An end-to-end based air traffic control primary radar target detection method

Abstract

The application discloses an air traffic control primary radar target detection method based on end-to-end, relates to the field of air traffic control primary radar, and solves the problem that the prior art relies on artificial experience and an existing deep learning model is difficult to effectively utilize radar space-time information. The method comprises the following steps: acquiring and pre-processing real-time radar data to obtain a real-time matrix, combining the real-time matrix with historical matrices of two continuous time points into a space-time three-dimensional data block; inputting the data block into a trained target segmentation model to obtain a binary segmentation graph; extracting a multi-dimensional feature vector of each connected region through connected domain analysis; inputting the multi-dimensional feature vector into a trained target classification model to obtain a probability value of each set belonging to a real target; and finally outputting a target track list through threshold screening, non-maximum suppression and coordinate mapping. The application realizes end-to-end automation of the detection process, can adapt to complex environments, significantly improves detection performance and reduces dependence on artificial experience.

Description

Technical Field

[0001] This invention belongs to the field of primary air traffic control radar and relates to radar target detection technology, specifically an end-to-end primary air traffic control radar target detection method. Background Technology

[0002] Air traffic control primary radar is one of the main sensors in the air traffic control system. In the target detection process of primary air traffic control radar, constant false alarm rate (CFAR) processing is a crucial step, the core of which lies in setting the detection threshold. Setting the threshold too high will cause weak target signals to be overwhelmed by noise and clutter, resulting in missed detections; setting the threshold too low will cause a large amount of noise fluctuations and residual clutter to be misjudged as targets, generating intolerable false alarms and severely interfering with the stability of subsequent track establishment and tracking.

[0003] Existing technologies for threshold adjustment mainly rely on two methods: one is manual setting by operators based on experience, which is cumbersome, lacks real-time performance, and requires high technical skills; the other is automatic CFAR algorithms based on local statistics, but these methods depend on the ideal assumption of a uniform background and sparse targets. In actual complex air traffic control surveillance scenarios, factors such as geographical environment, meteorological clutter, and multi-target interference often lead to non-uniform backgrounds, causing a decline in the performance of traditional CFAR and making it difficult to simultaneously address false alarms and missed detections.

[0004] Although deep learning has provided new ideas for radar detection, directly applying general image detection models to radar data often fails to fully extract the spatiotemporal coherence information unique to radar echoes. Furthermore, its generalization ability in complex clutter environments, the efficiency of utilizing radar-specific prior knowledge, and the interpretability of the decision-making process remain insufficient. Summary of the Invention

[0005] This application provides an end-to-end air traffic control primary radar target detection method, which solves the technical problems of existing technologies such as reliance on human experience, insufficient utilization of spatiotemporal coherent information and radar-specific prior knowledge, and insufficient interpretability of feature extraction and decision-making processes.

[0006] To achieve the above objectives, this application adopts the following technical solution: In the first aspect, an end-to-end air traffic control single radar target detection method is provided, including: preprocessing the real-time acquired radar data to obtain a real-time radar matrix, and combining it with the historical radar matrix corresponding to the two consecutive time units to obtain a spatiotemporal three-dimensional data block. The spatiotemporal 3D data block is input into the target segmentation model to obtain a binary segmentation map; the spatiotemporal 3D data block and the binary segmentation map are input into the target classification model to obtain several connected regions and extract multidimensional feature vectors; the multidimensional feature vectors are processed by a classification network to obtain the probability value of each connected region being the target region; the target segmentation model and the target classification model are trained by a 3D convolutional neural network and a multidimensional feature encoding neural network, respectively. Connected regions with probability values ​​higher than a preset probability threshold are selected and marked as a preliminary target set; non-maximum suppression is applied to the preliminary target set to obtain the final target set; the center pixel position of the final target set is mapped to distance-azimuth coordinates to obtain a list of target point information.

