A Tunneling Millimeter-Wave Radar Target Detection Method Based on Beamforming

Abstract

This invention belongs to the field of radar detection technology, specifically a tunnel millimeter-wave radar target detection method based on beamforming. It first divides the horizontal monitoring area of ​​the tunnel into multiple sectors and establishes a spatial index. Narrow beams pointing towards each sector are transmitted sequentially, and digital beamforming is performed on the received echoes. At the transmitting end, unnecessary energy radiation towards the tunnel sidewalls is reduced, and at the receiving end, echoes from non-target directions are suppressed. The spatially filtered signals of each sector are output, and after feature tensor transformation, cross-sector feature fusion is performed to generate joint spatial features. Based on the joint spatial features, candidate targets are detected and false alarms are determined. Finally, the position, size, and category information of the target are regressed and output in a bird's-eye view coordinate system. This invention actively weakens the basis for multipath interference at the physical level of signal propagation, overcoming the passivity and delay defects of existing technologies that rely on post-processing algorithms for filtering.

Description

Technical Field

[0001] This invention belongs to the field of radar detection technology, specifically a tunnel millimeter-wave radar target detection method based on beamforming. Background Technology

[0002] In the enclosed environment of a tunnel, strong wall reflections cause severe multipath interference for millimeter-wave radar, resulting in high false alarm rates and inaccurate positioning. Millimeter-wave radar, due to its strong penetration and resistance to environmental interference, has become the primary sensing method in tunnel scenarios.

[0003] Currently, the mainstream technical solutions for dealing with multipath problems in tunnels mainly include: First, adopting a camera-millimeter-wave radar multimodal fusion strategy, such as the method for removing multipath false targets in tunnels based on multi-radar video, which is published in Chinese Patent Publication No. CN115343700B. This method determines whether the target is a real object by associating radar point clouds with video semantic information, thereby filtering out multipath false alarms.

[0004] Secondly, post-processing algorithms are used to remove isolated or non-compliant clutter points. For example, the tunnel scene detection method and millimeter-wave radar with Chinese patent publication number CN111699408A can identify whether a vehicle has entered a tunnel and dynamically determine and remove false targets based on the intensity of the reflected echo signal received after the vehicle enters the tunnel.

[0005] However, all of the above technical solutions have obvious drawbacks: multimodal fusion is limited by the sharp drop in camera performance in low-light and high-fog environments, making it impossible to achieve stable operation in all weather conditions. Algorithm denoising is essentially a post-processing correction, and it has high computational overhead and poor real-time performance, which can easily lead to target loss or delayed response in high-speed moving scenes.

[0006] Existing millimeter-wave radar systems still lack active suppression mechanisms for multipath propagation characteristics in tunnel scenarios. They cannot actively suppress the generation of multipath effects from the physical level of radar signal transmission and reception. In particular, when multiple reflection paths overlap to form false targets, it is difficult to distinguish between real echoes and image interference, which affects the safety and reliability of autonomous driving or intelligent transportation. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, embodiments of the present invention provide a tunnel millimeter-wave radar target detection method based on beamforming, which can effectively solve the problems involved in the prior art.

[0008] The objective of this invention can be achieved through the following technical solution: a tunnel millimeter-wave radar target detection method based on beamforming, comprising: dividing the horizontal monitoring area in the tunnel in front of the radar into multiple sectors, and establishing a spatial index mapping between the sectors and the azimuth angle.

[0009] According to the spatial index mapping, phase compensation and amplitude weighting are applied to the transmitted signals of each transmitting element in the radar transmitting antenna array, and narrow beams pointing to each sector are formed and transmitted in sequence.

[0010] For each transmit narrow beam, digital beamforming is performed on the echo signals of each channel of the receiving antenna array. The received beam is then synthesized through complex weighting, and the spatial filtering signal corresponding to each sector is output.

[0011] The spatial filtering signals of each sector are converted into feature tensors, and cross-region feature fusion is performed on the feature tensors to generate joint spatial features.

[0012] Based on the joint airspace features, candidate targets are detected, and false alarms are identified and filtered out by combining tunnel physical boundary information. The final target's position, size, and category information are then regressed and output in the bird's-eye view coordinate system.

[0013] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) The present invention first divides the tunnel horizontal monitoring area into sectors and establishes a spatial index, and then generates focused beams pointing to each sector in sequence through the coordinated beamforming of the transmitting end and the receiving end. This mechanism reduces unnecessary energy radiation to the tunnel sidewall at the transmitting end and suppresses echoes from non-target directions at the receiving end, actively weakening the basis for the generation of multipath interference from the physical level of signal propagation, and overcoming the passivity and delay defects of the prior art that rely on post-processing algorithm filtering.

[0014] (2) The present invention converts the signals of each sector into feature tensors and performs cross-sector fusion. The fusion process introduces a dynamic weight adjustment mechanism based on signal reliability, which effectively integrates multi-view information and can automatically reduce the feature contribution of sectors contaminated by local strong reflections, thereby improving robustness in non-uniform tunnel environments.

