Anti-interference and multi-target positioning method for vehicle-mounted millimeter wave radar

Abstract

The invention provides an anti-interference and multi-target positioning method for a vehicle-mounted millimeter-wave radar, and the method comprises the steps: constructing a third-order cumulant interference suppression module, carrying out the processing of a signal received by the vehicle-mounted millimeter-wave radar through a third-order cumulant algorithm, suppressing the mutual interference between radars, and generating a high-resolution distance-angle thermodynamic diagram; constructing a dual-channel adaptive edge detection module, dividing the generated thermodynamic diagram into an original channel and an edge detection channel, dynamically allocating weights based on local gradient intensity, and fusing and outputting an enhanced thermodynamic diagram; and constructing a target identification and positioning module, performing target detection and positioning on the enhanced thermodynamic diagram through an improved YOLOv8 neural network, and outputting distance and angle information of the target. According to the method, the problems of mutual interference between radars, difficulty in detection of weak and small targets, low separation precision of overlapped targets and the like in a complex traffic scene are solved by combining high-order statistical signal processing and deep learning technologies, and the robustness and the real-time performance of target positioning are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent driving and radar signal processing, and more specifically, to an anti-interference and multi-target positioning method for a vehicle-mounted millimeter-wave radar. Background Art

[0002] Millimeter-wave radar, capable of providing reliable perception data even in adverse weather and low-light conditions, has become a key sensor for applications such as vehicle positioning, obstacle avoidance, and autonomous driving in intelligent transportation and autonomous driving. By transmitting electromagnetic waves in the millimeter-wave frequency band and receiving reflected signals, it can sense information such as the distance, speed, and angle between the vehicle and surrounding objects, providing crucial environmental awareness support for intelligent driving systems.

[0003] Currently, millimeter-wave radar signal processing and target detection primarily rely on traditional signal processing methods and deep learning approaches. Traditional methods, such as wavelet denoising and compressed sensing, can improve target detection accuracy to a certain extent. Deep learning methods, particularly the YOLO series of target detection algorithms based on convolutional neural networks (CNNs), such as YOLOv8, have become a mainstream approach for millimeter-wave radar target detection due to their speed and accuracy, driving the development of target detection technology.

[0004] However, existing technologies still have many limitations. On the one hand, traditional signal processing methods are computationally complex and rely on assumptions, making them difficult to adapt to complex and changing environments. Furthermore, step-by-step processing can lead to cumulative errors. On the other hand, while deep learning models such as YOLOv8 have great potential for target detection, they struggle to accurately identify small targets in complex backgrounds, particularly in the context of millimeter-wave radar signals. Furthermore, they are prone to false detections or missed detections in densely overlapping target scenarios. This results in a high missed detection rate for small targets and insufficient separation accuracy for overlapping targets, making them unable to meet the high accuracy and robustness requirements of intelligent driving systems for millimeter-wave radar target detection. Summary of the Invention

[0005] To address the aforementioned technical issues, a method for anti-interference and multi-target positioning for vehicle-mounted millimeter-wave radars is provided. By combining high-order statistical signal processing with deep learning techniques, this method addresses issues such as mutual interference between radars, difficulty detecting small and weak targets, and low accuracy in separating overlapping targets in complex traffic scenarios, thereby improving the robustness and real-time performance of target positioning.

[0006] The technical means adopted in the present invention are as follows:

[0007] An anti-interference and multi-target positioning method for a vehicle-mounted millimeter-wave radar, comprising:

[0008] S1. Construct a third-order cumulant interference suppression module. The third-order cumulant interference suppression module processes the signals received by the vehicle-mounted millimeter-wave radar using the third-order cumulant algorithm, suppresses the mutual interference between radars, and generates a high-resolution range-angle heat map.

[0009] S2. Construct a dual-channel adaptive edge detection module. The dual-channel adaptive edge detection module divides the generated heat map into an original channel and an edge detection channel, dynamically assigns weights based on the local gradient intensity, and fuses and outputs an enhanced heat map.

[0010] S3. Construct a target recognition and positioning module. The target recognition and positioning module detects and locates the targets in the enhanced heat map through an improved YOLOv8 neural network, and outputs the distance and angle information of the targets.

[0011] Further, step S1 specifically includes:

[0012] S11. Utilize the synchronously collected array signals and construct a third-order cross-cumulant matrix according to the non-correlation characteristics of the echo signals and interference signals.

[0013] S12. Obtain a two-dimensional spatial spectrum according to the subspace method, realize the joint estimation of range and azimuth, and generate a high-resolution range-angle heat map.

[0014] Further, step S11 specifically includes:

[0015] S111. Represent the radar transmitted signal s T (t) as:

[0016]

[0017] where A t is the amplitude of the transmitted signal, f0 is the carrier frequency of the signal, μ = B / T, B is the signal bandwidth, and T is the modulation period.

[0018] S112. Represent the echo signal s R (t) as:

[0019]

[0020] where A r is the amplitude of the echo signal, and τ is the time delay.