[0007] Based on the above technical solutions, this application provides an end-to-end air traffic control primary radar target detection method. By constructing an end-to-end deep learning detection framework, the entire process of radar target detection, from data preprocessing to point generation, is automated, eliminating the deep reliance on manual parameter tuning experience. This method combines continuous temporal data to form a three-dimensional input and utilizes a three-dimensional convolutional network to deeply mine the spatiotemporal coherence of the target, effectively enhancing the detectability of weak targets in complex clutter backgrounds. Embedding the inherent physical prior knowledge of the radar into the network in an adaptive weighted form significantly improves the model's ability to suppress non-uniform clutter and its scene adaptability. By designing an interpretable decision path of segmentation-feature extraction-classification and extracting multi-dimensional features for comprehensive discrimination, not only is the reliability of false alarm filtering improved, but the transparency and analyzability of the system output results are also enhanced. This provides a complete solution for stable and reliable automated detection of primary air traffic control radar in complex surveillance environments such as strong clutter and multiple targets.

[0008] In conjunction with the first aspect above, in one possible implementation, the method for acquiring the spatiotemporal three-dimensional data block includes: From the real-time radar matrix and the two historical radar matrices, M identical range cells are continuously extracted along the range dimension, and N identical azimuth channels are continuously extracted along the azimuth dimension to obtain complex submatrices; Two-dimensional Fourier transforms are performed on the complex submatrices respectively to extract the low-frequency region of the corresponding spectrum and obtain the corresponding principal phase component by calculating the statistical average of the phase angles; The global phase rotation of the complex submatrix of the historical radar matrix is ​​obtained by calculating the difference between the principal phase component of the complex submatrix corresponding to the historical radar matrix and the principal phase component of the complex submatrix corresponding to the real-time radar matrix. Based on the global phase rotation, the complex submatrix of the historical radar matrix is ​​subjected to reverse phase compensation, and stacked with the complex submatrix corresponding to the real-time radar matrix in chronological order to obtain a spatiotemporal three-dimensional data block with a size of 3×M×N.

[0009] By adaptive phase compensation of historical data, the system phase difference and random phase noise between radar echoes at different times are effectively eliminated, ensuring the coherence of the spatiotemporal three-dimensional data blocks in the time dimension. This enables the subsequent neural network to extract the continuous features of the target in the time domain more accurately, enhancing the ability to identify real moving targets in environments with strong clutter or low signal-to-noise ratio.

[0010] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the binary segmentation map includes: The spatiotemporal three-dimensional data block is subjected to three-dimensional convolution and nonlinear activation processing to obtain the first feature map; The first feature map is subjected to cross-dimensional adaptive weighting to obtain the second feature map; The second feature map is subjected to spatial downsampling, three-dimensional convolution, and non-linear activation to obtain the third feature map; The third feature map is upsampled and then connected to the first feature map along the channel dimension to obtain a fused feature map. The fused feature map is subjected to 3D convolution and nonlinear activation processing to obtain a single-channel feature map; The single-channel feature map is processed by the Sigmoid activation function to obtain a probability map; each pixel value in the probability map represents the probability that a target exists at that location; Pixels with a probability value greater than a preset segmentation threshold are marked as target pixels, and non-target pixels are marked as background pixels, thus obtaining a binary segmentation map.

[0011] By employing a cross-dimensional adaptive weighting mechanism, prior knowledge of the radar system is transformed into feature enhancement weights, significantly strengthening the feature response of the real target region. Combined with a multi-scale feature fusion structure, detailed information of small targets is fully preserved while compressing noise, thus providing a high-precision initial target region for subsequent processing.

[0012] In conjunction with the first aspect mentioned above, in one possible implementation, the method for obtaining the second feature map includes: Calculate the signal attenuation curve based on the inherent system parameters of the radar, generate the compensation coefficient corresponding to each range cell of the first feature map, and mark it as a range dimension weight vector; The radar antenna pattern data is obtained based on the inherent antenna design parameters of the radar, and the gain ratio of each directional channel relative to the beam center is calculated to obtain the azimuth weight vector. Based on spatiotemporal three-dimensional data blocks, the time series statistical variance of each distance orientation unit is calculated, and nonlinear mapping processing is performed to obtain normalized compensation coefficients, which are marked as time dimension weight vectors. Based on the inherent hardware parameters of the radar receiving channel, the calibration gain value of each channel is obtained and the gain normalization coefficient of each receiving channel is calculated to generate a channel-dimensional weight vector. The distance dimension weight vector, orientation dimension weight vector, time dimension weight vector, and channel dimension weight vector are fused and then multiplied element-wise on the first feature map to obtain the second feature map.