[0015] (3) In the bird's-eye view coordinate system, the present invention directly regresses the position, size and category information of the target based on the fused joint spatial features. It avoids complex coordinate transformation, can complete perception and false alarm removal from a unified perspective, and provides intuitive results for easy downstream planning and control. It also integrates multipath false alarm discrimination steps to improve the overall accuracy of target detection. Attached Figure Description

[0016] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0018] Figure 2 This is a schematic diagram of the spatial filtering signal output process for each sector of the present invention.

[0019] Figure 3 This is a schematic diagram of the feature tensor transformation of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figure 1 As shown, this invention provides a tunnel millimeter-wave radar target detection method based on beamforming, comprising: Given that radar signals in a tunnel environment are reflected multiple times by the side walls, generating dense multipath interference, which severely affects the reliability of target detection based on a single antenna pattern, this invention improves anti-interference capability by using beamforming control to directionally focus the transmitted energy in the horizontal dimension. To this end, a discretized spatial reference frame needs to be established for beam pointing.

[0022] Therefore, the first step (S1) of the method of the present invention is: dividing the horizontal monitoring area in the tunnel in front of the radar into multiple sectors, and establishing a spatial index mapping between the sectors and the azimuth angle. Specifically, this includes: establishing a cross-sectional coordinate system with the vertical projection point of the radar installation point on the cross-section of the tunnel as the origin, where the horizontal axis is the horizontal direction and the direction of tunnel movement is positive. The vertical axis is the vertical direction and upward is positive.

[0023] Based on the radar installation height and a given azimuth angle, the equation of the horizontal detection ray in the cross-sectional coordinate system is constructed. The principle for constructing the equation is: for any point on the horizontal detection ray, the coordinates... Its ordinate The value is equal to the radar installation height, plus the x-coordinate of that point. With a given azimuth tangent value The product of.

[0024] Based on the tunnel's geometric topology, independent calculation processes are performed for the first and second sidewalls of the tunnel: For each sidewall, a theoretical contour line mathematical expression describing the geometric profile of that sidewall is obtained by pre-inputting tunnel design drawing data or by scanning and measuring with radar during the initial calibration phase. Among them, targeting Independent variable The range of values ​​is defined by the lateral range of the sidewall in the cross-sectional coordinate system.

[0025] By combining the equations of the straight line and the contour lines of each sidewall, a combined equation is obtained. This combined equation expresses the mathematical relationship that must be satisfied when points on the horizontal probe ray and points on the sidewall contour lines have the same abscissa and ordinate.

[0026] In the independent variable Within the range of values ​​for , solve the combined equations. The purpose of this solution is to find, for a given azimuth angle, the equations that satisfy the combined equations. The coordinate value corresponds to a potential intersection point between the probe ray projection and the sidewall profile.

[0027] If for a certain azimuth angle, in the independent variable... If there are one or more real solutions within the range of values ​​of , then a valid intersection point is determined to exist. The valid intersection point must satisfy the following conditions: its position is located on the actual physical contour of the tunnel cross section, and it is the first contour contact point encountered along the ray direction from the radar installation point projection.

[0028] By analyzing the conditions under which effective intersection points exist as the azimuth changes, the limiting azimuth angle corresponding to the sidewall is determined. The value of the limiting azimuth angle is determined by one of the following two geometric conditions: (1) Tangency condition: the azimuth angle value corresponding to when the horizontal probe ray is tangent to the sidewall contour line. This condition is equivalent to the existence of repeated roots in the range of values ​​of the combined equation, or it is obtained by solving the contour line derivative, which is equal to the azimuth tangent value, and simultaneously satisfying the combined equation.

[0029] (2) Boundary intersection condition: The azimuth angle value corresponding to the first intersection of the horizontal probe ray with a specific linear boundary segment in the sidewall contour (such as the top boundary of the vertical sidewall). This condition is obtained by directly substituting the equation describing the specific boundary (such as x being a constant or y being a constant) into the equation of the straight line and solving for it.

[0030] From the azimuth angle values ​​calculated under the above two conditions, select the azimuth angle value that minimizes the detection sector range, i.e., the absolute value of the azimuth angle, as the final limiting azimuth angle for that side.

[0031] The extreme azimuth angle calculated for the first sidewall is defined as the maximum effective detection boundary angle on the right, and the extreme azimuth angle calculated for the second sidewall is defined as the maximum effective detection boundary angle on the left.

[0032] The total detection field of view in the horizontal direction is determined based on the difference between the maximum effective detection boundary angles on the left and right sides.

[0033] Based on the radar normal direction, the total detection field of view is divided into N non-overlapping and continuously distributed predefined angle intervals.

[0034] As a specific example of the present invention, the number of predefined angle intervals N=4. Correspondingly, the sector division method is as follows: the total effective detection field of view is divided into four equal parts, forming four sectors continuously distributed from left to right. For example, when the total effective detection field of view is 90°, the angle intervals corresponding to the four sectors are: [-45°, -22.5°], [-22.5°, 0°], [0°, 22.5°], [22.5°, 45°], respectively.