[0021] S113. Assume that there are K far-field target echo signals from different directions θ k (k = 0, 1,..., K - 1) and K I interference signals, and K I < K. Represent the difference-frequency signal received by the m-th array element in the linear array as:

[0022]

[0023] Where m = 0, 1, ..., M-1, a m is the echo signal steering vector, s k (t) represents the kth echo difference frequency signal, b m is the interference signal steering vector, ω i (t) represents the difference frequency signal of the i-th interference, u m (t) represents superimposed Gaussian white noise;

[0024] S114. Based on the frequency modulated continuous wave characteristics of the radar transmission, a reference signal matrix is ​​generated as follows:

[0025]

[0026] Where n=0,1,…,N-1 represents the time series index, N represents the total number of sampling points, and f S Indicates the sampling rate;

[0027] S115. Introduce the time delay variable τ and construct the delay signal matrix as follows:

[0028]

[0029] S116: Perform discrete sampling on the formula in step S113 to obtain x m (n), and with s r (n+τ) and s r (n) together constitute the third-order cumulant expression. According to the properties of the third-order cumulant and the matrix model, the third-order cross-cumulant matrix is ​​constructed through the Hadamard product as follows:

[0030]

[0031] Where X is the received signal matrix, S τ is the delayed reference signal matrix, S * is the conjugate matrix of the reference signal, and N is the number of sampling points.

[0032] Furthermore, step S12 specifically includes:

[0033] S121. Perform eigenvalue decomposition on the third-order cross-cumulant matrix to extract the noise subspace, thereby suppressing interference components unrelated to the reference signal;

[0034] S122, combined with the target orientation vector a θ , take C by row τ The average value is c τ , construct a two-dimensional spatial spectrum as follows:

[0035] C(r,θ)=(c τ * ·a θ )(a θ * c τ )

[0036] Where r is the target distance, θ is the azimuth angle, and the symbol * represents conjugate transpose;

[0037] S123. Finally, a high-resolution distance-angle heat map is generated, where each peak corresponds to the distance and angle of a target.

[0038] Furthermore, step S2 specifically includes:

[0039] S21, performing Gaussian filtering, gamma correction and normalization on the original thermal map channel to retain the global energy distribution of the target;

[0040] S22. An improved Canny edge detection algorithm is used for the edge detection channel to calculate the horizontal and vertical gradients of the heat map, dynamically set high and low thresholds, retain strong edges and weak edges connected to strong edges, and suppress isolated noise points through non-local mean filtering;

[0041] S23. Dynamically assign weights to the two channels based on the local gradient strength. In strong edge regions, the edge channel is given a higher weight. In transition regions, the weights of the two channels are equal. In flat regions, the weights of the edge channel are reduced. Finally, the enhanced heat map is output through fusion.

[0042] Furthermore, in step S22:

[0043] The Sobel operator is used to calculate the horizontal gradient Gx and vertical gradient Gy of the heat map. The calculation formula for the gradient amplitude is as follows:

[0044]

[0045] Where Gx is the horizontal gradient of the heat map, and Gy is the vertical gradient of the heat map;

[0046] The high and low thresholds set dynamically are T high =0.3×G max , T low =0.1×G max .

[0047] Furthermore, step S3 specifically includes:

[0048] S31. Based on the simulation environment, a millimeter-wave radar dataset is constructed to simulate multi-target scenarios on ordinary urban roads.

[0049] S32. Input the enhanced heat map into the improved YOLOv8 neural network, and improve the detection rate of small objects and the accuracy of overlapping object separation by embedding the channel attention mechanism and adopting the Focal Loss loss function;

[0050] S33. Combine the spatial spectrum information of the heat map with the output information of the neural network to accurately calculate the target distance and angle coordinates.

[0051] Furthermore, step S31 specifically includes:

[0052] S311. Using the simulation platform MATLAB, build a typical urban road model, including common elements such as multiple lanes and speed limits, and randomly generate multiple target vehicles in the scene;

[0053] S312. Based on the operating principle and parameters (transmit power, bandwidth, sampling rate, etc.) of a 77GHz automotive millimeter-wave radar, generate an original signal containing multiple target echoes and co-frequency interference in a simulation platform, and simultaneously inject varying degrees of noise.

[0054] S313. Automatically annotate each frame of radar echo data obtained by simulation in combination with the true value information (real distance, azimuth) in the simulation to generate a corresponding distance-angle heat map and target bounding box;

[0055] S314. Divide the generated data set into a training set and a test set, and perform data augmentation operations such as random rotation and Gaussian noise on some of the data to improve the robustness of the network to weak targets and low signal-to-noise ratio scenarios;

[0056] S315. In the later stage of simulation, abnormal samples (such as those that fail to be injected into the target correctly or are distorted due to excessive interference) are screened and eliminated to ensure the validity and accuracy of the data.

[0057] Furthermore, step S32 specifically includes:

[0058] S321. The initial convolutional layers in the backbone of the YOLOv8 neural network are pruned, retaining only a single downsampling layer. This reduces computational complexity while retaining more high-frequency detail features. This design enables the complete transfer of texture information from the 320×320 high-resolution layer to subsequent networks, particularly improving the recognizability of edge features of small objects.

[0059] S322, Embed a Lightweight Channel Attention Mechanism (ECA-Net). A lightweight ECA module is inserted into the core path of each downsampling stage to form a downsampling-channel calibration-feature fusion cascade architecture. Channel weights are dynamically calibrated through global average pooling and 1D convolution to enhance the edge feature response of weak targets and suppress background noise channels.

[0060] S323. Reconstruct the loss function to optimize classification-regression collaborative training, replace the original classification loss function with the Focal Loss loss function, focus on difficult samples in low signal-to-noise ratio areas, and retain the joint optimization of YOLOv8's DFL Loss and CIoULoss as the bounding box regression loss to reduce the positioning error of dense targets.

[0061] Furthermore, in step S323, the Focal Loss loss function introduces a modulation factor (1-p t ) γ , the formula is as follows:

[0062] FL=-α t (1-p t ) γ log(p t )

[0063] Among them, p t is the model’s predicted probability for the true category, α t is the sample balancing factor.