[0013] By explicitly encoding the inherent physical constraints and statistical characteristics of the radar system into multidimensional weights, the feature map is adaptively calibrated. This effectively suppresses background interference introduced by the radar's physical characteristics and hardware non-ideals, enhances the response of the target area, and improves the model's feature discrimination ability and robustness in real-world complex scenarios.

[0014] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the probability values ​​of each connected region being the target region includes: The binary segmentation map and the spatiotemporal 3D data block are input into the target classification model; Through connected component analysis, several connected regions are obtained, and multidimensional feature vectors are extracted; the multidimensional feature vectors include: geometric features, statistical features, motion features, and motion compliance features; The multidimensional feature vector is split into four feature sub-vectors and input into four independent fully connected coding sub-networks for linear mapping to obtain four corresponding deep feature representations. The four deep feature representations are concatenated and input into the cross-attention layer. The attention weights between the four deep feature representations are calculated to obtain a four-dimensional weight vector. The four feature sub-vectors are then weighted and fused based on the weight vector to obtain the aggregated feature representation; The aggregated feature representation is mapped to a classification score through a fully connected layer. The classification score is then nonlinearly compressed and normalized using the Sigmoid function to obtain the probability value that the connected region belongs to the target region.

[0015] By splitting feature sub-vectors and using a cross-attention fusion mechanism, refined modeling of multi-dimensional target features is achieved. The cross-attention layer can dynamically evaluate and weight the contributions of different feature subclasses, enabling the model to adaptively strengthen key discriminative features and suppress interfering features in complex scenarios, thereby significantly improving the accuracy of real and false target classification and the robustness of decision-making.

[0016] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the multidimensional feature vector includes: Based on the pixel position of each connected region in the binary segmentation map, the number of pixels in the connected region and the aspect ratio of the bounding rectangle are calculated and marked as geometric features; Extract the values ​​of all matrix elements within the corresponding spatial region from the spatiotemporal three-dimensional data block, and calculate the mean and variance of the matrix elements as statistical features; Matrix elements of the corresponding spatial regions are extracted from three consecutive time slices of the spatiotemporal three-dimensional data block. The displacement is obtained by calculating the centroid coordinates of each slice and the difference between the centroid coordinates of adjacent time slices, and is marked as motion features. The radial acceleration of the target is calculated using second-order difference based on the displacement and compared with a preset target acceleration threshold to obtain a compliance flag bit marked as a motion compliance feature; The feature values ​​contained in the geometric features, statistical features, motion features, and motion compliance features are concatenated to obtain the multidimensional feature vector corresponding to the connected region.

[0017] By extracting multi-dimensional physical features of the target, a rich set of discrimination criteria was constructed. In particular, the introduction of motion compliance judgment can effectively identify and eliminate false targets caused by clutter jitter or noise and whose motion patterns violate common sense, thereby significantly enhancing the false alarm suppression capability at the feature level and improving the credibility of the overall detection results.

[0018] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the target point information list includes: Calculate the mean row index and mean column index of all pixels in each connected region of the final target set in the binary segmentation image; and convert them into the target range value and azimuth value respectively according to the inherent spatial calibration parameters of the radar system. The distance, azimuth, probability, and radar data timestamp of each target are encapsulated into structured data records, and the results are aggregated to produce a list of target point information.

[0019] By combining pixel-level center positioning with system calibration parameters, sub-pixel-level coordinate transformation accuracy is achieved, and the output structured point data can be directly used for track association, improving system integration efficiency.