[0035] Assign a unique sector number to each predefined angle interval and establish a spatial index mapping from the sector number to the center azimuth angle within its corresponding angle interval.

[0036] Based on the established discrete sectors and spatial index mapping, beamforming control of the radar transmitting array is required to achieve directional energy projection to each sector, according to the spatial index mapping. Therefore, in the second step (S2) of this invention, phase compensation and amplitude weighting are applied to the transmitted signals of each transmitting element in the radar transmitting antenna array according to the spatial index mapping, thereby sequentially forming and transmitting narrow beams pointing to each sector.

[0037] Specifically, for the current sector, the phase compensation of each transmitting element is calculated based on its center azimuth, element spacing, and signal wavelength. The calculation formula is as follows: .

[0038] In the formula, For the first The phase compensation amount is calculated for each transmitting element, and the unit is radians. The wavelength is the signal wavelength, measured in meters. This is the azimuth angle of the current sector center, measured from the normal direction of the array, in radians; The distance between adjacent elements in the transmitting antenna array, in meters; The index number of the transmitting array element represents the transmitting array element relative to the phase reference array element (usually indexed). The position number is a dimensionless integer.

[0039] This is the sine value of the azimuth angle of the current sector center, used to convert the azimuth angle into a spatial path difference scaling factor along the array axis.

[0040] The negative sign in the formula indicates phase delay compensation. When the target direction is... At this time, the array elements that are ahead of the reference array element need to be given a negative phase to compensate for their path lead caused by their spatial position, so as to finally align the phase of the radiation signal of all array elements in the target direction.

[0041] In the above formula, the first The phase compensation amount of each transmitting element is related to its index number, element spacing, and... The product is proportional. The formula is used to compensate for the path difference when electromagnetic waves are radiated from different array elements to the same direction in the far field, so that the transmitted signals of all array elements in the azimuth direction of the current sector center can be superimposed in phase.

[0042] To meet the sidelobe suppression requirement towards the tunnel sidewalls, in the spatial energy distribution pattern of the transmitted beam in the current sector formed by the transmitting antenna array, the sidelobe gain located outside the main lobe at the azimuth angle of the current sector center and towards the left and right sidewalls of the tunnel must be less than or equal to a preset safety threshold. The amplitude weights of each transmitting array element are optimized and solved.

[0043] The preset safety threshold is a maximum permissible sidelobe gain value set based on multipath interference caused by sidewall reflections in the tunnel environment. This value ensures that when the gain of the transmitted beam in the direction towards the tunnel sidewall does not exceed this value, the intensity of the multipath reflected signal is lower than the interference tolerance of the radar receiver. The specific value is determined through electromagnetic simulation software modeling based on the reflection coefficient of the tunnel sidewall material and the radar operating frequency. In the simulation model, the maximum sidelobe gain corresponding to an input signal-to-interference ratio (SIR) of not less than 10 dB is the preset safety threshold.

[0044] For each transmitting element, its phase compensation is converted into a complex phase factor, and its amplitude weighting factor is used as the amplitude scaling factor. The two are multiplied to obtain the transmitting beamforming weight under the complex shape of the corresponding transmitting element.

[0045] According to sector order, the corresponding transmit beamforming weights are applied in different transmit cycles to cyclically transmit narrow beams.

[0046] It should be noted that the second step of this invention transforms discrete angle commands into a physically achievable directional transmission mode. Through a narrow-beam scanning mechanism with sector-based time-division multiplexing, energy is concentrated and projected onto the current target sector during the transmission phase. Simultaneously, radiation towards non-target areas, especially tunnel sidewalls, is actively suppressed. This reduces the primary multipath signal energy generated by strong wall reflections at the physical source and provides clear echo data by sector and time slot for subsequent received signal processing.

[0047] After directional scanning of the transmit beam is achieved through step S2, time-division transmitted signals pointing to each sector are obtained. However, even if the transmitted energy is directionally focused, the echoes collected by the receiving antenna array may still contain residual multipath reflections and environmental clutter from non-target directions. To achieve spatial selectivity at the receiving end that matches the transmit beam and further improve signal quality, appropriate spatial filtering processing of the received signal is required.

[0048] Therefore, the third step (S3) of the method of the present invention is as follows: for each transmitted narrow beam, digital beamforming is performed on the echo signals of each channel of the receiving antenna array, and the received beam is synthesized by complex weighting to output the spatial filtering signal corresponding to each sector. Specifically, refer to... Figure 2 As shown: Analog-to-digital conversion is performed on the echo signals of each channel of the receiving antenna array to recover the multi-channel digital signal containing distance and Doppler information.