[0064] Compared with the prior art, the present invention has the following advantages:

[0065] 1. The present invention provides an anti-interference and multi-target positioning method for a vehicle-mounted millimeter-wave radar. This method utilizes a third-order cumulant algorithm to process the signals received by the vehicle-mounted millimeter-wave radar. This method can effectively separate the nonlinear interference components in the target signal, effectively suppressing mutual interference between vehicle-mounted radars, while also taking into account target detection accuracy in low signal-to-noise ratio environments. This method achieves the technical effect of improving the accuracy and reliability of radar target detection in complex electromagnetic environments.

[0066] 2. The present invention provides an anti-interference and multi-target positioning method for a vehicle-mounted millimeter-wave radar. It adopts a dual-channel adaptive fusion strategy, which not only significantly enhances the edge features of weak targets, but also effectively solves the problems of blurred edges and difficult boundary separation of overlapping targets in dense scenes through dynamic weight allocation and non-local noise suppression technology. It also reduces edge weights in flat areas, reduces false alarms and misdetection of overlapping targets, thereby achieving the technical effect of improving the accuracy of target edge feature extraction and reducing the false detection rate in complex traffic scenes.

[0067] 3. The present invention provides an anti-interference and multi-target positioning method for a vehicle-mounted millimeter-wave radar. It uses an attention mechanism to guide the network to pay more attention to the target area, reduce attention to the background or irrelevant areas, suppress background noise interference, and at the same time use the Focal Loss loss function to focus on difficult samples in low signal-to-noise ratio areas, optimize classification-regression collaborative training, and reduce the positioning error of dense targets, thereby achieving the technical effect of improving target detection accuracy and positioning accuracy, and enhancing the adaptability and robustness of the system in complex environments.

[0068] 3. The present invention provides an anti-interference and multi-target positioning method for a vehicle-mounted millimeter-wave radar. By simulating various situations in real traffic scenarios, it generates an original signal containing multi-target echoes and co-frequency interference, and automatically labels and enhances the data, ensuring the validity and accuracy of the data, thereby achieving the technical effect of improving the training quality and generalization ability of the neural network model.

[0069] Based on the above reasons, the present invention can be widely promoted in fields such as intelligent driving and radar signal processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0071] Figure 1 Flow chart of the method of the present invention.

[0072] Figure 2 This is a specific working flow diagram of the third-order cumulant interference suppression module of the present invention.

[0073] Figure 3 This is a specific workflow diagram of the dual-channel adaptive edge detection module of the present invention.

[0074] Figure 4 This is a specific workflow diagram of the target identification and positioning module of the present invention.

[0075] Figure 5 This is the improved YOLOv8 network structure diagram of the present invention.

[0076] Figure 6 A schematic diagram of an interference scenario simulation provided by an embodiment of the present invention.

[0077] Figure 7 This is a distance-angle heat map provided by an embodiment of the present invention.

[0078] Figure 8This is a diagram of multi-target positioning results provided by an embodiment of the present invention.

[0079] Figure 9 This is a comparative relationship diagram of RMSE and SNR under different interference numbers provided by an embodiment of the present invention.

[0080] Figure 10 This is an evaluation indicator provided by the embodiment of the present invention. DETAILED DESCRIPTION

[0081] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0082] It should be noted that the terms "including" and "having" and any variations thereof in the specification and claims of the present invention and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products or apparatuses.

[0083] like Figure 1 As shown, the present invention provides an anti-interference and multi-target positioning method for a vehicle-mounted millimeter-wave radar, comprising:

[0084] S1. Construct a third-order cumulant interference suppression module. The third-order cumulant interference suppression module uses a third-order cumulant algorithm to process the signals received by the vehicle-mounted millimeter-wave radar, suppress mutual interference between radars, and generate a high-resolution distance-angle heat map.

[0085] S2. Construct a dual-channel adaptive edge detection module, which divides the generated heat map into an original channel and an edge detection channel, dynamically assigns weights based on local gradient strength, and fuses and outputs an enhanced heat map;

[0086] S3. Build a target recognition and positioning module. The target recognition and positioning module detects and locates the target on the enhanced thermal map through the improved YOLOv8 neural network, and outputs the distance and angle information of the target.

[0087] In specific implementation, as a preferred implementation manner of the present invention, when strong interference signals in the echo signals received by vehicle-mounted radars are aliased with the reflection signals of adjacent targets, mutual interference between radars will occur. Excessive interference will lead to false alarms or missed detections of targets and trigger traffic accidents. In this embodiment, the constructed third-order cumulant interference suppression module suppresses the mutual interference between vehicle-mounted radars based on the third-order cumulant. By using the synchronously collected array signals and according to the non-correlation characteristics of the echo signals and interference signals, a third-order cross-cumulant matrix is constructed, and then a two-dimensional spatial spectrum is obtained according to the subspace method to finally achieve interference suppression. The implementation process is as Figure 2 shown, that is, step S1, which specifically includes:

[0088] S11. Use the synchronously collected array signals and according to the non-correlation characteristics of the echo signals and interference signals, construct a third-order cross-cumulant matrix; signal modeling and reference signal generation are the basis of interference suppression. To model the signals in the interference scenario, step S11 specifically includes:

[0089] S111. Considering the general radar transmit signal and echo signal models, represent the radar transmit signal s T (t) as:

[0090]

[0091] where A t is the amplitude of the transmit signal, f0 is the carrier frequency of the signal, μ = B / T, B is the signal bandwidth, and T is the modulation period;

[0092] S112. Represent the echo signal s R (t) as:

[0093]

[0094] where A r is the amplitude of the echo signal, and τ is the time delay;

[0095] S113. Assume that there are K far-field target echo signals from different directions θ k (k = 0, 1,..., K - 1) and K I interference signals, and K I < K. Represent the difference-frequency signal received by the m-th array element in the linear array as:

[0096]

[0097] where m = 0, 1,..., M - 1, a m is the echo signal steering vector, s k (t) represents the k-th echo difference-frequency signal, b m is the interference signal steering vector, ωi (t) represents the difference frequency signal of the i-th interference, u m (t) represents superimposed Gaussian white noise;

[0098] S114. In order to accurately separate the target signal, a reference signal matrix matching the radar emission waveform characteristics needs to be constructed. Based on the frequency modulated continuous wave characteristics of the radar emission, the reference signal matrix is ​​generated as follows:

[0099]

[0100] Where n=0,1,…,N-1 represents the time series index, N represents the total number of sampling points, and f S Indicates the sampling rate;

[0101] S115. Introduce the time delay variable τ and construct the delay signal matrix as follows:

[0102]

[0103] S116: Perform discrete sampling on the formula in step S113 to obtain x m (n), and with s r (n+τ) and s r (n) together constitute the third-order cumulant expression. According to the properties of the third-order cumulant and the matrix model, the third-order cross-cumulant matrix is ​​constructed through the Hadamard product as follows:

[0104]

[0105] Where X is the received signal matrix, S τ is the delayed reference signal matrix, S * is the conjugate matrix of the reference signal, and N is the number of sampling points. In this embodiment, the third-order cumulant can effectively separate the nonlinear interference components in the target signal by capturing the high-order statistical characteristics of the signal, laying the foundation for subsequent interference suppression and joint parameter estimation.

[0106] S12, using the subspace method to obtain a two-dimensional spatial spectrum, achieve joint estimation of distance and orientation, and generate a high-resolution distance-angle heat map. Step S12 specifically includes:

[0107] S121. Perform eigenvalue decomposition on the third-order cross-cumulant matrix to extract the noise subspace, thereby suppressing interference components unrelated to the reference signal;

[0108] S122, combined with the target orientation vector a θ , take C by row τ The average value is c τ , construct a two-dimensional spatial spectrum as follows:

[0109] C(r,θ)=(c τ * ·a θ )(a θ * c τ )

[0110] Where r is the target distance, θ is the azimuth angle, and the symbol * represents conjugate transpose;

[0111] S123. Finally, a high-resolution distance-angle heat map is generated, where each peak corresponds to the distance and angle of a target.

[0112] In this embodiment, the third-order cumulant interference suppression module achieves efficient suppression of nonlinear interference through high-order statistical modeling and joint parameter estimation, while maintaining target detection accuracy in low signal-to-noise ratio environments. This module, through high-order statistical modeling, interference suppression, and joint parameter estimation, enables precise extraction of millimeter-wave radar signals in complex scenarios, laying the core foundation for target detection and positioning throughout the system.

[0113] During specific implementation, as a preferred embodiment of the present invention, the distance angle heat map generated will enter the dual-channel adaptive edge detection module. By fusing the original heat map with the edge detection channel, not only the edge features of small targets are significantly enhanced, but also the dynamic weight allocation and non-local noise suppression technology are used to effectively solve the problem of blurred edges and difficult boundary separation of overlapping targets in dense scenes, and reduce edge weights in flat areas to reduce false alarms and misdetection of overlapping targets. The implementation process can be divided into two key stages: dual-channel division and preprocessing, and dynamic weight fusion. The implementation process is as follows: Figure 3 As shown, step S2 specifically includes:

[0114] S21. Gaussian filtering, gamma correction, and normalization are performed on the original heat map channel to retain the global energy distribution of the target. In this embodiment, high-frequency noise is suppressed by Gaussian filtering to avoid noise amplification in the subsequent enhancement process and to ensure a balance between noise suppression and edge preservation. Gamma correction is then applied to enhance weak signals in low-contrast areas, and the data is normalized to the [0, 1] range through normalization to retain the global energy distribution of the target.

[0115] S22, using an improved Canny edge detection algorithm for the edge detection channel, calculating the horizontal and vertical gradients of the heat map, dynamically setting high and low thresholds, retaining strong edges and weak edges connected to strong edges, and suppressing isolated noise points through non-local mean filtering; in step S22:

[0116] First, use the Sobel operator to calculate the horizontal gradient Gx and vertical gradient Gy of the heat map. The calculation formula of the gradient amplitude is as follows:

[0117]

[0118] Where Gx is the horizontal gradient of the heat map, and Gy is the vertical gradient of the heat map;

[0119] Then the high and low thresholds set dynamically are T high =0.3×G max , T low =0.1×G max , retaining strong edges and weak edges connected to strong edges.

[0120] Finally, isolated noise points are suppressed by non-local means filtering (NLM).

[0121] S23. Dynamically assign weights to the two channels based on the local gradient strength. Strong edge regions give the edge channel a higher weight, while transition regions give the two channels equal weights. In flat regions, the edge channel weight is reduced. Finally, the fusion output is an enhanced heat map with both high-contrast edges and low-noise floor, providing significantly optimized input data for subsequent target recognition and positioning modules. The fusion formula is as follows:

[0122] I f =w o I r +w e ·E f

[0123] Among them, I r is the original heat map channel, which is gamma-corrected and normalized to preserve the global energy distribution of the target; E f For edge detection channel, refined edge information is extracted by improving Canny algorithm. In this embodiment, dynamic weight fusion further optimizes the balance between target edge and background. In this stage, the weights of the two channels are dynamically assigned according to the local gradient strength. n ≥0.3) gives 70% weight to the edge channel and 30% to the original channel (w e =0.7,w o =0.3), highlighting the target boundary by sharpening the outline; transition area (0.1≤G n <0.3), the two channels each account for 50% (w e =w o =0.5), taking into account both detail preservation and smooth transition; flat area (G n <0.1) reduce the edge channel weight to 20% and increase the original channel weight to 80%, (w o =0.8, w e =0.2), effectively filtering random noise.

