length. This not only makes the detection range fit the real road profile and completely gets rid of the constraints of the symmetric rectangle, but more importantly, it realizes on-demand allocation of computing power: only the far distance weak target area with low signal-to-noise ratio and the most need for improvement is subjected to the adaptive beam synthesis processing with large amount of calculation, while the near distance area with strong target is subjected to efficient regular detection. In this way, while significantly improving the weak target detection ability in the far distance, the huge calculation redundancy brought by full-range beam synthesis is effectively avoided, the processing efficiency is guaranteed from the source, and a solid foundation is laid for meeting the high refresh rate requirement.

[0022] The application effectively suppresses the high sidelobe interference caused by the physical sparse array through high-resolution angle estimation processing based on the differential virtual array principle, and obtains the initial angle estimation with sharp main lobe and high resolution. The correction process of "generating multiple groups of candidate signals, obtaining multiple angle spectra through parallel processing, and comparing peak values to make decisions" is creatively designed. This mechanism fundamentally corrects the systematic angle error introduced by speed ambiguity, ensuring that the corrected angle output has high precision and high reliability. The distance-speed information of the preliminary detection point cloud is associated with the original array signal, and the accurate angle is calculated through signal processing closed loop, so that the complete and reliable "distance-speed-angle" three-dimensional target information is output.

[0023] Since the angle used for cross-frame matching is a high-precision value that has been corrected, the trial calculation and ambiguity in the matching process are greatly reduced, making the association of fast and slow frame point clouds fast and accurate, and laying a solid foundation for subsequent speed de-ambiguity. The processing object of the application is the fast and slow frame point clouds generated alternately in a specific time sequence, indicating that the underlying system uses efficient system architectures such as sliding frame cache and pipeline scheduling, making it possible to stably and quickly fuse multi-beam data and output high frame rate point clouds under limited resources, thereby solving the fundamental contradiction between fusion time consumption and system refresh rate.

[0024] In the application, the adaptive detection provides accurate "distance-speed" guidance information for angle correction, and the high-precision corrected angle is the fundamental prerequisite for efficient fusion to be completed quickly and accurately. Ultimately, it ensures that targets from near to far and from strong to weak can be stably perceived, accurately measured and continuously tracked, and high-quality target trajectories are continuously output in complex traffic scenarios, thereby realizing systematic and high-reliability improvement of the capture rate.

[0025] Further, according to the road geometry parameters, the detection matrix generated based on the radar echo signal is divided into a first area and a second area, comprising:

[0026] performing Fourier transform on the radar echo signal in the range dimension and the velocity dimension to generate a range-Doppler detection matrix;

[0027] setting a range division threshold according to a lane length in the road geometry parameter;

[0028] dividing a region with a range dimension index less than or equal to the range division threshold in the range-Doppler detection matrix into the first region, and dividing a region with a range dimension index greater than the range division threshold into the second region.

[0029] The present application divides the detection matrix into a near field (first region) and a far field (second region) by setting a range division threshold, allowing the use of differentiated detection algorithms for the two regions. This avoids uniformly executing algorithms with high computational complexity (such as road geometry-based beam synthesis) on the entire detection matrix, significantly reducing the overall computational load and improving system processing speed and real-time performance.

[0030] The use of more direct or more refined detection methods in the first region (near field) can better cope with the characteristics of dense target distribution, high signal-to-noise ratio, and complex multipath and occlusion effects in the near field region, helping to reduce missed detection and false alarms and improve the detection reliability of key near-field targets. Road geometry-based beam synthesis and detection, which are computationally intensive, are focused on the second region (far field). Far-field targets require high angular resolution, and road geometry constraints are strong. This division ensures high-precision and targeted angle estimation and detection for far-field targets, while avoiding unnecessary complex operations in the near field, achieving the best match between algorithm resources and scene requirements.

[0031] Further, performing road geometry parameter-based beam synthesis and detection on data in the second region, comprising:

[0032] calculating an angle scanning range based on a lane width in the road geometry parameter and a preset range division threshold;

[0033] generating a set of discrete beam steering vectors according to the angle scanning range;

[0034] performing beam synthesis on multi-channel complex signals extracted from the second region using the beam steering vectors to obtain complex signals in multiple beam directions;

[0035] performing modulus square operation on each complex signal in the beam direction to obtain signal power;

[0036] performing non-coherent accumulation on the signal power within a coherent processing interval to obtain accumulated energy distribution;

[0037] Perform constant false alarm rate detection on the accumulated energy distribution, and output a target detection result of the second region.

[0038] The application dynamically calculates the angle scanning range based on prior information of lane width, and generates discrete beam steering vectors in the range to perform focused beam synthesis, so that limited signal processing energy can be concentrated on the spatial direction where the target is most likely to appear. This greatly suppresses noise and interference in irrelevant directions, thereby obtaining higher angle resolution and detection signal-to-noise ratio in the far field region.

[0039] Compared with traditional full-space scanning, limiting the angle scanning range means that the number of beams to be generated and calculated is significantly reduced. At the same time, subsequent non-coherent accumulation further smooths the noise fluctuation and enhances the continuity of the target. Combined with constant false alarm rate detection, the flow greatly reduces the overall operation burden and false alarm probability of the system under the premise of ensuring detection sensitivity, and realizes efficient and reliable far field target detection.

[0040] The application deeply integrates the key prior knowledge of road geometry into the signal processing chain. By directly converting parameters such as lane width and distance threshold into constraint conditions for beam synthesis, the radar processing is highly adapted to specific road scenes, enhancing the self-adaptation ability in different road environments (such as different numbers of lanes and curves), and improving the practicality and reliability of the detection results.

[0041] Further, according to the angle scanning range, a group of discrete beam steering vectors are generated, including:

[0042] Within the angle scanning range, K discrete angle points are selected at a preset interval;

[0043] According to the physical position of the radar receiving antenna array, the beam steering vector of each discrete angle point is calculated.

[0044] The application can flexibly adjust the angle resolution and computational load of beam synthesis by selecting discrete angle points at a preset and controllable interval within the limited angle scanning range. Smaller intervals can achieve finer angle coverage, and reasonable interval settings can avoid generating redundant beams while ensuring lane coverage accuracy, significantly improving operation efficiency.

[0045] The steering vector of each angle is calculated based on the real physical position of the radar receiving antenna array, which strictly follows the basic principles of array signal processing and ensures the accuracy of beam synthesis in spatial direction. At the same time, the application method is adapted to any configuration (such as uniform or sparse) of the receiving array, enhancing the generality and engineering practicability of the algorithm.

[0046] Further, based on the speed information of each target in the preliminary detection point cloud and the maximum unambiguous speed of the radar, at least one speed hypothesis is generated, including:

[0047] obtaining the speed information of each target in the preliminary detection point cloud and the maximum unambiguous speed of the radar ;

[0048] generating the at least one speed hypothesis according to the speed information of each target in the preliminary detection point cloud and the maximum unambiguous speed of the radar :

[0049] if the speed information is greater than 0, the speed hypothesis is ;

[0050] if the speed information is less than or equal to 0, the speed hypothesis is ;

[0051] and the speed information of each target in the preliminary detection point cloud is directly taken as the speed hypothesis.