[0049] The center azimuth angle of the sector pointed to by the current transmit narrow beam is set as the desired receiving beam pointing angle. Based on the desired receiving beam pointing angle, the element spacing of the receiving antenna array, and the signal wavelength, the required phase compensation amount for each receiving element is calculated as the complex exponential amplitude. This calculation process is consistent with the calculation process of the phase compensation amount for each transmitting element, and will not be elaborated here.

[0050] The initial complex weight values ​​of the corresponding array elements of the receiving antenna array are calculated using Euler's formula, and then arranged in the order of the array elements to form the receiving complex weight vector.

[0051] The cross-correlation coefficient between each channel signal pair of the multi-channel digital signal is obtained by using the Pearson product-moment correlation coefficient calculation formula. The average cross-correlation coefficient between all channel signal pairs is obtained by arithmetic mean. If the average cross-correlation coefficient is less than the preset cross-correlation standard value, it is determined that the channel echo phase is inconsistent. The received complex weight vector is adaptively adjusted to form null in the non-target direction.

[0052] The specific implementation process of adaptively adjusting the received complex weight vector is as follows: Based on all channel echo data received by the current sector in this scanning cycle, the sample covariance matrix is ​​calculated and eigenvalue decomposition is performed to obtain multiple eigenvalues ​​and their corresponding eigenvectors under the number of corresponding receiving array elements.

[0053] Arrange the features in descending order of their values, and identify the space spanned by the feature vectors corresponding to the preset number of feature values ​​before sorting as the signal subspace. The signal subspace contains components of the desired target signal and strong interference signals.

[0054] Within the signal subspace, candidate angles are traversed at a fixed step size within a preset azimuth range. For each candidate angle, an array steering vector is constructed to calculate its spatial spectral power. The spatial spectral power sequence is integrated, and its mean and standard deviation are calculated. The mean and three times the standard deviation are defined as the significant spectral peak reference threshold. If the spatial spectral power reaches or exceeds the significant spectral peak reference threshold, it is marked as a significant peak. One or more significant peaks other than the desired target direction are identified, and the angles corresponding to each significant peak are estimated as the direction of strong interference waves in the current tunnel environment.

[0055] Based on the estimated direction of strong interference, an adaptive receiving weight vector is calculated using a linearly constrained minimum variance criterion. This linearly constrained minimum variance criterion is based on the expected direction... Maintain unity gain in the direction of strong interference Forming zero traps as constraints, constructing a constraint matrix and response matrix .

[0056] The adaptive receiving weight vector is solved using the following mathematical expression: .

[0057] In the formula, The inverse matrix of the sample covariance matrix is ​​represented by the matrix in which the sample covariance matrix is ​​expressed. This represents the conjugate transpose of the constraint matrix, which transforms the projection relationship between the weight vector and the constraint matrix into a mathematical expression of the constraint conditions.

[0058] The constraint matrix is ​​in The inverse of the projection matrix under the metric. Used to ensure that the output power is minimized while satisfying the constraints.

[0059] The physical meaning of this mathematical expression is: Under the constraint conditions are satisfied... Under the premise of maintaining unity gain in the desired direction and forming nulls in the interference direction, the array output power is minimized. To find the optimal weight vector Minimizing output power is equivalent to maximally suppressing interference and noise, since the target signal direction has been constrained and protected.

[0060] It should be noted that the aforementioned preset cross-correlation standard value is a statistical benchmark for the inter-channel cross-correlation coefficients calculated by collecting background echo signals after radar installation or during periodic calibration, under the condition that there are no strong interfering targets in the current tunnel environment. This statistical benchmark is used to characterize the correlation level when the receiving channels are in a normal, consistent state.

[0061] As an optional implementation, a time period with no real moving targets and no radio interference in the current tunnel environment is selected, and multiple frames (e.g., 100 frames) of raw echo data from the radar receiving channels are continuously acquired. For each frame of data, the complex signal cross-correlation coefficient or the Pearson correlation coefficient based on signal amplitude is calculated between all pairs of receiving channels, and the average cross-correlation coefficient of all channel pairs in that frame is calculated. The average cross-correlation coefficients of all channel pairs in each frame are integrated to construct an average cross-correlation coefficient sequence, and the mean and standard deviation of this sequence are calculated. A preset cross-correlation standard value is set as the difference between the sequence mean and a preset multiple of the standard deviation, which is typically 2 or 3.

[0062] The Euler formula and the Pearson product-moment correlation coefficient calculation formula mentioned above are existing formulas and will not be elaborated here.

[0063] After spatial filtering of the received signals from each sector, isolated echo data for each sector was obtained. However, this type of data has two limitations: first, the signals from each sector are isolated from each other and cannot directly support overall situational awareness across sectors; second, the data is still represented in a radar-centric spherical coordinate system, which does not match the physical layout of the tunnel environment. To achieve full-section collaborative target detection in tunnel scenarios, the signals from each sector must be converted into a unified feature representation and effective cross-sector information fusion must be performed.