[0124] In this embodiment, this step overcomes the rigidity of traditional fixed-weight fusion through a gradient-guided dynamic weight allocation mechanism. The dual-channel design balances global energy distribution with local edge details, avoiding the limitations of single-channel information and accurately adapting to diverse scenarios, such as small and overlapping targets. The dual-channel adaptive edge detection module, through layered optimization and collaborative design, significantly improves millimeter-wave radar's target detection capabilities in complex environments.

[0125] In specific implementation, as a preferred embodiment of the present invention, the target recognition and positioning module extracts target features based on the enhanced heat map output by dual-channel adaptive fusion, and introduces the attention mechanism and Focal Loss loss function optimization to the deep learning model to achieve improved detection rate of weak targets and separation accuracy of overlapping targets, and combines the spatial spectrum information of the heat map with the neural network output to accurately calculate the target distance and angular coordinates, such as Figure 4 As shown, step S3 specifically includes:

[0126] S31. Based on the simulation environment, a millimeter-wave radar data set is constructed to simulate a multi-target scenario on an ordinary urban road. Step S31 specifically includes:

[0127] S311. Using the simulation platform MATLAB, build a typical urban road model, including common elements such as multiple lanes and speed limits, and randomly generate multiple target vehicles in the scene;

[0128] S312. Based on the operating principle and parameters (transmit power, bandwidth, sampling rate, etc.) of a 77GHz automotive millimeter-wave radar, generate an original signal containing multiple target echoes and co-frequency interference in a simulation platform, and simultaneously inject varying degrees of noise.

[0129] S313. Automatically annotate each frame of radar echo data obtained by simulation in combination with the true value information (real distance, azimuth) in the simulation to generate a corresponding distance-angle heat map and target bounding box;

[0130] S314. Divide the generated data set into a training set and a test set, and perform data augmentation operations such as random rotation and Gaussian noise on some of the data to improve the robustness of the network to weak targets and low signal-to-noise ratio scenarios;

[0131] S315. In the later stage of simulation, abnormal samples (such as those that fail to be injected into the target correctly or are distorted due to excessive interference) are screened and eliminated to ensure the validity and accuracy of the data.

[0132] The dataset constructed through the above simulation method can well reproduce the multi-target and interference conditions in ordinary urban roads, laying the foundation for subsequent neural network training and evaluation.

[0133] S32, inputting the enhanced heat map into the improved YOLOv8 neural network, by embedding the channel attention mechanism and adopting the Focal Loss loss function, improving the detection rate of small targets and the accuracy of overlapping target separation; Step S32 specifically includes:

[0134] S321. The initial convolutional layers in the backbone of the YOLOv8 neural network are pruned, retaining only a single downsampling layer. This reduces computational complexity while retaining more high-frequency detail features. This design enables the complete transfer of texture information from the 320×320 high-resolution layer to subsequent networks, particularly improving the recognizability of edge features of small objects.

[0135] S322, Embed a Lightweight Channel Attention Mechanism (ECA-Net). A lightweight ECA module is inserted into the core path of each downsampling stage to form a downsampling-channel calibration-feature fusion cascade architecture. Channel weights are dynamically calibrated through global average pooling and 1D convolution to enhance the edge feature response of weak targets and suppress background noise channels.

[0136] S323. Reconstruct the loss function to optimize classification-regression collaborative training, replace the original classification loss function with the Focal Loss loss function, focus on difficult samples in low signal-to-noise ratio areas, and retain the joint optimization of YOLOv8's DFL Loss and CIoULoss as the bounding box regression loss to reduce the positioning error of dense targets. In the target detection task, the background area usually occupies the vast majority of the image. The traditional loss function is easily dominated by the incorrect prediction of the background area, resulting in the model being unable to accurately identify the target. In this embodiment, the Focal Loss loss function improves detection accuracy by assigning a smaller loss weight to samples that are easy to classify, focusing on targets that are difficult to classify.

[0137] In step S323, the Focal Loss loss function is based on the cross entropy loss function by introducing a modulation factor (1-p t ) γ , the formula is as follows:

[0138] FL=-α t (1-p t ) γ log(p t )

[0139] Among them, p t is the model’s predicted probability for the true category, α t is the sample balancing factor.

[0140] The improved YOLOv8 network structure is as follows Figure 5As shown in the figure, the core idea of ​​the attention mechanism is to help the network focus on the most critical parts of the image by assigning different weights to different regions. In this system, the attention mechanism guides the network to pay more attention to the target area and less attention to background or irrelevant areas, thereby suppressing background noise interference. This mechanism helps effectively identify multiple objects in complex environments, improving the model's sensitivity and accuracy in detecting objects.

[0141] S33. Combine the spatial spectrum information of the heat map with the output information of the neural network to accurately calculate the target distance and angle coordinates.

[0142] After completing the above steps, the multi-object detection phase begins. In this phase, the optimized network outputs the location, category, and bounding box information for all objects in the image. Specifically, the network generates a bounding box that precisely locates each object, along with its distance and angle. For each detected object, the system outputs a result containing information such as the object's center coordinates, width, height, and category label.

[0143] Overall, the target recognition and localization module achieves high-precision target detection and localization by enhancing target features through the coordinated optimization of the attention mechanism and the Focal Loss loss function. This module improves target detection accuracy, ensuring the system can quickly and accurately locate target information. It achieves a balance between precision and real-time performance in millimeter-wave radar target detection, providing reliable target perception capabilities for intelligent driving systems.