[0020] In conjunction with the first aspect above, in one possible implementation, the preprocessing of the real-time acquired radar data to obtain the real-time radar matrix includes: The real-time acquired radar echo signals are pulse compressed to obtain a one-dimensional range image sequence; The one-dimensional range image sequence is windowed and coherently accumulated, and then subjected to fast Fourier transform along the azimuth dimension to obtain a two-dimensional range-Doppler spectrum. The Doppler spectrum of each distance unit is divided into several sub-bands according to a preset Doppler frequency interval. Based on the statistical distribution of the spectral amplitude in each sub-band, the background level is obtained through ordered statistical processing. Spectral units with amplitudes higher than the background level by a preset multiple are marked as candidate significant units. Non-continuous candidate significant units in the distance dimension are removed to obtain significant spectral energy. The amplitude of the significant spectral energy is detected and logarithmically compressed to obtain an amplitude data matrix, which is labeled as the real-time radar matrix.

[0021] Through coherent accumulation and subband ordered statistical processing, complex clutter is effectively suppressed and the signal-to-noise ratio is improved. The generated real-time radar matrix highlights potential targets while reducing background interference, providing cleaner and more salient input data for subsequent deep learning models.

[0022] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the target segmentation model and the target classification model includes: Construct a training dataset; the training dataset includes: spatiotemporal three-dimensional data blocks corresponding to simulated and real radar data and ground truth labels for target locations; Using the spatiotemporal three-dimensional data block as input and the target location ground truth label as supervision, a predicted segmentation map is obtained through forward propagation calculation; The spatial adaptive weighted focus loss and topological consistency loss are calculated based on the predicted segmentation map and then summed in weight to obtain the final error value. Based on the final error value, the parameters are optimized through the backpropagation algorithm until the model converges. The training dataset is processed by the target segmentation model to generate a binary segmentation map and perform connected component analysis to obtain the multidimensional feature vector of each connected region. Calculate the overlap between each connected region and the spatial range defined by the corresponding target location ground value, and label connected regions that exceed the preset labeling threshold as positive samples, and label the rest as negative samples; Using the multidimensional feature vector as input, and the labeled positive and negative samples as supervision labels, the class prediction value is calculated through forward propagation; the prediction error is calculated through the cross-entropy loss function; and the network parameters are iteratively optimized through the backpropagation algorithm until the model converges, thus obtaining the target classification model.

[0023] By training with a mix of simulation and real data, combined with a targeted loss function and a rigorous positive and negative sample labeling strategy, the model's generalization ability under complex clutter, sensitivity to small targets, and combined accuracy of segmentation and classification are effectively improved.

[0024] In conjunction with the first aspect above, in one possible implementation, the method for obtaining the training dataset includes: Simulated radar data is generated by a radar data simulator and the target position true value label is output synchronously. The real data is received by the radar antenna, and the target position detection result is obtained by ordered statistical constant false alarm detection, and marked as the target position true value label; Simulated radar data and real radar data are processed into spatiotemporal three-dimensional data blocks, and paired with corresponding ground truth labels to form the training dataset.

[0025] By combining the precise ground truth of simulation data with the actual noise characteristics of real data, the problem of obtaining and labeling the ground truth of training samples in real complex scenarios is effectively solved, providing a high-quality and high-coverage data foundation for model training.

[0026] This application provides an end-to-end air traffic control radar target detection method that automates the entire detection process, significantly reducing reliance on human experience. By mining the target's coherence in the continuous time domain, it effectively improves the detection probability of weak targets in low signal-to-noise ratio environments. Simultaneously, embedding radar physical characteristics as prior knowledge into the network significantly enhances the model's ability to suppress non-uniform clutter and its scene adaptability. Furthermore, its interpretable feature engineering and decision-making process improves the reliability of the results and the maintainability of the system, providing a reliable technical guarantee for the stable operation of air traffic control radar in complex environments.