[0064] Therefore, the fourth step (S4) of this invention performs spatial transformation and feature fusion from sector-specific signals to global joint features, which is specifically divided into the following two sub-steps: (Refer to...) Figure 3 As shown, S41. Convert the spatial filtering signal of each sector into a feature tensor, including: sequentially performing Fast Fourier Transform (FFT) on the spatial filtering signal of each sector in the range dimension, Doppler dimension, and angle dimension to analyze the radial range, radial velocity, and horizontal angle of arrival of the target, generating a three-dimensional complex data array. The number of FFT points in the range dimension is equal to the number of radar signal sampling points, the number of FFT points in the Doppler dimension is equal to the number of pulses within a coherent processing interval, and the number of FFT points in the angle dimension is equal to the number of receiving antenna array elements. Before performing the FFT, Hamming windows need to be applied in each dimension to reduce spectral leakage.

[0065] The magnitude of each cell in the three-dimensional complex data array is calculated to obtain a three-dimensional energy matrix. The three dimensions of the three-dimensional energy matrix correspond to the distance index, velocity index, and angle index, respectively. Each element value represents the signal strength within the distance-velocity-angle resolution cell.

[0066] The three-dimensional energy matrix is ​​normalized by using the global maximum-minimum normalization method, which linearly scales all elements of the matrix to the interval [0, 1].

[0067] Under the condition that the total length of the corresponding dimension is less than the requirement, the size of the custom sliding window in the distance, velocity and angle dimensions is set, the sliding step size is 1, the median of all elements in each sliding window is calculated, and the value of the element at the center position of the window is replaced with the median value to achieve sliding window statistical filtering and noise reduction, and the remaining cells are marked as valid cells.

[0068] Radar installation height With beam pitch angle The physical location represented by each effective unit in the three-dimensional energy matrix is ​​mapped from a spherical coordinate system with the radar as the origin to a horizontal bird's-eye view coordinate system with the tunnel floor as the reference. Specifically, the mapping process involves mapping the slant range corresponding to each effective unit. Horizontal azimuth The coordinates of the effective element in the horizontal bird's-eye view coordinate system are calculated using the following set of formulas. : .

[0069] In the formula, The cosine of the beam pitch angle is used as a geometric projection factor to convert the slant range into the ground distance on the horizontal plane.

[0070] These are the sine and cosine values ​​of the horizontal azimuth, respectively, used to decompose the horizontal distance into two mutually perpendicular coordinate components in the bird's-eye view coordinate system.

[0071] This set of formulas completes the mapping from a spherical coordinate system with the radar as the origin to a horizontal bird's-eye view coordinate system with the tunnel floor as the reference. Its physical essence is: projecting the three-dimensional spatial units detected by the radar vertically onto the horizontal ground to solve for the corresponding two-dimensional planar coordinates. .

[0072] The mapped data are stitched and aligned along the channel dimension according to the azimuth angle of the beam center of each sector to form a multi-channel feature tensor.

[0073] S42. Perform cross-interval feature fusion on the feature tensor to generate joint spatial features. This process can be executed by designing a multi-branch feature fusion network, specifically including: S421. Splitting the multi-channel feature tensor into sub-tensors corresponding to different sectors according to its channel dimension, and inputting them into multiple parallel convolutional backbone branches of the network. Each backbone branch consists of at least one convolutional layer, a batch normalization layer, and a ReLU activation function layer connected sequentially, used to extract primary features from the input sub-tensors respectively. The primary features include local morphological features learned by the convolutional kernel that can characterize the shape of the target within the corresponding sector, as well as target motion intensity change features captured by convolution operations.

[0074] S422. At the same level depth in each backbone branch, the initial features output by each branch are spliced ​​along the channel dimension to form the preliminary fused features of that level. The preliminary fused features maintain consistent resolution in the spatial dimension, and their number of channels is the sum of the number of channels of the initial features of each branch.

[0075] S423. Based on the maximum signal energy value generated when the original three-dimensional energy matrix is ​​mapped to the bird's-eye view coordinate system, the spatial reliability weight of each two-dimensional pixel position is obtained by global maximum value normalization.

[0076] S424. Monitor in real time the ratio of the average background noise intensity to the peak energy of the echo signal of each sector channel group, and label it as the signal-to-noise ratio.

[0077] When the signal-to-noise ratio of a sector exceeds the preset signal-to-noise ratio interference threshold, it is determined that the sector has strong specular reflection interference. Penalty coefficients corresponding to different threshold amplitude ranges are predefined. Based on the threshold amplitude range of the sector with strong specular reflection interference, the corresponding penalty coefficient is called and multiplied by the spatial reliability weight to obtain the corresponding weight reduction value of the sector in feature fusion. At the same time, the weight of the adjacent sector channel group in feature fusion is increased.

[0078] As a specific example, the preset signal-to-noise ratio interference threshold is 15dB; when the sector signal-to-noise ratio exceeds the preset signal-to-noise ratio interference, the penalty coefficient is determined according to the following rules: 0.9 for exceeding the threshold by 5dB; 0.7 for exceeding by 10dB; and 0.5 for exceeding by more than 10dB.