[0144] Example

[0145] In order to verify the feasibility of the method of the present invention, based on Figure 6 The method of the present invention was tested in the dynamic traffic scenario shown in Figure 1. Vehicle A, equipped with an array radar, detected surrounding targets. However, strong interference from vehicle D prevented the detection of vehicle B. In this test example, the onboard radar parameters were set to a center frequency of 77 GHz, a bandwidth of 520 MHz, and an 8-element linear array, as shown in Table 1.

[0146] Table 1 Parameter settings

[0147]

[0148] The simulation dataset construction strategy is as follows:

[0149] A city road model was built in the simulation platform MATLAB. Assuming that the radar is installed in front of the vehicle, the detection range is 0 to 50 meters and the azimuth angle is ±60 degrees. 640 points are sampled in each of the distance and angle directions to generate a 640×640 heat map. Among them, the radar cross section σ is divided into weak targets (σ = 0.5 to 5m 2) and conventional targets (σ=5~20m 2 Several static or moving targets can be randomly placed in the scene to simulate a common urban driving environment.

[0150] Approximately 50,000 frames of radar echo data were generated in the simulation, covering a variety of traffic densities and relative speed differences. In each simulation scenario, 1 to 6 targets were randomly generated. For some of these targets, a smaller radar cross-section (RCS) was set or they were placed at a long distance (e.g., 40 to 50 meters) to reflect weak target signals. By placing 0 to 2 targets close to each other in range and azimuth (angle difference of 3° to 5°, distance difference ≤ 5m), target overlap in dense scenarios was simulated. Regarding the number of interfering radars, 0 to 5 interfering radars were added in the same frequency band (77GHz) to simulate strong co-frequency interference between vehicle-mounted radars, with a signal-to-interference ratio (SIR) of -5dB to 5dB.

[0151] The distance between the jamming radar and the primary radar can be randomly distributed within a range of 0 to 50 meters, and different modulation parameters and relative speeds can be set to generate jamming signals of varying amplitudes. Subsequently, the echo signal of each target is modeled based on the principle of FMCW frequency-modulated continuous wave radar. The jamming radar signal is superimposed on the primary radar receiver, and the power, phase, and angle of arrival of the jamming signal are controlled to present strong interference or partially overlapping false peaks in the simulation. Random perturbations such as Gaussian white noise are added to the signal layer to enhance the fit to the real environment. The signal-to-noise ratio (SNR) is set to -10dB to 25dB.

[0152] In the simulation system, the true range (r) and azimuth (θ) of each target are directly obtained. A range-angle heatmap is generated based on the echo intensity, range, and angle information. Combined with the ground-truth labels (r, θ), each target in the heatmap frame is automatically annotated with a bounding box or center point coordinates. Interference sources are also recorded to facilitate subsequent evaluation of interference suppression effectiveness. Each heatmap frame corresponds to a text file, which records the label information for each target line by line.

[0153] The generated raw data can be divided into training and test sets in an 8:2 ratio, ensuring that the test set contains a variety of interference intensities and different numbers of targets. Data augmentation operations are performed on some training data, such as randomly rotating heat maps or adding additional noise, to improve the network's robustness to weak and overlapping targets. The proportion of "strong interference" scenarios is appropriately increased (for example, the number of interfering radars is ≥3) to ensure that the model can still learn interference suppression features when co-frequency interference is severe. Ultimately, the true label of each target is ensured to be accurate, and the annotation file clearly records information such as the number of targets, distance, angle, and RCS.

[0154] The simulation dataset constructed using the above strategy effectively reproduces the scenarios of strong co-frequency interference and multiple overlapping targets encountered by on-board radars in typical urban road environments within a range of 0 to 50 meters and an azimuth angle of ±60°. It particularly effectively simulates small and weak targets (those at long range or with low RCS). This dataset provides targeted samples for the subsequent deep learning network training and evaluation of the present invention, ensuring highly robust detection performance despite multiple interferences, multiple overlapping targets, and small and weak targets.

[0155] Subsequently, the reference signal matrix S is generated based on the radar transmission waveform characteristics, the modulation slope μ=B / T=520MHz / 11μs, and the delay variable τ is introduced to construct the delay reference matrix S τ , use MATLAB to calculate the third-order cross-cumulant matrix And C τ Perform singular value decomposition to extract the noise subspace, and combine it with the steering vector to generate a distance-angle heat map, such as Figure 7 The results show that the peak of the strong interference signal (vehicle D) is suppressed by more than 90%, and the weak target (vehicle B) is clearly visible. Subsequently, the heat map enters the dual-channel adaptive fusion stage: the original channel uses Gaussian filtering (3×3 kernel, σ=1.0) and gamma correction (γ=0.5) to enhance the target energy; the edge detection channel extracts the gradient through the Sobel operator and dynamically sets the high and low thresholds (T high =0.3×G max , T low =0.1×G max ), combined with non-local mean filtering (5×5 window, h=0.8) to suppress noise, and finally dynamically fused the two channels according to the local gradient strength (weights 0.7:0.3 in strong edge areas and 0.8:0.2 in flat areas), the gradient strength of the edges of weak targets was increased by 2.3 times.

[0156] The enhanced heatmap input improves the YOLOv8 network. Using the PyTorch framework, the network removes two downsampling layers to retain the 160×160 input resolution. An ECA-Net module (1D convolution kernel k=3) is embedded to dynamically calibrate channel weights. Multi-scale features (160×160, 80×80, and 40×40) are fused. A focal loss (α=0.25, γ=2.0) and a combined loss (DFL+CIoU) are used to optimize classification and regression. During training, both synthetic and real-world scene data are loaded, augmented with random rotations (±15°) and Gaussian noise (σ=0.01). SGD optimizer (learning rate 0.01, momentum 0.937) is used with a cosine annealing strategy for 300 epochs.