[0052] The present application is directly aimed at the speed folding phenomenon caused by the maximum unambiguous speed limit inherent to the radar. By a simple and certain mathematical rule (±2 shift according to the original speed sign), a most likely real speed candidate value (i.e. speed hypothesis) is generated for each target that may have speed ambiguity. This provides a clear and efficient entry point for solving the speed ambiguity problem. This generation rule only depends on the original speed sign, completely avoids complex pattern recognition or iterative search, and the logic is extremely simple, certain, and has no ambiguity or branch nesting, which makes it stable and reliable to execute under various signal-to-noise ratios and scenes, and has strong engineering robustness. The generated speed hypothesis is not the final result, but provides a unique and necessary input for subsequent Doppler phase compensation. This step is a key bridge connecting "speed detection" and "high-resolution angle estimation", ensuring that subsequent precise algorithms such as differential virtual array can perform phase correction under the correct speed hypothesis, thereby ultimately achieving accurate angle calculation.

[0053] Further, based on the speed hypothesis, a plurality of candidate signals are generated through Doppler phase compensation, including:

[0054] for each speed hypothesis generated for each target, a corresponding Doppler phase compensation value is calculated;

[0055] using the Doppler phase compensation values, the original multi-channel received signals are respectively phase compensated;​

[0056] wherein each phase-compensated signal constitutes a candidate signal of the plurality of candidate signals, and the original multi-channel received signal is a signal extracted from the probe matrix according to the distance and velocity information of the target.

[0057] In the embodiment, the core is to convert the abstract velocity hypothesis into a specific "phase compensation value" that can act on the received signal through the Doppler effect formula. This enables the phase error between the received channels caused by velocity ambiguity to be accurately corrected in the digital domain, laying a precise mathematical and physical foundation for subsequent recovery of correct spatial angle information using array signal processing techniques.

[0058] A plurality of phase-compensated candidate signals are generated in parallel by calculating a plurality of phase compensation values for the same target velocity hypothesis. This essentially pre-corrects the phase distortion of each possible velocity scenario at the signal level, so that the subsequent angle estimation module can objectively determine the velocity hypothesis that best fits the actual situation through comparison (such as peak amplitude) on a fair signal basis, greatly improving the accuracy and reliability of velocity de-ambiguity.

[0059] Further, high-resolution angle estimation processing based on the principle of differential virtual array is performed on each candidate signal, including:

[0060] For each candidate signal, a corresponding spatial autocorrelation matrix is calculated;

[0061] The spatial autocorrelation matrix is subjected to a vectorization operation to obtain an initial differential signal vector;

[0062] The initial differential signal vector is subjected to element position de-redundancy and reordering processing to obtain a corresponding equivalent uniform virtual array signal;

[0063] The equivalent uniform virtual array signal is subjected to spatial spectrum estimation to obtain an angle spectrum corresponding to the candidate signal.

[0064] In the embodiment, the core is the differential virtual array technology. By calculating the autocorrelation matrix and vectorizing, an equivalent continuous uniform virtual array with a longer aperture can be reconstructed from the signals of the physically sparse array. This is equivalent to significantly expanding the effective aperture of the array without increasing the number of physical antennas, thereby fundamentally breaking through the angle resolution limit of the original array and achieving ultra-high resolution angle estimation capability far superior to the physical array.

[0065] The "de-redundancy and reordering" step before generating the equivalent signal vector eliminates the repetition and discontinuity of the array elements in the virtual array, and finally obtains a continuous and uniform virtual array. The beam pattern of the uniform array has a lower sidelobe level. Therefore, based on the equivalent signal vector for spatial spectrum estimation, the false sidelobes and grating lobes in the angle spectrum can be significantly suppressed, the main lobe of the real target is more sharp and prominent, and the accuracy of angle estimation, single target resolution and reliability of multi-target detection are greatly improved.

[0066] The process of "autocorrelation→vectorization→virtual array reconstruction→spatial spectrum estimation" ingeniously converts the non-linear optimization problem of super-resolution angle measurement based on sparse array into a linear problem of standard spatial spectrum estimation for an equivalent uniform linear array. This enables the direct application of classical spectrum estimation methods (such as FFT) with high computational efficiency and good stability, avoiding the complex eigenvalue decomposition and search required by algorithms such as MUSIC and ESPRIT, greatly reducing the computational complexity and sensitivity to signal-to-noise ratio while ensuring high performance, and having stronger engineering implementation.

[0067] Further, based on the target distance and the corrected angle, the fast and slow frame point clouds of different beam types alternately generated by the radar at a specific time sequence are cross-frame matched and velocity deblurring processed, including:

[0068] A buffer is maintained for buffering target point cloud data of at least one historical subframe, and the target point cloud data at least includes a distance index, a corrected angle and a velocity index of the target;

[0069] The target point cloud data of the current subframe is obtained, and the target point cloud data of a historical subframe with the same beam type and different frame type and the nearest neighbor of the current subframe is selected from the buffer for matching;

[0070] Based on the distance index and the corrected angle, the target point cloud data of the current subframe and the historical subframe are matched;

[0071] For the matched point cloud pair, the velocity deblurring processing is performed based on the velocity index to determine the real velocity of the target.

[0072] In this embodiment, by maintaining a buffer and formulating a clear matching rule of "same beam, different frame type and nearest neighbor", the method of the application establishes a stable and reliable cross-frame association mechanism for the fast and slow frame point clouds generated continuously and asynchronously, ensures that the point cloud matching can achieve the best balance between time continuity and waveform difference, and lays a foundation for subsequent joint calculation using different waveform characteristics.

[0073] The core purpose of the matching rule is to pair the observation results of the same target under different waveforms (fast frame / slow frame). Since the fast and slow frames have different speed resolutions and blur characteristics, this pairing directly provides two inherent related but different form speed observations of the same target, which is the necessary and sufficient data prerequisite for accurate deblurring processing using the speed extension algorithm. When matching, not only the target distance is considered, but also the accurate corrected angle after processing by the foregoing high resolution algorithm, which constitutes a two-dimensional (distance-angle) correlation constraint. Compared with one-dimensional matching using only distance, the method greatly improves the uniqueness and accuracy of point cloud pairing, effectively overcomes the matching errors caused by target density or close position, and enhances the robustness of the system. The "nearest neighbor" condition in the matching rule ensures that the system always uses the latest and most relevant historical frame data for matching, which maximizes the reduction of errors caused by data lag and effectively controls the data size and processing delay of the buffer.

[0074] Further, the target point cloud data of the current subframe and the historical subframe are matched by using a weighted nearest neighbor algorithm, comprising:

[0075] The normalized distance difference and the normalized angle difference between the i th point in the target point cloud data of the current subframe and the j th point in the target point cloud data of the historical subframe are calculated respectively.

[0076] A weighted fusion value is calculated according to the normalized distance difference and the normalized angle difference.

[0077] If the weighted fusion value is less than a preset matching threshold, it is determined that the i th point and the j th point are matched successfully, and the speed deblurring processing is entered.

[0078] In this embodiment, the two key spatial parameters of distance and angle are combined by weighted fusion, overcoming the limitation that single-dimensional matching is prone to error when the target is dense or the trajectory is crossed. Normalization ensures that parameters of different dimensions and different ranges can be compared and fused fairly, so that the matching criterion can adaptively consider the proximity of the target in both distance and angle directions, significantly improving the comprehensiveness and accuracy of matching.

[0079] By introducing a preset matching threshold, an objective and unified threshold is set for successful matching. This effectively filters out accidental close points caused by noise, measurement error or non-associated targets, greatly reducing the false matching rate. The existence of the weighting coefficient allows the relative importance of distance and angle in matching to be adjusted according to the actual scene (such as urban road, highway), enhancing the adaptability and robustness of the algorithm under different road conditions.

[0080] The matching method aims to screen point cloud pairs highly suspected to come from the same physical target for subsequent velocity deambiguity steps. High-quality matching is a prerequisite for successful velocity deambiguity. The application ensures that only points with high spatial consistency are executed deambiguity through strict two-dimensional weighted threshold judgment, which fundamentally improves the data quality of the input of the velocity deambiguity processing, thereby guaranteeing the reliability and accuracy of the final velocity calculation result.