[0079] Optimized weights are used to weight the preliminary fusion features at each level in terms of both channel and spatial dimensions.

[0080] S425. Perform upsampling and multi-level information aggregation on the weighted optimized features to output joint spatial features.

[0081] After obtaining the joint spatial features that integrate information from multiple sectors, these features need to be decoded into specific target parameters. To this end, the fifth step (S5) of this invention detects candidate targets based on the joint spatial features and, in conjunction with tunnel physical boundary information, performs false alarm discrimination and filtering on the candidate targets. The final target's position, size, and category information are then regressed and output in the bird's-eye view coordinate system. This is achieved through a pre-defined single-stage target detection network. The execution process of this network is as follows: the single-stage target detection network takes the joint spatial features as input and calls a detection head that has been learned during the training phase to decode target parameters from the features.

[0082] Before the detection head executes, a series of anchor frames of different scales and aspect ratios need to be predefined in the bird's-eye view coordinate system. The anchor frames are set a priori based on the typical physical dimensions of common targets (such as pedestrians and vehicles) in the tunnel scene, and are evenly laid on the spatial units characterized by the joint spatial features. Each anchor frame is defined by its center coordinates, width, and length.

[0083] The detection head contains two parallel sub-network branches: a classification sub-branch consisting of one or more convolutional layers, responsible for outputting the confidence score of the presence of a target within each anchor box, and the probability distribution of the target belonging to each predefined category, which includes at least people, vehicles, and background.

[0084] The regression sub-branch consists of convolutional layers and is responsible for outputting the geometric correction parameters for each anchor frame, including the center point offset and the size scaling. The center point offset is divided into the horizontal and vertical axis offsets of the center point, and the size scaling is divided into the length scaling and the width scaling.

[0085] Normalization is applied to the category probability distribution of each anchor frame to obtain its probability value belonging to the categories of people, vehicles, and background.

[0086] The highest class probability value is used as the confidence score of the anchor box, and targets with confidence scores exceeding a set confidence threshold are marked as candidate targets.

[0087] For each candidate target, the following false alarm detection is performed: Sector energy consistency test: Extract the energy value of the corresponding spatial location of the candidate target on each sector channel, calculate the ratio of the standard deviation of the energy value to the mean, and define it as the energy discretization coefficient of the corresponding spatial location of the candidate target. If the energy discretization coefficient exceeds the false alarm sample calibration threshold, it indicates that the energy distribution of each sector is inconsistent, and the candidate target is judged to have a false alarm. The false alarm sample calibration threshold can be determined by collecting the energy discretization coefficients of several (e.g., 1000) false alarm samples generated by a single strong reflection from the sidewall and taking their average value.

[0088] Motion trajectory rationality check: Combining the detection results of previous multiple frames, the motion trajectory of the candidate target is fitted and the maximum angle between the motion trajectory and the tunnel axis is calculated. If the maximum angle exceeds the angle between the tunnel physical boundary and the tunnel axis as marked by the tunnel design drawings, it indicates that the motion trajectory is unreasonable, and the candidate target is judged to have a false alarm. Special note: For newly appearing candidate targets that do not yet have historical trajectories, the motion trajectory rationality check is skipped.

[0089] Boundary compliance check: Calculate the minimum distance between the circumscribed rectangle of the candidate target in the bird's-eye view and the physical boundary of the tunnel. If the minimum distance is less than 0, it indicates that part of the candidate target is located outside the boundary, and the candidate target is judged to have a false alarm.

[0090] For candidate targets identified as having false alarms, their confidence scores are reduced according to a preset penalty rule. The preset penalty rule is as follows: a penalty factor is pre-defined for each of the above false alarm criteria. The penalty factor is a positive number less than 1, and the magnitude of the penalty factor corresponding to different conditions inversely reflects the severity of the false alarm indicated by that condition. For each candidate target, its triggered false alarm criteria are checked sequentially. For each triggered condition, its confidence score is multiplied by the corresponding penalty factor to obtain the reduced confidence score.

[0091] The penalty factors corresponding to the three false alarm detection conditions can be set, in descending order of severity, as follows: 0.2 for boundary compliance test, 0.5 for trajectory rationality test, and 0.7 for sector energy consistency test. When a candidate target triggers multiple false alarm detection conditions simultaneously, its confidence score will be continuously multiplied by the corresponding penalty factor.

[0092] If the reduced confidence score still exceeds the set confidence threshold, the target is deemed valid and the specific category is recorded.

[0093] For each anchor frame deemed valid, its geometric correction parameters are applied to the anchor frame's preset center coordinates, length, and width: the center point's horizontal axis offset is multiplied by the anchor frame's width and superimposed onto the anchor frame's preset center horizontal coordinates to obtain the target's two-dimensional center horizontal coordinates in the bird's-eye view coordinate system.