[0157] This test covers the entire process from signal processing to network inference, fully verifying the effectiveness of the method of this invention. The third-order cumulant algorithm performs excellently in suppressing dynamic interference, successfully suppressing more than 90% of false peaks. The dual-channel fusion strategy increases the edge gradient strength of weak targets by 2.3 times through dynamic weight allocation, improves the accuracy of overlapping target separation by 45%, and combines the improved Canny algorithm with non-local mean filtering (NLM) to reduce the false alarm rate to 0 when SNR = 5dB.

[0158] In terms of target detection and positioning, the target positioning results are shown in the figure below. Figure 8 As shown in the figure, the coordinates represent the target's distance angle and the error (in meters) from the actual result. The actual angle and distance are [0, 48, 26, -3, -30, -38] degrees and [32, 12, 42, 35, 14, 25] meters, respectively. After embedding ECA-Net channel attention and focal loss, the localization error of dense targets is reduced by 37%, and dense targets can be clearly distinguished. The azimuth angle estimation accuracy reaches 3°.

[0159] Figure 9 The algorithm's performance was evaluated under different numbers of interference sources. This example investigates how different interference conditions and SNR levels affect the positioning accuracy measured by RMSE. Different markers and line styles represent interference conditions, ranging from 0 to 5 interference sources. The figure clearly shows that when there are no interference sources, the RMSE value is the lowest, indicating the best performance and highest positioning accuracy. The positioning accuracy shows a controllable and gradual change, and the RMSE curves under various interference conditions maintain stable convergence, proving that the algorithm can still effectively maintain positioning accuracy in multi-interference coupling scenarios.

[0160] like Figure 10 As shown, the bounding box prediction losses (box_loss and dfl_loss) both steadily decreased and eventually converged to below 0.1. Precision, recall, and mAP metrics all steadily increased from 0, ultimately approaching or reaching 0.95, demonstrating the model's excellent object detection capabilities. Combined with the collaborative design of the MATLAB, Python, and PyTorch toolchain, this model provides a complete and efficient technical solution for millimeter-wave radar, from interference suppression to high-precision positioning in complex traffic scenarios. It breaks through the bottlenecks of small target detection and dense target separation, and provides a new perception paradigm for Level 4 and higher autonomous driving systems that combines high precision, low latency, and strong anti-interference capabilities.

[0161] In summary, this invention combines high-order statistical analysis of third-order cumulants with deep neural network technology, effectively breaking through the key bottlenecks faced by vehicle-mounted millimeter-wave radars in complex traffic environments: interference suppression, weak target recognition, and overlapping target separation. By introducing a dynamic weight allocation mechanism, an edge feature enhancement strategy, and a channel attention module, it not only achieves high-resolution detection, but also significantly improves the accuracy and robustness of target positioning. Experimental results fully verify the excellent performance of the method of the present invention in reducing the risk of false detection and missed detection, providing a solid technical foundation for intelligent driving and autonomous driving systems.

[0162] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A vehicle-mounted millimeter-wave radar anti-interference and multi-target positioning method, characterized in that: include: S1. Construct a third-order cumulant interference suppression module. The third-order cumulant interference suppression module uses a third-order cumulant algorithm to process the signals received by the vehicle-mounted millimeter-wave radar, suppress mutual interference between radars, and generate a high-resolution distance-angle heat map. S2. Construct a dual-channel adaptive edge detection module, which divides the generated heat map into an original channel and an edge detection channel, dynamically assigns weights based on local gradient strength, and fuses and outputs an enhanced heat map; S3. Build a target recognition and positioning module. The target recognition and positioning module detects and locates the target on the enhanced thermal map through the improved YOLOv8 neural network, and outputs the distance and angle information of the target.

2. The anti-interference and multi-target positioning method of a vehicle-mounted millimeter-wave radar according to claim 1, characterized in that: Step S1 specifically includes: S11, using the synchronously collected array signals and according to the non-correlated characteristics of the echo signal and the interference signal, constructing a third-order cross-cumulant matrix; S12. Obtain a two-dimensional spatial spectrum based on the subspace method, realize joint estimation of distance and orientation, and generate a high-resolution distance-angle heat map.

3. The anti-interference and multi-target positioning method of a vehicle-mounted millimeter-wave radar according to claim 2, characterized in that: Step S11 specifically includes: S111, the radar transmits a signal s T (t) is expressed as: Among them, A t is the amplitude of the transmitted signal, f0 is the carrier frequency of the signal, μ=B / T, B is the signal bandwidth, and T is the modulation period; S112, the echo signal s R (t) is expressed as: Among them, A r is the echo signal amplitude, τ is the time delay; S113. Assume that there are K far-field target echo signals from different directions θ k (k = 0, 1, …, K - 1) and K I interference signals, and K I < K. The difference frequency signal received by the m-th array element in the linear array is expressed as: Where m = 0, 1, ..., M-1, a m is the echo signal steering vector, s k (t) represents the kth echo difference frequency signal, b m is the interference signal steering vector, ω i (t) represents the difference frequency signal of the i-th interference, u m (t) represents superimposed Gaussian white noise; S114. Based on the frequency modulated continuous wave characteristics of the radar transmission, a reference signal matrix is ​​generated as follows: Where n=0,1,…,N-1 represents the time series index, N represents the total number of sampling points, and f S Indicates the sampling rate; S115. Introduce the time delay variable τ and construct the delay signal matrix as follows: S116: Perform discrete sampling on the formula in step S113 to obtain x m (n), and with s r (n+τ) and s r (n) together constitute the third-order cumulant expression. According to the properties of the third-order cumulant and the matrix model, the third-order cross-cumulant matrix is ​​constructed through the Hadamard product as follows: Where X is the received signal matrix, S τ is the delayed reference signal matrix, S * is the conjugate matrix of the reference signal, and N is the number of sampling points.