[0081] Further, for the matched point cloud pairs, velocity deambiguity processing is performed based on the velocity index, including:

[0082] For the matched point cloud pairs, a velocity expansion matrix of the i-th point is generated according to the velocity index, the maximum value of the velocity dimension index of the i-th point in the target point cloud data of the current subframe, and the preset expansion multiple; a velocity expansion matrix of the j-th point is generated according to the velocity index, the maximum value of the velocity dimension index of the j-th point in the target point cloud data of the historical subframe, and the preset expansion multiple;

[0083] The absolute value of the difference between all elements of the velocity expansion matrix of the i-th point and the velocity expansion matrix of the j-th point is calculated, and the minimum value is selected;

[0084] If the minimum value is less than the preset velocity threshold, it is determined that the velocity matching of the point cloud pair is successful, and the velocity corresponding to the velocity index in the velocity expansion matrix of the i-th point corresponding to the minimum value is taken as the true velocity;

[0085] If the minimum value is greater than or equal to the preset velocity threshold, it is determined that the velocity matching of the point cloud pair fails, and is first retained until the i-th point in the target point cloud data of the current subframe and all points in the target point cloud data of the historical subframe fail to match in velocity, and the i-th point is discarded.

[0086] In this embodiment, by generating a velocity expansion matrix containing a main lobe and multiple side lobes (velocity ambiguity terms), all possible values of each velocity index within multiple velocity ambiguity periods are systematically enumerated. By calculating the minimum difference between all element pairs of the two expansion matrices, only one pair of physically closest true velocity candidate values can be accurately and automatically found out, thereby efficiently solving the ambiguity problem in radar speed measurement and accurately restoring the true radial velocity of the target.

[0087] The application introduces a "preset velocity threshold" as the final criterion. This not only effectively filters out false "minimum difference" caused by noise, matching errors or non-associated targets, improves the reliability of decision-making, but also enables the algorithm to adapt to different signal-to-noise ratios and waveform parameters. At the same time, the mechanism of "first retention until all matching fails and then discard" avoids premature misjudgment of a single uncertain target, and improves the integrity of the point cloud data and the fault tolerance of the system.

[0088] The core of the deblurring process of the present application is to compare the speed observations from the fast and slow frames of two different types. Because the speed resolution and blurring characteristics of the fast and slow frames are different, there is a fixed mathematical relationship (such as a 2-fold relationship) in the "folding" of the same real speed in the speed domain. The present application converts the constraint of this waveform difference into a computable and strongly constrained matching search problem through the generation and comparison of the speed extension matrix, so that the unique real speed can be reliably solved.

[0089] Based on the same concept, the present application also provides an electronic device comprising a memory, a processor, and a computer program or instructions stored on the memory, wherein the processor executes the computer program or instructions to implement the traffic radar-based high-capture-rate target detection method as described above.

[0090] Based on the same concept, the present application also provides a computer-readable storage medium having a computer program or instructions stored thereon, wherein the computer program or instructions are executed by a processor to implement the traffic radar-based high-capture-rate target detection method as described above.

[0091] Compared with the prior art, the present application has the following beneficial effects:

[0092] The present application systematically solves the three core problems in traffic radar detection: insufficient angle resolution in the far field, inaccurate information caused by speed blurring, and difficulty in point cloud matching and fusion through a complete and coherent signal processing chain. The final output includes the final target point cloud with accurate distance, real speed, and super-high resolution corrected angle, realizing the full-parameter and high-precision synchronous measurement of the position, speed, and azimuth angle of the road target, and greatly improving the perception ability of the radar.

[0093] Through regional differential processing, the present application implements road-prior-based focused beam synthesis in the far field (second region), effectively improving the signal-to-noise ratio and angle resolution of the far field target. Combined with subsequent differential virtual array super-resolution processing, the aperture limitation of the physical array is further broken, enabling the clear differentiation and stable detection of small targets in the far field and dense targets with close angles, thereby significantly improving the detection probability (capture rate) of the system in the key far field region.

[0094] The present application creatively introduces the processing flow of "generating speed hypothesis → Doppler phase compensation → parallel angle estimation → optimal angle spectrum decision", deeply fuses and synchronously solves the problems of speed deblurring and high-resolution angle estimation, which are traditionally separate or sequential. Using the quality (peak amplitude) of the angle spectrum as the basis for decision, the correct speed is determined in reverse, realizing mutual verification and joint optimization of the speed and angle information, fundamentally solving the subsequent angle estimation error caused by speed blurring, and greatly improving the overall reliability of parameter estimation in complex scenarios.

[0095] The present application can efficiently and accurately correlate and fuse fast and slow frame point cloud data with different waveform characteristics alternately transmitted by the radar by designing the "cross-frame matching based on accurate distance and angle" and "deblurring based on speed extension matrix" processes. This not only outputs more accurate instantaneous speed, but also generates a stable target point cloud sequence that is consistent in time and space and accurate in parameters, providing input data of much higher quality than traditional methods for subsequent target tracking, classification and behavior analysis.

[0096] The whole scheme embodies the intelligent resource allocation and algorithm cooperation idea: avoid global complex operation through area division; compensate for the lack of hardware aperture through differential virtual array technology; improve parameter accuracy by using time dimension information through sliding frame matching and deblurring. These designs enable the present application method to realize a qualitative leap in detection performance without significantly increasing the cost of radar hardware and the power consumption of data processing, and have high engineering application value and market competitiveness. BRIEF DESCRIPTION OF DRAWINGS

[0097] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only one embodiment of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0098] Figure 1 is the flow chart of the high capture rate target detection method based on traffic radar in the embodiment of the present application;

[0099] Figure 2 is the schematic diagram of the receiving antenna array and the differential virtual array in the embodiment of the present application.

[0100] Figure 3 is the angle measurement schematic diagram of the receiving antenna array in the embodiment of the present application;

[0101] Figure 4 is the angle measurement schematic diagram of the differential virtual array in the embodiment of the present application;

[0102] Figure 5 is the schematic diagram of the radar beam arrangement sequence in the embodiment of the present application. DETAILED DESCRIPTION

[0103] The technical solutions in the present application will be described clearly and completely in combination with the drawings in the embodiment of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the scope of protection of the present application.

[0104] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0105] Example 1

[0106] like Figure 1 As shown, the high acquisition rate target detection method based on traffic radar (such as traffic radar operating in the 90GHz band) provided in this embodiment of the invention includes the following steps:

[0107] Step S1: Based on the road geometry parameters, the detection matrix generated based on the radar echo signal is divided into a first region and a second region.

[0108] In the initial stage of the detection process, this invention first performs a crucial detection matrix region partitioning operation. The core objective of step S1 is to intelligently partition the entire detection matrix based on the prior road geometry information of the radar deployment scenario, so as to achieve optimized allocation of subsequent processing resources and accurate adaptation of algorithm strategies. The specific implementation process is as follows:

[0109] Step S1.1: Radar raw data preprocessing and detection matrix generation.

[0110] The raw echo signal received by the radar (i.e., the radar echo signal) is preprocessed. Specifically, for the multi-channel complex signal within each coherent processing interval, a range-dimensional Fast Fourier Transform (FFT) and a velocity-dimensional (Doppler-dimensional) FFT are performed sequentially. Through these two FFT processes, the time-domain echo signal is transformed into the range-Doppler two-dimensional frequency domain, generating a complete range-Doppler detection matrix. The size of the range-Doppler detection matrix is ​​mmdoppler × numsample, where mmdoppler is the number of points in the velocity dimension (Doppler dimension) and numsample is the number of points in the range dimension. Each cell (r, v) in the range-Doppler detection matrix contains signal energy (or complex amplitude) information at the range gate r and the velocity cell v.