[0094] Multiply the center point's vertical axis offset by the anchor frame length and superimpose it onto the anchor frame's preset center vertical coordinates to obtain the target's two-dimensional center vertical coordinates in the bird's-eye view coordinate system.

[0095] Substitute the width scaling factor into the standard exponential function, and use the sum of the function output value and the anchor frame width as the width of the target in the bird's-eye view coordinate system.

[0096] Similarly, by substituting the length scaling factor into the standard exponential function, the sum of the function output value and the anchor frame length is taken as the length of the target in the bird's-eye view coordinate system.

[0097] Based on the confidence scores of all valid anchor boxes, they are sorted. A non-maximum suppression algorithm is applied to eliminate redundant spatially overlapping detection boxes; that is, the valid anchor box with the highest confidence score is selected sequentially, and the remaining boxes with an intersection-union ratio (IUU) greater than 0.5 are discarded. This process is repeated until all valid anchor boxes have been processed, and the final target's category, location, and size information are output.

[0098] It should be noted that during the training phase of the single-stage object detection network, a dataset labeled with the real object's location, size, and category is used to iteratively optimize the parameters of the classification and regression sub-branches through a loss function that includes classification loss and regression loss.

[0099] In the application stage of the method of this invention, the pre-trained parameters of the network can be directly called to perform forward calculation on the input joint spatial features, and directly output the class probability, confidence score and geometric correction parameters of all anchor boxes. Then, through the above decoding process and non-maximum suppression steps, the final target list is obtained.

[0100] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined by the present invention, they should all fall within the protection scope of the present invention.

Claims

1. A target detection method for tunneling millimeter-wave radar based on beamforming, characterized in that, include: The horizontal monitoring area inside the tunnel in front of the radar is divided into multiple sectors, and a spatial index mapping between the sectors and the azimuth angle is established. According to the spatial index mapping, phase compensation and amplitude weighting are applied to the transmitted signals of each transmitting element in the radar transmitting antenna array, and narrow beams pointing to each sector are formed and transmitted in sequence. For each transmit narrow beam, digital beamforming is performed on the echo signals of each channel of the receiving antenna array. The received beam is synthesized by complex weighting and outputs the spatial filtering signal corresponding to each sector. The spatial filtering signals of each sector are converted into feature tensors, and cross-interval feature fusion is performed on the feature tensors to generate joint spatial features. Based on the joint airspace features, candidate targets are detected, and false alarms are identified and filtered out by combining tunnel physical boundary information. The final target's position, size, and category information are then regressed and output in the bird's-eye view coordinate system.

2. The tunnel millimeter-wave radar target detection method based on beamforming according to claim 1, characterized in that, The horizontal monitoring area within the tunnel in front of the radar is divided into multiple sectors, including: Based on the tunnel's geometric topology and radar installation height, the azimuth angle of the horizontal detection ray emitted from the radar installation point intersects with the theoretical contour lines of the tunnel's two side walls, thus obtaining the maximum effective detection boundary angles on the left and right sides respectively. The total detection field of view in the horizontal direction is determined based on the difference between the maximum effective detection boundary angles on the left and right sides. Based on the radar normal direction, the total detection field of view is divided into several non-overlapping and continuously distributed predefined angle intervals; Assign a unique sector number to each predefined angle interval and establish a spatial index mapping from the sector number to the center azimuth angle within its corresponding angle interval.

3. The tunneling millimeter-wave radar target detection method based on beamforming according to claim 2, characterized in that, The sequential formation and transmission of narrow beams directed at each sector includes: For the current sector, the phase compensation amount of each transmitting element is calculated based on its center azimuth, element spacing and signal wavelength; The amplitude weights of each transmitting array element are optimized by taking the requirement of sidelobe suppression towards the tunnel sidewall as a constraint. Generate the transmit beamforming weights for the current sector based on the phase compensation amount and amplitude weights; According to sector order, the corresponding transmit beamforming weights are applied in different transmit cycles to cyclically transmit narrow beams.

4. The tunnel millimeter-wave radar target detection method based on beamforming according to claim 3, characterized in that, The sidelobe suppression requirements toward the tunnel sidewall include: In the spatial energy distribution pattern of the transmitted beam for the current sector formed by the transmitting antenna array, the side lobe gain located outside the main lobe at the azimuth angle of the current sector center and towards the tunnel sidewall must be less than or equal to a preset safety threshold.

5. A tunneling millimeter-wave radar target detection method based on beamforming according to claim 2, characterized in that, The digital beamforming of the echo signals from each channel of the receiving antenna array includes: The echo signals from each channel of the receiving antenna array are converted from analog to digital to recover the multi-channel digital signal containing distance and Doppler information; The center azimuth angle of the sector pointed to by the current transmitting narrow beam is set as the desired receiving beam pointing angle. Based on the desired receiving beam pointing angle, the element spacing of the receiving antenna array, and the signal wavelength, the required phase compensation amount for each receiving element is calculated as the complex exponential argument. The initial complex weight values ​​of the corresponding elements of the receiving antenna array are calculated using Euler's formula and arranged in the order of the elements to form the receiving complex weight vector. Calculate the average cross-correlation coefficient between all channel signal pairs of the multi-channel digital signal. If the average cross-correlation coefficient is less than the cross-correlation standard value, it is determined that the channel echo phases are inconsistent. Adaptively adjust the received complex weight vector to form nulls in non-target directions. Using the adjusted received complex weight vector, complex multiplication operations are performed on each digital signal, and the results of all complex multiplication operations are accumulated to output the spatial filtering signal of the current sector.