[0027] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0028] Figure 1 A system architecture diagram of an end-to-end air traffic control primary radar target detection method provided in this application embodiment; Figure 2 A flowchart illustrating a preprocessing method provided in an embodiment of this application; Figure 3 A schematic diagram illustrating the process of obtaining a binary segmentation map provided in an embodiment of this application; Figure 4 A schematic diagram illustrating the process of obtaining multidimensional feature vectors provided in an embodiment of this application; Figure 5 This is a flowchart illustrating the process of obtaining the probability value of a connected region being a target region, as provided in an embodiment of this application. Figure 6 A schematic diagram illustrating the process of obtaining a list of target point information provided in an embodiment of this application; Figure 7 A schematic flowchart illustrating a model training process provided in an embodiment of this application; Detailed Implementation

[0029] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0030] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0031] To address the technical problems in existing technologies, such as reliance on human experience, insufficient utilization of spatiotemporal coherence information and radar-specific prior knowledge, and insufficient interpretability of feature extraction and decision-making processes, this application provides an end-to-end air traffic control single-target detection method. This method includes: preprocessing real-time acquired radar data to obtain a real-time radar matrix, and combining it with historical radar matrices corresponding to the two consecutive time units to obtain a spatiotemporal three-dimensional data block; inputting the spatiotemporal three-dimensional data block into a target segmentation model to obtain a binary segmentation map; inputting the spatiotemporal three-dimensional data block and the binary segmentation map into a target classification model to obtain several connected regions and extracting multi-dimensional feature vectors; processing the multi-dimensional feature vectors using a classification network to obtain the probability value of each connected region being a target region; the target segmentation model and the target classification model are trained by a three-dimensional convolutional neural network and a multi-dimensional feature encoding neural network, respectively; selecting connected regions with probability values ​​higher than a preset probability threshold and marking them as a preliminary target set; performing non-maximum suppression processing on the preliminary target set to obtain a final target set; and mapping the center pixel position of the final target set to range-azimuth coordinates to obtain a target point information list. Based on this, intelligent and automated air traffic control radar target detection has been achieved, effectively improving the detection capability and false alarm suppression effect of weak targets in complex environments, and ensuring the credibility of the results and the maintainability of the system through an interpretable decision-making mechanism.

[0032] like Figure 1 As shown in the embodiment of this application, an end-to-end air traffic control primary radar target detection method includes: S101. The real-time radar data is preprocessed to obtain a real-time radar matrix, and then combined with the historical radar matrices corresponding to the two consecutive time units to obtain a spatiotemporal three-dimensional data block.

[0033] Among them, the real-time acquired radar data refers to the raw echo signal data received by the radar antenna in real time without subsequent preprocessing. It usually exists in the form of radio frequency digital signals and contains mixed information such as target reflection signals, clutter signals, and electromagnetic interference signals. It is the data input basis for the entire detection process. The real-time radar matrix refers to the radar data carrier presented in the form of a two-dimensional matrix after the real-time radar echo signal has been preprocessed. The matrix elements represent the radar echo energy or amplitude information of the corresponding range-azimuth unit. The historical radar matrix refers to the preceding radar data matrix that is temporally continuous with the real-time radar data and has undergone the same preprocessing process. In this embodiment, the time unit is set to 1 millisecond, corresponding to the radar pulse repetition period. The spatiotemporal three-dimensional data block is a data set that integrates time and spatial dimension information. It is formed by stacking two-dimensional radar matrices with different time units in chronological order.

[0034] In some implementations, air traffic control radar is used to acquire raw echo data at a fixed scanning period. Through pulse compression and moving target detection and data compression preprocessing, the raw echo data is converted into a two-dimensional range-Doppler spectrum and labeled as a real-time radar matrix. The real-time radar matrix is ​​then combined with the historical radar matrices corresponding to the two consecutive time units using a spatiotemporal sliding window method, ultimately forming a spatiotemporal three-dimensional data block of size 3×M×N (time×range×Doppler), which serves as the standard input for subsequent deep learning models.

[0035] S102. Input the spatiotemporal three-dimensional data block into the target segmentation model to obtain a binary segmentation map.

[0036] The target segmentation model is an encoder-decoder architecture 3D convolutional neural network specifically designed for processing spatiotemporal sequence data. The core of this network utilizes 3D convolutional kernels that slide across data blocks. Unlike ordinary 2D convolutions, which only process the length and width of an image, 3D convolutions operate additionally in the temporal dimension. The binary segmentation map is a visual output of the model after performing binary classification of target and background at each spatial location in the input data. It is a single-channel 2D matrix with the same size as the spatial dimension of the input data. Each pixel value in the map is either 0 or 1.