[0111] Step S1.2: Set distance thresholds based on lane length.

[0112] The key to step S1.2 lies in utilizing prior knowledge of the road scenario, namely lane length. Lane length The lane length is set during system initialization based on the actual traffic scenario (such as road segment mode). The nearest end usually corresponds to the radar detection blind zone, and the farthest end is set to the radar's maximum effective detection range.

[0113] Based on lane length , calculate or directly set a distance division threshold In a preferred embodiment, the distance division threshold is set as the distance dimension index corresponding to half of the lane length. For example, if the lane length is 50 meters and the distance resolution is 0.5 meters, the distance division threshold may be set as 50 (corresponding to 25 meters). The physical meaning of the distance division threshold is that it divides the detection range of the radar into a near-field region and a far-field region in the distance, where the near-field region usually has strong target echoes but complex distribution, and the far-field region has weak target echoes and is more consistent with the road geometry constraints.

[0114] Step S1.3: Perform region division of the distance-Doppler detection matrix.

[0115] Using the distance division threshold determined in step S1.2, divide the distance-Doppler detection matrix generated in step S1.1:

[0116] First region: defined as all data cells in the detection matrix whose distance dimension index is less than or equal to the distance division threshold L. This region corresponds to the near-field region. Since the target distance is close, the signal energy is usually strong, but may be greatly affected by factors such as multipath, occlusion, etc., and the angle spread is obvious. Therefore, in the present application, instead of using computationally complex beam synthesis, efficient conventional CFAR detection is directly performed on this region to quickly output reliable targets.

[0117] Second region: defined as all data cells in the detection matrix whose distance dimension index is greater than the distance division threshold L. This region corresponds to the far-field region. The signal energy of the target in this region is severely attenuated, which is the key region for weak target detection, and its azimuth angle distribution is more strictly limited by the lane geometry. Therefore, the present application will implement beam synthesis and focusing detection based on road geometry parameters on this region, which has a large amount of calculation but significant gain, to improve the detection probability and angle measurement accuracy of far-field weak targets.

[0118] Through the above steps S1.1 to S1.3, the present application completes the scene-adaptive-based region division of the detection matrix, laying a key data foundation for the subsequent differentiated, efficient and high-precision combined parallel processing process. This division method ensures that the computing resources are preferentially used for far-field weak target detection which needs to improve performance the most, while avoiding unnecessary complex operations in the near-field, thereby achieving the optimal balance between detection performance and system real-time on the whole.

[0119] Step S2: Perform beam synthesis and detection based on road geometry parameters on the data in the second region, and perform detection on the signals in the first region.

[0120] Step S2 performs parallel and differentiated detection processes on the two regions of the probe matrix to simultaneously ensure the efficiency of near-field processing and the performance of far-field detection. The specific implementation of step S2 includes the following two parallel processing branches:

[0121] Branch 1: Detect the signal in the first region (near field region).

[0122] From the first region obtained in step S1.3, the multi-channel complex signal corresponding to each range-Doppler cell is extracted directly. Instead of complex beamforming, the extracted multi-channel complex signals are directly subjected to constant false alarm rate (CFAR) detection on the signal of each receiving channel (or after incoherent energy accumulation of the multi-channel complex signals). This detection method has low computational complexity and high speed, and is suitable for near-field regions with high signal-to-noise ratios and strong target echoes. It aims to quickly and reliably output the range and velocity information (i.e., range-Doppler cell subscripts) of near-field targets.

[0123] Output: All point targets detected by CFAR (including distance dimension subscripts, velocity dimension subscripts, and preliminary amplitude information) are used as the detection results for the first region and temporarily stored for fusion.

[0124] Branch 2: Perform beamforming and detection on the data of the second region (far field region) based on road geometry parameters.

[0125] This branch is the core of improving the performance of far-field weak target detection. It is implemented for the second region, and the specific steps are as follows:

[0126] Step S2.1: Calculate the angle scanning range based on road geometry parameters.

[0127] Based on the known lane width The actual physical distance R corresponding to the preset distance threshold L. L (R) L =L×Δr, where Δr is the range resolution), to calculate the maximum scanning angle of beamforming. In a preferred embodiment, the maximum scanning angle is calculated using the following formula:

[0128] (1)

[0129] in, is the maximum scanning angle; atan is the arctangent function. The physical meaning of formula (1) is: at the actual physical distance R L At this point, the arctangent of the angle opened by half the lane width. Therefore, the angle scan range is limited to [ This ensures that the beam energy is focused within the effective area covered by the lane width, greatly suppressing clutter and interference outside the lane.

[0130] Step S2.2: Generate a set of discrete beam steering vectors.

[0131] Within the defined scanning range [ Within a given range, K discrete angle points are uniformly selected at preset angular intervals (e.g., 0.5° or 1°). }

[0132] For a radar receiving antenna array, the physical position of its nth element relative to the first element is known to be d. n (Using half-wavelength λ / 2 as the normalized unit, n = 1, 2, ..., , (Number of receiving channels). For each discrete angle point. Calculate its corresponding beam steering vector:

[0133] (2)

[0134] in, For the k-th discrete angle point The corresponding beam steering vector; j is the imaginary unit, and the superscript T indicates transpose. This yields a set of discrete beam steering vectors. .

[0135] Step S2.3: Perform beamforming on the signal in the second region.

[0136] For each distance-Doppler unit in the second region Extract its corresponding Channel complex signals (one (a column vector of ×1); using the beam steering vector generated in step S2.2, for the range-Doppler unit corresponding to... Channel complex signals Beamforming is performed. For the k-th beam direction, the synthesized output is:

[0137] (3)

[0138] in, This represents the beamforming result in the k-th beam direction, i.e., the complex signal in the k-th beam direction; the superscript H indicates the conjugate transpose. After calculating for all K beam directions, the complex signals of this range-Doppler element in the K beam directions are obtained. .

[0139] Repeat the above operation for all range-Doppler cells in the second region, thereby generating the detection matrix (size of mmdoppler×(numsample-L)×) corresponding to the second region. The transformation is into K beam domain detection matrices (each with a size of mmdoppler×(numsample-L)).

[0140] Step S2.4: Beam domain signal processing and detection.

[0141] For each element in each beam domain detection matrix Calculate its signal power Then, within a coherent processing interval, the signal power is... Perform incoherent accumulation (such as direct summation) to obtain each Energy value after unit accumulation Then, consider the energy distribution corresponding to the k-th beam direction. Independent CFAR detection is performed to detect potential target points in the beam direction. Each target point contains information: range subscript r, velocity subscript v, and beam direction angle. Accumulate energy value .

[0142] Since the same physical target may appear in multiple adjacent beam directions, the detection results of K beams need to be fused. For a specific... If a cell is detected in multiple beam directions, the energy value of that cell is selected for accumulation. The direction of the highest beam is used as the estimated angle of the target. Different The target of the unit is directly retained. All target points after fusion processing (including distance index, velocity index, preliminary estimated angle, and accumulated energy) are used as the detection results for the second region.

[0143] Step S3: Fuse the detection results of the second region and the first region to output a preliminary detection point cloud containing target distance and velocity information.

[0144] The detection results of the first region output from branch one and the detection results of the second region output from branch two in step S2 are merged to form a preliminary detection point cloud. For the very few possible instances where the same target is detected repeatedly at the boundary of two regions (with the same or very close distance and velocity indices), the point with the higher amplitude (or energy) after incoherent accumulation is selected as the output, while the other point is suppressed. Finally, the preliminary detection point cloud obtained after fusion contains the distance information (indices) and velocity information (indices) of all detected targets, as well as the preliminary estimated angle from the target in the second region, providing basic data for subsequent high-resolution angle correction and velocity deblurring processing.