6. The tunnel millimeter-wave radar target detection method based on beamforming according to claim 1, characterized in that, The process of converting the spatial filtering signal of each sector into a feature tensor includes: Fast Fourier transforms are performed sequentially on the spatial filtered signals of each sector in the range dimension, Doppler dimension, and angle dimension to analyze the radial distance, radial velocity, and horizontal angle of arrival of the target, generating a three-dimensional complex data array. The magnitude of each cell in the three-dimensional complex data array is calculated to obtain a three-dimensional energy matrix. The three dimensions of the three-dimensional energy matrix correspond to the distance index, velocity index and angle index, respectively. Each element value represents the signal intensity within the distance-velocity-angle resolution cell. The three-dimensional energy matrix is ​​normalized and noise is reduced using sliding window statistical filtering. Based on the radar's installation height and beam elevation angle, the physical location represented by each effective unit in the three-dimensional energy matrix is ​​mapped from the spherical coordinate system with the radar as the origin to the horizontal bird's-eye view coordinate system with the tunnel floor as the reference. The mapped data are stitched and aligned along the channel dimension according to the azimuth angle of the beam center of each sector to form a multi-channel feature tensor.

7. The tunnel millimeter-wave radar target detection method based on beamforming according to claim 1, characterized in that, The method of performing cross-interval feature fusion on feature tensors to generate joint spatial features includes: Feature extraction is performed on the channel groups corresponding to each sector in the multi-channel feature tensor to obtain the local morphological features representing the target shape and the motion intensity features representing the target velocity changes in the signal of each sector. Features extracted from different sectors are spliced ​​across channels to form preliminary fused features; Based on the maximum signal energy value recorded at each two-dimensional position when the original three-dimensional energy matrix is ​​mapped to the bird's-eye view coordinate system, a normalized spatial reliability weight is generated. The initial fusion features are optimized by weighting spatial reliability. The weighted and optimized features are upsampled and multi-level information is aggregated to output joint spatial features.

8. The tunnel millimeter-wave radar target detection method based on beamforming according to claim 7, characterized in that, The step of weighting and optimizing the preliminary fusion features using spatial reliability weights includes: The ratio of the average background noise intensity to the peak energy of the echo signal in each sector channel group is monitored in real time and denoted as the signal-to-noise ratio. When the signal-to-noise ratio of a sector exceeds the preset signal-to-noise ratio interference threshold, it is determined that the sector has strong specular reflection interference. The weight of the sector in feature fusion is dynamically reduced according to the threshold amplitude, while the weight of its adjacent sectors in feature fusion is increased simultaneously, and the sum of the weights of all sectors in feature fusion remains unchanged.

9. The tunnel millimeter-wave radar target detection method based on beamforming according to claim 1, characterized in that, The regression output of the final target's position, size, and category information in the bird's-eye view coordinate system includes: Associating a set of anchor frames of predefined sizes on the spatial units characterized by the joint spatial features; Through convolution operations, the class probability distribution of each anchor box and the corresponding geometric correction parameters are generated by decoding. The geometric correction parameters include center point offset and size scaling. Normalization is applied to the category probability distribution of each anchor frame to obtain its probability value belonging to the categories of people, vehicles and background, respectively. The maximum category probability value is used as the confidence score of the anchor box. The existence of a valid target and its specific category within the anchor box is determined based on whether the confidence score exceeds a set confidence threshold. For each anchor frame that is determined to be valid, its geometric correction parameters are applied to the preset center coordinates and dimensions of the anchor frame to calculate the two-dimensional center coordinates, length and width of the target in the bird's-eye view coordinate system; The target is sorted according to the confidence scores of all valid anchor boxes, and a non-maximum suppression algorithm is applied to eliminate redundant spatially overlapping detection boxes. Finally, the target's category, location, and size information are output.

10. A tunneling millimeter-wave radar target detection method based on beamforming according to claim 9, characterized in that, The determination of whether there is a valid target and its specific category within the anchor frame includes: Targets with confidence scores exceeding a set confidence threshold are marked as candidate targets; Based on the consistency of energy distribution in each sector of the joint airspace features, and the conformity between the motion trajectory of the candidate target and the physical boundary of the tunnel across multiple frames, false alarms are identified for the candidate targets. For candidate targets identified as false alarms, their confidence scores are reduced according to preset penalty rules. If the reduced confidence score still exceeds the set confidence threshold, the target is deemed valid and its specific category is recorded.

Images