[0037] In some implementations, spatiotemporal 3D data blocks are input into the target segmentation model. By performing multi-level 3D convolution and nonlinear activation function processing on the spatiotemporal 3D data blocks, and introducing cross-dimensional adaptive weighting processing, the physical prior knowledge of the radar system is embedded into the network to enhance the target region and suppress clutter background, resulting in a probability map. Each pixel value in the probability map represents the probability of a target existing at that location. A final binary segmentation map is generated by thresholding the probability map.

[0038] S103. Input the spatiotemporal three-dimensional data block and the binary segmentation map into the target classification model to obtain several connected regions and extract multi-dimensional feature vectors.

[0039] The target classification model is a multi-dimensional feature-encoding neural network integrating feature extraction, feature fusion, and classification decision. Based on multi-dimensional features, this model dynamically fuses the importance of different feature sub-vectors through a cross-attention mechanism, ultimately calculating and outputting the probability value that each candidate region is the real target. A connected region is an independent contiguous block composed of all spatially adjacent (e.g., 8-neighborhood) target pixels in a binary segmentation image. Each connected region corresponds to a pre-segmented, physically complete candidate target instance, serving as the basic unit for subsequent independent feature extraction and classification. The multi-dimensional feature vector is a comprehensive numerical description extracted for a connected region, containing sub-features in four dimensions: geometric, statistical, motion, and motion compliance. This vector serves as the direct input to the classification model; after decomposition and encoding, it is weighted and fused through attention to form aggregated features for the final probability calculation.

[0040] In some implementations, spatiotemporal 3D data blocks and binary segmentation maps are input into the target classification model. Based on the binary segmentation map, an 8-neighborhood labeling algorithm is used for connected component analysis to identify all independent connected regions. For each connected region, geometric features, statistical features, motion features, and motion compliance features are simultaneously extracted from the corresponding spatiotemporal 3D data block. These features together constitute the multidimensional feature vector of the region.

[0041] S104. The multidimensional feature vectors are processed by a classification network to obtain the probability values ​​of each connected region being the target region.

[0042] The classification network processing is a deep computational process integrating feature encoding, interaction, and decision-making. It receives a multi-dimensional feature vector extracted from a single connected region as input, and uses a multi-path encoding-attention fusion neural network structure to analyze and fuse this vector, ultimately outputting a quantified confidence score, which is the probability that the region is the true target region. The core purpose of this process is to comprehensively utilize the multi-dimensional attributes of the target to achieve the transformation from coarse-grained candidate targets to refined scoring and verification.

[0043] In some implementations, the multidimensional feature vector is split into four feature sub-vectors, and each is input into three independent fully connected encoding sub-networks for nonlinear mapping and dimensionality reduction. The resulting three deep feature representations are then weighted and fused after the interaction weights are calculated by a cross-attention layer to generate an aggregated feature representation. This representation is passed through a fully connected layer and a sigmoid activation function, outputting a scalar value between 0 and 1 as the probability value that the connected region is the target region.

[0044] S105. Select connected regions with probability values ​​higher than a preset probability threshold and mark them as the initial target set; perform non-maximum suppression processing on the initial target set to obtain the final target set.

[0045] The initial target set refers to the set of connected regions whose probability values ​​exceed a preset threshold after being scored by the target classification model. It is a coarse-grained list of targets that has not undergone spatial deduplication and contains potential overlapping or neighbor detection results. The final target set is a refined list of targets that are spatially non-overlapping and have the highest confidence level, obtained by applying a non-maximum suppression algorithm to remove redundant detections based on the initial target set. It represents the final valid targets output by this detection process.