[0145] Step S4: generating at least one velocity hypothesis based on the velocity information of each target in the preliminary detection point cloud and the maximum unambiguous velocity of the radar, and generating a plurality of groups of candidate signals by Doppler phase compensation based on the velocity hypothesis.

[0146] Step S4 is the key innovative link of the present application to realize joint solution of velocity and angle. The purpose is to generate a candidate real velocity hypothesis for each target that may have velocity ambiguity, and use the hypothesis to compensate the original array signal in phase, so as to provide a plurality of candidate, phase-consistent signal versions for subsequent high-resolution angle estimation. The specific implementation process is as follows:

[0147] Step S4.1: generating at least one velocity hypothesis for each target in the preliminary detection point cloud.

[0148] For each target point i in the preliminary detection point cloud, the key parameters are obtained:

[0149] Original velocity information : This value comes from the preliminary detection point cloud, which is the velocity value corresponding to the target in the velocity dimension (Doppler dimension). The velocity information may be folded in the range of [ ] due to velocity ambiguity.

[0150] Maximum unambiguous velocity of the radar : This is a system constant determined by the current radar waveform parameters.

[0151] According to the sign of the velocity information , a velocity hypothesis is generated as the most likely candidate value of the real velocity of the target. The generation rule is as follows:

[0152] If the velocity information > 0, the velocity hypothesis is ;

[0153] If the velocity information ≤ 0, the velocity hypothesis is .

[0154] This generation rule is based on the physical principle that the velocity ambiguity under TDMA (Time Division Multiple Access) waveform is folded with a period of 2 . For an observed positive velocity , the real velocity may be itself, or (a smaller positive or negative velocity).

[0155] Step S4.2: calculating the corresponding Doppler phase compensation value for each velocity hypothesis.

[0156] For each target, the velocity hypothesis , the Doppler phase difference it causes, which needs to be compensated among the receiving channels (i.e., the Doppler phase compensation value) is calculated as:

[0157] (4)

[0158] where denotes the velocity hypothesis corresponding Doppler phase compensation value (unit: m / s); denotes the center frequency of the current radar transmit waveform (unit: Hz); denotes the duration of a single transmit pulse or the equivalent coherent processing time (unit: s); denotes the speed of light (unit: m / s). Formula (4) calculates the additional phase term introduced on the channel of different transmit antenna and receive antenna combination due to the target radial velocity .

[0159] Step S4.3: Compensate the original multi-channel received signal of the target with the Doppler phase compensation value, and generate a candidate signal.

[0160] Step S4.3 is the core operation at the signal level, which generates multiple phase-corrected signal copies for each target.

[0161] Extract the original multi-channel received signal: according to the distance index and the velocity index (corresponding to the observed velocity ) of the target i in the preliminary detection point cloud, extract the original multi-channel complex signal corresponding to the cell from the original, un-beamformed range-Doppler detection matrix. . is a column vector with a dimension of × (the number of transmit antennas and the number of receive antennas are and respectively for MIMO virtual array), or simplified to dimensional receiving channel signal according to the processing flow.

[0162] Apply the phase compensation value calculated in step S4.2 to the extracted original signal to generate a phase-compensated candidate signal, and the specific formula is:

[0163] (5)

[0164] where, denotes the phase compensated candidate signal; denotes the transmit antenna index; denotes the receive antenna index.

[0165] The physical meaning of equation (5) is: if the true velocity of the target is exactly , then this compensation will exactly cancel out the Doppler phase difference introduced between the receive channels , so that the compensated signal is, in the ideal case, equal to the array signal of a stationary target or a target without velocity blurring.

[0166] For the same target i, its original velocity itself also implies an assumption of = (no velocity blurring). Therefore, for each target, we generate at least two groups of candidate signals:

[0167] Candidate signal group A: the signal obtained after phase compensation based on the velocity assumption (generated by step S4.1).

[0168] Candidate signal group B (optional but implied): the signal obtained after phase compensation based on the original velocity (i.e. the velocity assumption is , corresponding to calculation).

[0169] Finally, each target corresponds to multiple groups (at least two groups) of candidate signals after phase compensation based on different velocity assumptions { , ,...}. These candidate signals are sent to the subsequent step S5 for high-resolution angle estimation based on the differential virtual array.

[0170] Step S4 generates a velocity assumption through a simple rule and uses this assumption to accurately compensate the original array signal. This creatively transforms the problem of "which velocity is true" into the problem of "which group of phase-compensated signals can produce a better angle spectrum" in terms of signal quality, providing a clear and calculable basis for the subsequent joint decision.

[0171] Step S5: Perform high-resolution angle estimation processing based on the differential virtual array principle for each group of candidate signals to obtain multiple angle spectra.

[0172] ​​Step S5 receives multiple sets of candidate signals output from step S4 (each set corresponding to a velocity hypothesis), and uses Differential Virtual Array (DVA) technology to generate a high-resolution spatial angle spectrum for each set of signals, thereby providing a basis for subsequent selection of the correct velocity and angle. The specific implementation is as follows: the following process is executed independently for each set of candidate signals:

[0173] Suppose a non-uniform receiving antenna array has There are n array elements, and the physical position of the nth array element is a. n ×d, where n=1,2,..., d = λ / 2 is the reference antenna spacing, λ is the radar operating wavelength, and a n Let a be the relative position coefficient of the nth array element in units of d. For a uniform array, a n = n-1; for sparse arrays, a n This constitutes a non-uniform sequence.

[0174] Assume that the receiving antenna array simultaneously receives signals from M uncorrelated far-field targets. The incident signal vector is... The noise vector is The noise vector statistically follows a Gaussian distribution with a mean of 0 and a variance of . At this time, the receiving antenna array receives... ×1D quick capture signal It can be represented as:

[0175] (6)

[0176] in, for A ×M-dimensional guidance matrix; for the m-th direction... The signal, its corresponding ×1-dimensional guiding vector for:

[0177] (7)

[0178] For each group of candidate signals from step S4 (which can be viewed as a snapshot estimate on a specific distance-Doppler unit), The following steps S5.1 to S5.4 are executed independently:

[0179] Step S5.1: Calculate the spatial autocorrelation matrix of the candidate signal.

[0180] For a set of candidate signals for target i (one Estimate the spatial autocorrelation matrix of a 1×1 dimensional column vector:

[0181] (8)

[0182] in, This represents the estimated spatial autocorrelation matrix; This indicates the number of snapshots used for estimation, which in radar is typically the number of pulses within a coherent processing interval or the number of multichannel samples within a range-Doppler cell; the superscript H indicates a conjugate device. Let represent the candidate vector for the l-th snapshot. For point target detection, Typically, 1 or a small number is used, or a single snapshot estimate is used directly. .

[0183] The calculated spatial autocorrelation matrix It is × The Hermitian matrix. According to the signal model, the theoretical spatial autocorrelation matrix can be expressed as:

[0184] (9)

[0185] in, This represents the theoretical spatial autocorrelation matrix; This indicates the operation of calculating the mean; for A ×M-dimensional guiding matrix; Indicates the m-th direction The guiding vector; This represents the autocorrelation matrix corresponding to the incident signal; Indicates the order is The identity matrix; This represents the noise power (i.e., variance) on each receiving channel. The covariance matrix represents the entire noise vector; Represents the m-th incident signal The power (variance).

[0186] Step S5.2: Vectorize the spatial autocorrelation matrix to obtain the initial differential signal vector.