4. The anti-interference and multi-target positioning method of a vehicle-mounted millimeter-wave radar according to claim 2, characterized in that: Step S12 specifically includes: S121. Perform eigenvalue decomposition on the third-order cross-cumulant matrix to extract the noise subspace, thereby suppressing interference components unrelated to the reference signal; S122, combined with the target orientation vector a θ , take C by row τ The average value is c τ , construct a two-dimensional spatial spectrum as follows: C(r,θ)=(c τ * ·a θ )(a θ * ·c τ ) Where r is the target distance, θ is the azimuth angle, and the symbol * represents conjugate transpose; S123. Finally, a high-resolution distance-angle heat map is generated, where each peak corresponds to the distance and angle of a target.

5. The anti-interference and multi-target positioning method of a vehicle-mounted millimeter-wave radar according to claim 1, characterized in that: Step S2 specifically includes: S21, performing Gaussian filtering, gamma correction and normalization on the original thermal map channel to retain the global energy distribution of the target; S22. An improved Canny edge detection algorithm is used for the edge detection channel to calculate the horizontal and vertical gradients of the heat map, dynamically set high and low thresholds, retain strong edges and weak edges connected to strong edges, and suppress isolated noise points through non-local mean filtering; S23. Dynamically assign weights to the two channels based on the local gradient strength. Strong edge regions give the edge channel a higher weight, while transition regions give the two channels equal weights. Flat regions reduce the edge channel weight. Finally, the fusion output is an enhanced heat map with both high-contrast edges and low-noise floor. The fusion formula is as follows: I f =w o I r +w e ·HAVE BEEN f Among them, I r is the original heat map channel, which is gamma-corrected and normalized to preserve the global energy distribution of the target; E f It is the edge detection channel, which extracts refined edge information through the improved Canny algorithm.

6. The anti-interference and multi-target positioning method of a vehicle-mounted millimeter-wave radar according to claim 5, characterized in that: In step S22: The Sobel operator is used to calculate the horizontal gradient Gx and vertical gradient Gy of the heat map. The calculation formula for the gradient amplitude is as follows: Where Gx is the horizontal gradient of the heat map, and Gy is the vertical gradient of the heat map; The high and low thresholds set dynamically are T high =0.3×G max , T low =0.1×G max .

7. The anti-interference and multi-target positioning method of a vehicle-mounted millimeter-wave radar according to claim 1, characterized in that: Step S3 specifically includes: S31. Based on the simulation environment, a millimeter-wave radar dataset is constructed to simulate multi-target scenarios on ordinary urban roads. S32. Input the enhanced heat map into the improved YOLOv8 neural network, and improve the detection rate of small objects and the accuracy of overlapping object separation by embedding the channel attention mechanism and adopting the Focal Loss loss function; S33. Combine the spatial spectrum information of the heat map with the output information of the neural network to accurately calculate the target distance and angle coordinates.

8. The anti-interference and multi-target positioning method of a vehicle-mounted millimeter-wave radar according to claim 7, characterized in that: Step S31 specifically includes: S311. Using the simulation platform MATLAB, build a typical urban road model, including common elements such as multiple lanes and speed limits, and randomly generate multiple target vehicles in the scene; S312. Based on the operating principle and parameters of the 77GHz automotive millimeter-wave radar, generate an original signal containing multiple target echoes and co-frequency interference in the simulation platform, and inject different levels of noise at the same time; S313. Automatically annotate each frame of radar echo data obtained by simulation in combination with true value information in the simulation to generate a corresponding distance-angle heat map and target bounding box; S314. Divide the generated data set into a training set and a test set, and perform data augmentation operations such as random rotation and Gaussian noise on some of the data to improve the robustness of the network to weak targets and low signal-to-noise ratio scenarios; S315. In the later stage of simulation, abnormal samples are screened and eliminated to ensure the validity and accuracy of the data.

9. The anti-interference and multi-target positioning method of a vehicle-mounted millimeter-wave radar according to claim 7, characterized in that: Step S32 specifically includes: S321. Delete the initial convolutional layer in the backbone of the YOLOv8 neural network and retain only the single downsampling layer, which reduces the amount of computation while retaining more high-frequency detail features. S322, Embed a Lightweight Channel Attention Mechanism (ECA-Net). A lightweight ECA module is inserted into the core path of each downsampling stage to form a downsampling-channel calibration-feature fusion cascade architecture. Channel weights are dynamically calibrated through global average pooling and 1D convolution to enhance the edge feature response of weak targets and suppress background noise channels. S323. Reconstruct the loss function to optimize classification-regression collaborative training, replace the original classification loss function with the FocalLoss loss function, focus on difficult samples in low signal-to-noise ratio areas, and retain the joint optimization of YOLOv8's DFL Loss and CIoULoss as the bounding box regression loss to reduce the positioning error of dense targets.

10. The anti-interference and multi-target positioning method of a vehicle-mounted millimeter-wave radar according to claim 7, characterized in that: In step S323, the FocalLoss loss function introduces a modulation factor (1-p t ) γ , the formula is as follows: FL=-a t (1-p t ) γ log(p t ) Among them, p t is the model’s predicted probability for the true category, α t is the sample balancing factor.

Images