[0046] In some implementations, a fixed probability threshold is set, the selection of which must strike a balance between ensuring the target detection rate and suppressing false alarms. Regions with probability values ​​higher than this threshold in all connected regions are filtered out to form an initial target set. Non-maximum suppression is applied to the initial target set. The spatial location and range of each target are calculated based on the bounding rectangle of its corresponding connected region; the intersection-union ratio (IUR) between any two rectangles is also calculated. If the IUR of two rectangles exceeds a set overlap threshold, they are considered duplicate detections of the same real target, and the one with the higher probability value is retained while the other is discarded. This process is iterated until the overlap between all target pairs in the set is below the threshold. Finally, a high-confidence target set without redundancy is output as the list of valid targets detected in this radar scan cycle.

[0047] For example, a probability threshold of 0.65 is set. This threshold is a critical point in the probability distribution of the model output that can effectively distinguish high-confidence targets from a large number of low-confidence clutter. Below this value, the number of false alarms will increase significantly; above this value, some weak but real signals may be filtered out.

[0048] With an overlap threshold of 0.3, the air traffic control radar has a wide field of view, and the projections of real targets on the range-Doppler plane are typically sparse and independent. A lower overlap threshold is sufficient to distinguish and separate different targets, while effectively merging multiple highly overlapping detection boxes caused by target energy diffusion or small fluctuations in the detection boundary. This setting avoids misclassifying a single target as multiple targets and also prevents excessive suppression from causing nearby small real targets to be incorrectly deleted.

[0049] S106. Map the center pixel position of the final target set to distance-azimuth coordinates to obtain a list of target point information.

[0050] The target point information list is the final output of this detection method, and is a structured data list. Each item in the list corresponds to a confirmed real target in the final target set, and includes the target's range and azimuth coordinates in the radar polar coordinate system. This list transforms the pixel-level detection results output by the deep learning model into accurate target location information with clear physical meaning, conforming to the air traffic control surveillance system interface specifications.

[0051] In some implementations, the connected region corresponding to each target in the final target set is obtained, the centroid of its minimum bounding rectangle or pixel set is calculated, and this position is determined as the center pixel coordinate of the target on the binary segmentation map; coordinate mapping is performed according to the calibration parameters predetermined by the radar system to obtain a list of target point information.

[0052] Based on the above technical solutions, this application provides an end-to-end air traffic control primary radar target detection method. By constructing an end-to-end deep learning detection framework encompassing preprocessing, segmentation, feature extraction, classification verification, and post-processing, the entire process from raw echoes to a list of target points is automated, significantly reducing reliance on manual intervention and feature engineering experience. By combining continuous temporal data to form a spatiotemporal three-dimensional input and utilizing a three-dimensional convolutional network to fully exploit the joint correlation of targets in the temporal and spatial domains, the detectability of weak targets under strong clutter is effectively improved. By explicitly embedding the inherent physical constraints of the radar system into the segmentation network in an adaptive weighted form, the model's ability to suppress complex non-uniform clutter and its scene adaptability are significantly enhanced. By designing a multi-dimensional manual feature extraction method after segmentation, including geometric, statistical, motion, and compliance features, and employing an attention-based feature fusion network for refined verification, a clear and interpretable detection-discrimination decision path is formed, greatly improving the reliability of false alarm filtering and enhancing the credibility of the final output and the transparency of the system's decision-making process.

[0053] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 2 As shown, the above S101 can be implemented through the following S201 and S202, which are explained in detail below: S201. Perform pulse compression and moving target detection processing on the real-time acquired radar echo signal to obtain the real-time radar matrix.

[0054] In some implementations, this includes pulse compression processing of the real-time acquired radar echo signal to convert the original radar echo, which exists as a radio frequency digital signal, into a one-dimensional range image.

[0055] In some implementations, the radio frequency digital signal received by the radar antenna in real time is pulse compressed, and the radio frequency digital signal is convolved by a matched filter to compress the pulse width, thereby significantly improving the range resolution and outputting a high-resolution one-dimensional range image. It should be noted that the impulse response of the matched filter is the time-reversed complex conjugate of the transmitted signal. This process converts the continuous time-domain echo signal into a discrete sequence arranged according to range cells, thus obtaining a one-dimensional range image sequence.