[0187] The spatial autocorrelation matrix calculated in step S5.1 Perform vectorization operations:

[0188] (10)

[0189] The vec(·) operator converts the matrix... Stack them in columns to form a The column vector z of size 1 can be expressed as:

[0190] (11)

[0191] where is a matrix of size M x M whose m-th column is , denotes taking the conjugate, denotes the Kronecker product; is the equivalent steering vector of the differential virtual array corresponding to the direction ; is a signal power vector of size M x 1; is a vector of size 1 whose elements are 0 except for the diagonal position of the matrix which is 1.

[0192] The vector z is the initial differential signal vector, which contains the information of all virtual array elements that can be generated by the original sparse array through differential operation.

[0193] Step S5.3: De-redundancy and reordering of the initial differential signal vector to obtain the equivalent uniform virtual array signal.

[0194] Step S5.3 is to extract the available and continuous virtual array signal from the theoretical vector z, and the core is to construct and apply the equivalent steering vector of the differential virtual array .

[0195] According to , it can be further expressed as the following block structure:

[0196] (12)

[0197] where the nth sub-block of size x 1 is given by formula (13), which represents the phase relationship between the signals of the array elements when the nth array element is taken as the reference, where the k-th element depends on the relative position of the array element k and the reference array element n . Based on d = λ / 2, the k-th element is:

[0198] (13)

[0199] where denotes the position coordinates of the k-th array element; ​represents the position coordinate of the nth specific reference element.

[0200] The vector z obtained in step S5.2 has elements corresponding to the structure of Each element in vector z essentially maps to a position of a differential virtual element determined by a pair of physical elements with index (in half wavelength unit).

[0201] Due to the differential operation different pairs of physical elements may map to positions symmetric about the origin and , or different pairs of elements may map to the same differential position, so there are a large number of redundant elements in the virtual element set corresponding to vector z. These redundant elements need to be processed. The specific operation is: select one value to retain or take the arithmetic mean of all repeated position corresponding signal values as the final signal value at the unique virtual element position. This process removes redundant elements.

[0202] Reordering and extracting continuous uniform part: the virtual elements obtained after removing redundancy, with unique positions, are sorted from small to large according to the numerical value of their spatial coordinates . After sorting, observe the distribution of virtual element positions. Usually, there will be a continuous and uniformly distributed virtual element near the origin, with a half wavelength interval. Generally, the longest continuous and uniform virtual element subset located at the center is selected for subsequent processing. Let this continuous and uniform virtual array contain N v elements.

[0203] According to the sequence of continuous and uniform virtual element positions, extract the signal values corresponding to these positions from the set of signal values that have been removed redundancy, and strictly arrange them in order of their spatial positions to form an N v ×1 column vector z virtual . This z virtual is the equivalent signal vector corresponding to a continuous and uniform virtual array (i.e., the equivalent uniform virtual array signal). The effective physical aperture of z virtual is (N v -1)×(λ / 2), which is significantly larger than the physical aperture of the original sparse array, as shown in Figure 2 . Figure 2 In the figure, the top row is the receiving antenna array, and the bottom row is the differential virtual array. The original sparse array is reconstructed, and the middle continuous array is generally selected for subsequent angle spectrum estimation. The uniform element advantage suppresses the sidelobes generated by the original sparse array, as shown in Figure 3and Figure 4 As shown in FIG. 6, the side lobes are basically eliminated, and only the main lobe peak is used for angle measurement.

[0204] Step S5.4: Perform spatial spectrum estimation on the equivalent uniform virtual array signal to obtain an angle spectrum.

[0205] The conventional beamforming method based on FFT is used to perform spatial spectrum estimation on the equivalent uniform virtual array signal z virtual obtained in step S5.3.

[0206] The N virtual point (N FFT ≥ N FFT ) fast Fourier transform is performed on z v ; the FFT result is squared to obtain a power spectrum; and the FFT point number is mapped to a spatial angle, thereby obtaining a discrete angle spectrum, i.e., a high-resolution angle spectrum corresponding to the candidate signal.

[0207] Step S5.5: Generate multiple angle spectra for each target.

[0208] For each group of candidate signals of the same target i, steps S5.1 to S5.4 are repeatedly performed. Finally, multiple angle spectra are obtained for each target.

[0209] Step S5 transforms the observation of the sparse array into an equivalent large-aperture uniform array signal by using the differential virtual array technology, and performs high-resolution spectrum estimation. This process generates an independent angle spectrum for each velocity assumption, thereby converting the joint estimation problem of "velocity-angle" into a comparison problem of different angle spectrum qualities.

[0210] Step S6: Calculate the peak amplitude of each angle spectrum, and take the angle value corresponding to the angle spectrum with the maximum peak amplitude as the corrected angle of the target, thereby obtaining a detection point cloud containing the distance, velocity information, and corrected angle of the target.

[0211] The purpose of step S6 is to select a group that best matches the real physical scene from multiple angle spectra generated for the same target by different velocity assumptions by analyzing the quality (mainly using the peak amplitude as the criterion), thereby simultaneously determining the final accurate corrected angle and the real velocity of the target. The specific implementation process is as follows:

[0212] Step S6.1: Extract the peak amplitude and corresponding angle of multiple angle spectra for each target.

[0213] For the same target i from step S5, whose preliminary information includes the distance subscript r i , the observation velocity subscript v iperform the following operations:

[0214] For each angle spectrum, find its global maximum (peak) value and angle value in its valid angle scanning range. Establish a list for target i to record the peak information of each angle spectrum: (spectrum index, peak amplitude, peak corresponding angle, the speed hypothesis corresponding to this spectrum).

[0215] Step S6.2: Based on the peak amplitude, make a joint decision to determine the final angle and speed.

[0216] The correct speed hypothesis will make the array signal energy in the true target direction achieve the most ideal in-phase superposition through accurate phase compensation, so as to produce a main lobe peak with the sharpest and highest amplitude in its angle spectrum.

[0217] Compare all the peak amplitudes recorded for target i, select the angle spectrum with the largest peak amplitude, and get the spectrum index, peak corresponding angle, and speed hypothesis corresponding to this spectrum. The peak corresponding angle is the high-precision angle estimation value (i.e. the corrected angle) of the target obtained after super-resolution processing and Gaussian spectrum quality decision.

[0218] If the speed hypothesis corresponding to this spectrum is equal to the original observed speed , it is determined that the target speed has not been blurred, and the true speed is ; if the speed hypothesis corresponding to this spectrum is the speed hypothesis generated by ±2 rule, it is determined that the target speed has been blurred, and the true speed is the speed hypothesis corresponding to this spectrum. The distance information r i of the target remains unchanged.

[0219] Repeat steps S6.1 to S6.2 for each target in the preliminary detection point cloud. Collect the refined point cloud entries (distance information r i , speed information, and corrected angle) of all targets to form a new point cloud set.

[0220] Step S6 intelligently selects the most reasonable speed and angle combination by comparing the indirect method (angle spectrum peak) of "signal restoration quality" under different speed hypotheses. This method solves the problems of speed deblurring and angle super-resolution simultaneously, fully utilizes all the information of the array signal, and finally outputs a target point cloud with highly accurate and consistent parameters, laying the optimal data foundation for subsequent cross-frame matching and tracking.

[0221] Step S7: Based on the target distance and the correction angle, the fast and slow frame point clouds of different beam types alternately generated by the radar are matched and velocity deblurring processed across frames, and the final target point cloud with accurate parameters is output after fusion.

[0222] The purpose of step S7 is to generate a final target list containing high-precision distance, angle, and real velocity after deblurring by spatiotemporal correlation and velocity deblurring algorithm using the subframe point cloud data alternately emitted by the radar with different waveform characteristics. The specific implementation process is based on a specific beam scheduling strategy and matching deblurring algorithm, which is as follows:

[0223] Step S7.0: Beam scheduling and data buffering strategy.

[0224] The radar beams are alternately emitted in a fixed sequence of "narrow-beam fast frame, wide-beam slow frame, narrow-beam slow frame, and wide-beam fast frame", forming a complete cycle as shown in Figure 5 . Figure 5 In the figure, Frame1~Frame 6 represent the 1st frame to the 6th frame, and chirp represents a linear frequency modulation signal.

[0225] Buffer initialization: maintain a data buffer for caching historical subframe point cloud data processed.

[0226] Sliding frame processing logic: receive and process the kth subframe (k is the frame number).

[0227] According to the beam type (wide / narrow) and frame type (fast / slow) of the current subframe, the matching object is dynamically determined.

[0228] If the current frame is a wide-beam slow frame, select the cached wide-beam fast frame point cloud data with the same beam type (wide) but different frame type (fast) from the buffer for matching; if the current frame is a narrow-beam fast frame, select the cached narrow-beam slow frame point cloud data with the same beam type (narrow) but different frame type (slow) from the buffer for matching.

[0229] This "out-of-order arrangement and sliding frame matching" strategy enables the system to complete "fast-slow frame" matching of a wide beam and a narrow beam once within any time window (such as 80ms), achieving efficient point cloud output under a fixed refresh rate and optimizing the 160ms delay required by traditional sequential processing to 80ms.

[0230] Step S7.1: Obtain and prepare matching point cloud data.

[0231] Current subframe point cloud: obtain the refined detection point cloud corresponding to the current subframe (kth frame) output in step S6, where the data of each target point i is distance index, velocity index, and correction angle.

[0232] Historical subframe point cloud (frame k-2): According to the rules in step S7.0, the corresponding historical frame refined detection point cloud is retrieved from the buffer.

[0233] Step S7.2: Coarse matching of point clouds based on weighted nearest neighbor algorithm.

[0234] Step S7.2 aims to quickly identify point pairs from two point cloud frames that are most likely to correspond to the same physical target in terms of spatial location (distance, angle). Specifically, this includes:

[0235] Step S7.21: Calculate the normalized distance difference and normalized angle difference between the i-th point in the current subframe point cloud and the j-th point in the historical subframe point cloud.

[0236] (14)

[0237] (15)

[0238] in, These represent the distance indices of the i-th point in the current subframe point cloud and the j-th point in the historical subframe point cloud, respectively. This represents the maximum value in the distance dimension; These represent the angle values ​​of the i-th point in the current subframe point cloud and the j-th point in the historical subframe point cloud, respectively. Indicates the maximum field of view; This represents the normalized distance difference; This represents the difference in normalized angles.

[0239] Step S7.22: Calculate the weighted fusion value based on the normalized distance difference and the normalized angle difference. The specific formula is as follows:

[0240] (16)

[0241] in, Indicates the weighted fusion value; and This represents the weighting coefficients for the normalized distance difference and the normalized angle difference.

[0242] Step S7.23: Compare the weighted fusion value with the preset matching threshold.

[0243] If the weighted fusion value If the value is less than the preset matching threshold, the i-th point in the current subframe point cloud and the j-th point in the historical subframe point cloud are considered to be spatially close enough to be two observations of the same target, and the match is successful, proceeding to velocity deblurring processing. Otherwise, the match fails.

[0244] Step S7.3: Velocity defuzzification processing based on the velocity extension matrix.

[0245] For each coarsely matched point cloud pair (i,j), precise velocity calculation is performed using the different velocity resolution characteristics of fast and slow frames. Let the maximum value of the velocity dimension index in both fast and slow frames be... (Determined by waveform parameters), a preset even-number expansion factor P (e.g., P=4).

[0246] Generate the velocity index v of the i-th point in the target point cloud data of the current subframe. k Extended matrix :

[0247] (17)

[0248] Generate the velocity index v of the j-th point in the historical subframe point cloud. j Extended matrix :

[0249] (18)

[0250] These two extended matrices enumerate points i and j respectively, considering velocity ambiguity (in 2*). In the case of a period, all possible corresponding real velocity indices.

[0251] Calculate the extended matrix and Find the absolute value of the difference between all pairs of elements in the algorithm, and then find the minimum value among them. If the minimum value is less than a preset speed threshold, the speed match is considered successful. At this point, the value corresponding to the minimum value... The element in the matrix represents the correct velocity index of point i after defuzzification. Based on the minimum value... The elements in the calculation yield the target's true radial velocity.

[0252] If the minimum value is greater than or equal to the preset speed threshold, the speed matching is determined to be unsuccessful. Point i in the current subframe is temporarily retained and attempts to match it with other points in the historical subframe are continued. If point i in the current subframe fails to match with any points j in the historical subframe, point i is discarded after all matches in the current subframe have been processed.

[0253] S7.4: Fusion output final target point cloud

[0254] For each point pair (i,j) that successfully passes coarse matching in step S7.2 and velocity matching in step S7.3, generate a final target entry:

[0255] Distance: take r i or (r) i +r j) / 2 (or the better one according to the signal-to-noise ratio), and converted into a physical distance.

[0256] Angle: using the corrected angle from the refined point cloud.

[0257] Velocity: using the real velocity calculated by S7.3.

[0258] All the successfully matched and calculated final target entries in the current period (for example, in an 80ms window) are collected to form a final target point cloud.

[0259] The refined point cloud processed in the current subframe is stored in the buffer according to its beam type, replacing the old frame data of the same type, and preparing for the matching of the next subframe.

[0260] Step S7 efficiently integrates the coverage advantages of wide and narrow beams and the speed measurement advantages of fast and slow frames through innovative sliding frame scheduling and "same beam, different frame type" matching rules. The use of accurate correction angles for weighted matching greatly improves the correlation accuracy; and through the speed expansion matrix comparison, the speed ambiguity problem between fast and slow frames is robustly solved. The final output is a highly accurate and spatiotemporal consistent fusion point cloud in spatial position and velocity dimensions, greatly improving the radar system's target capture rate and parameter measurement accuracy, and providing high-quality input for upper layer perception and decision-making.

[0261] Embodiment two

[0262] The embodiment of the application also provides an electronic device, which comprises a memory, a processor and a computer program or instructions stored in the memory, and the processor executes the computer program or instructions to realize the traffic radar-based high-capture-rate target detection method in the embodiment of the application.

[0263] Although not shown, the electronic device includes a processor, which can perform various appropriate operations and processes according to programs and / or data stored in a read-only memory (ROM) or programs and / or data loaded from a storage section into a random access memory (RAM). The processor can be a multi-core processor or can include multiple processors. In some embodiments, the processor can include a general-purpose main processor and one or more special-purpose coprocessors, such as a central processing unit, a graphics processing unit (GPU), a neural network processing unit (NPU), a digital signal processor (DSP), etc. In the RAM, various programs and data required for device operation are also stored. The processor, ROM and RAM are connected to each other through a bus. An input / output (I / O) interface is also connected to the bus.

[0264] The processor and the memory are used together to execute programs / instructions stored in the memory, and the programs / instructions, when executed by a computer, can implement the methods, steps or functions described in the above embodiments.

[0265] Although not shown, the embodiments of the present application also provide a computer readable storage medium having stored thereon a computer program or instructions, which, when executed by a processor, implement the traffic radar based high capture rate target detection method in the embodiments of the present application.

[0266] The computer readable storage medium includes permanent and non-permanent, removable and non-removable media, which can be implemented by any method or technology to store information. The information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), flash memory or other memory technologies, compact disc read only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible by a computing device. According to the definition herein, computer readable medium does not include transitory computer readable medium, such as modulated data signals and carriers.

[0267] The above only discloses specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or modifications within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application.

Claims

1. A traffic radar-based high capture rate target detection method, characterized by, The detection method comprises: According to the road geometric parameters, the detection matrix generated based on the radar echo signal is divided into a first region and a second region; The data in the second region is subjected to beam synthesis and detection based on the road geometric parameters, while the signals in the first region are subjected to detection; The detection results of the second region and the first region are fused, and a preliminary detection point cloud containing target distance and speed information is output; Based on the speed information of each target in the preliminary detection point cloud and the maximum unambiguous speed of the radar, at least one speed hypothesis is generated, and based on the speed hypothesis, a plurality of groups of candidate signals are generated through Doppler phase compensation; Each group of candidate signals is subjected to high-resolution angle estimation processing based on the principle of differential virtual array to obtain a plurality of angle spectra; The peak amplitudes of each angle spectrum are calculated, and the angle value corresponding to the angle spectrum with the maximum peak amplitude is taken as the corrected angle of the target, thereby obtaining a detection point cloud containing target distance, speed information and corrected angle; Based on the target distance and the corrected angle, the fast and slow frame point clouds of different beam types alternately generated by the radar at a specific timing are subjected to cross-frame matching and speed unblurring processing, and the final target point cloud with accurate parameters is output after fusion; According to the road geometric parameters, the detection matrix generated based on the radar echo signal is divided into a first region and a second region, which comprises: The radar echo signal is subjected to Fourier transform in the distance dimension and the speed dimension to generate a range-Doppler detection matrix; According to the lane length in the road geometric parameters, a distance division threshold is set; The region with a distance dimension subscript less than or equal to the distance division threshold in the range-Doppler detection matrix is divided into the first region, and the region with a distance dimension subscript greater than the distance division threshold is divided into the second region; Wherein, the data in the second region is subjected to beam synthesis and detection based on the road geometric parameters, which comprises: Based on the lane width in the road geometric parameters and the preset distance division threshold, the angle scanning range is calculated; According to the angle scanning range, a set of discrete beam steering vectors are generated; The multi-channel complex signals extracted from the second region are subjected to beam synthesis using the beam steering vectors to obtain complex signals in multiple beam directions; The complex signals in each beam direction are subjected to modulus square operation to obtain signal power; The signal power is subjected to non-coherent accumulation within a coherent processing interval to obtain an accumulated energy distribution; The accumulated energy distribution is subjected to constant false alarm rate detection to output the target detection result of the second region. 2.The traffic-radar-based high-capture-rate target detection method of claim 1, wherein, According to the angle scanning range, a set of discrete beam steering vectors are generated, which comprises: K discrete angle points are selected at a preset interval within the angle scanning range; According to the physical position of the radar receiving antenna array, the beam steering vector of each discrete angle point is calculated. 3.The traffic-radar-based high-capture-rate target detection method of claim 1, wherein, Based on the speed information of each target in the preliminary detection point cloud and the maximum unambiguous speed of the radar, at least one speed hypothesis is generated, which comprises: acquiring speed information of each target in the preliminary detection point cloud and maximum unambiguous velocity of the radar ; generating speed information for each target in the point cloud based on the preliminary detection and the maximum unambiguous velocity of the radar generating the at least one speed hypothesis If the speed information is greater than 0, the speed hypothesis is ; If the speed information is less than or equal to 0, the speed hypothesis is ; and the speed information of each object in the preliminary detection point cloud directly as the speed hypothesis. 4.The traffic-radar-based high-capture-rate target detection method of claim 1, wherein, Based on the speed hypothesis, a plurality of groups of candidate signals are generated through Doppler phase compensation, which comprises: For each speed hypothesis generated for each target, a corresponding Doppler phase compensation value is calculated; Using each of the Doppler phase compensation values, the original multi-channel received signal is phase compensated respectively; Wherein each of the phase-compensated signals constitutes a candidate signal, and the original multi-channel received signal is a signal extracted from the detection matrix according to the distance and speed information of the target. 5.The traffic-radar-based high-capture-rate target detection method of claim 1, wherein, For each candidate signal, high-resolution angle estimation processing based on the principle of differential virtual array is performed, including: For each candidate signal, a corresponding spatial autocorrelation matrix is calculated; A vectorization operation is performed on the spatial autocorrelation matrix to obtain an initial differential signal vector; An element position de-redundancy and reordering processing is performed on the initial differential signal vector to obtain a corresponding equivalent uniform virtual array signal; A spatial spectrum estimation is performed on the equivalent uniform virtual array signal to obtain an angle spectrum corresponding to the candidate signal.

6. The high capture rate target detection method based on traffic radar according to any one of claims 1-5, characterized in that, Based on the target distance and the corrected angle, the fast and slow frame point clouds of different beam types alternately generated by the radar in a specific time sequence are cross-frame matched and velocity de-masking processed, including: A buffer is maintained for buffering target point cloud data of at least one historical subframe, wherein the target point cloud data at least includes the distance index, the corrected angle and the speed index of the target; The target point cloud data of the current subframe is obtained, and from the buffer, the target point cloud data of a historical subframe with the same beam type and different frame type and the closest neighbor of the current subframe is selected for matching; Based on the distance index and the corrected angle, the target point cloud data of the current subframe and the historical subframe is matched; For the matched point cloud pair, velocity de-masking processing is performed based on the speed index to determine the real speed of the target.

7. The high capture rate target detection method based on traffic radar according to claim 6, characterized in that, The weighted nearest neighbor algorithm is used to match the target point cloud data of the current subframe and the historical subframe, including: The normalized distance difference and the normalized angle difference between the i th point in the target point cloud data of the current subframe and the j th point in the target point cloud data of the historical subframe are calculated respectively; A weighted fusion value is calculated according to the normalized distance difference and the normalized angle difference; If the weighted fusion value is less than a preset matching threshold, it is determined that the i th point and the j th point are matched successfully, and the velocity de-masking processing is entered. 8.The traffic-radar-based high-capture-rate target detection method of claim 6, wherein, For the matched point cloud pair, velocity de-masking processing is performed based on the speed index, including: For the matched point cloud pair, a speed expansion matrix of the i th point is generated according to the speed index, the maximum speed index and a preset expansion multiple of the i th point in the target point cloud data of the current subframe; a speed expansion matrix of the j th point is generated according to the speed index, the maximum speed index and a preset expansion multiple of the j th point in the target point cloud data of the historical subframe; The absolute value of the difference between all elements of the speed expansion matrix of the i th point and the speed expansion matrix of the j th point is calculated, and the minimum value is selected; If the minimum value is less than a preset speed threshold, it is determined that the point cloud pair is matched successfully, and the speed corresponding to the speed index in the speed expansion matrix of the i th point corresponding to the minimum value is the real speed. If the minimum value is greater than or equal to a preset speed threshold value, it is determined that the point cloud fails in speed matching, and the point is reserved first until the i th point in the target point cloud data of the current subframe and all points in the target point cloud data of the historical subframe fail in speed matching, and the i th point is discarded.

9. An electronic device comprising a memory, a processor, and a computer program or instructions stored on the memory, wherein, The processor executes the computer program or instruction to implement the traffic radar-based high-capture-rate target detection method in any one of claims 1-8.

10. A computer readable storage medium having stored thereon a computer program or instructions, characterized in that, The computer program or instruction is executed by the processor to implement the traffic radar-based high-capture-rate target detection method in any one of claims 1-8.

Images