[0056] Windowing and coherent accumulation are performed on the one-dimensional range image, and a fast Fourier transform is performed along the azimuth dimension to generate a two-dimensional range-Doppler spectrum representing the target velocity information.

[0057] In some implementations, the one-dimensional range image sequence is multiplied by a window function (Hamming window, Taylor window) in the azimuth dimension. Within a single range unit, a coherent accumulation is performed on the complex sequence of echoes for multiple consecutive pulse repetition periods (e.g., 32 or 64 pulses). After accumulation, a Fast Fourier Transform (FFT) is performed along the azimuth dimension: the windowed and accumulated slow-time sequence corresponding to each range unit is transformed to the Doppler frequency domain. By repeating this operation for all range units, a two-dimensional range-Doppler spectrum is finally generated, where one dimension represents the range units and the other represents the Doppler frequency units. The value of each unit is a complex number containing the signal energy and phase information within that range-Doppler unit.

[0058] The generated two-dimensional distance-Doppler spectrum is divided into sub-bands, and the background level is estimated based on the ordered statistical constant false alarm rate method and a preset k value.

[0059] The division of subbands includes: through calculation formulas Calculate the lower limit of clutter Doppler broadening caused by antenna scanning modulation. ,in, This refers to the antenna scanning angular velocity. The current observation distance, The wavelength at which the radar operates is used; the correlation bandwidth of the background power spectrum is calculated by inversely proportional to the radar's coherent processing interval CPI. This width determines the maximum permissible width of the background within the subband that can be considered smooth; subband reference width You must select from the following range: ;like Then, a compromise value is selected within the interval, such as the geometric mean of the two: ;like This indicates that the system's phase noise performance limits its resolution. In this case, only sub-band stability can be prioritized, and a forced setting must be implemented. And accept that some clutter may cross subbands. Through calculation... The number of sub-bands N is calculated, where, It is the pulse repetition frequency.

[0060] Determining the value of k includes: through The number of protection units G is calculated; where, This represents the number of distance cells within this sub-band. The G value represents the number of Doppler units within this sub-band. It signifies the maximum number of potentially strong signal units that need to be excluded from the background estimation samples within this local frequency domain to ensure the background estimation remains uncontaminated. This is calculated using the formula... The ordinal number used to estimate the background level is calculated. ;in, If it is not satisfied, then it means that the current situation is... If the size is too small, the subbands need to be re-divided.

[0061] In some implementations, for the generated two-dimensional range-Doppler spectrum, for each range cell, its Doppler spectrum is uniformly divided into N sub-bands, and for each sub-band, the number of spectral cells it contains is... For the total number of Doppler units The amplitudes of all spectral units within the sub-band are sorted in ascending order; the amplitude corresponding to sorting k is selected as the local background level estimate for the sub-band.

[0062] It should be noted that key parameters such as antenna scanning angular velocity Ω, operating wavelength λ, pulse repetition frequency (PRF), and minimum detection signal-to-noise ratio are crucial. These values ​​all originate from radar system design specifications and are inherent or preset values.

[0063] The amplitude of each spectral unit within the sub-band is compared with its corresponding local background level estimate. Spectral units with amplitudes higher than the background level by a preset multiple T are marked as candidate significant units. Isolated false alarms are further eliminated through distance continuity test. Finally, the amplitude of the retained significant spectral energy is detected and logarithmically compressed to obtain the real-time radar matrix.

[0064] The preset multiplier T is based on the minimum detection signal-to-noise ratio required by the radar system design. Through calculation formula Calculated.

[0065] In some implementations, the resulting candidate salient cells are determined by the coordinates of each cell. and the original complex spectrum value Characterization; where m is the distance cell index and n is the Doppler cell index.

[0066] The distance cell indices of each cell are subtracted from each other. All candidate salient cells with a difference of 1 are retained, and their corresponding complex spectral values ​​are used to construct the salient spectral energy. Each cell in the salient spectral energy set is traversed, and for each cell, the original complex spectral value is calculated. The modulus is used to obtain its corresponding complex spectral value. This operation converts complex energy into a positive scalar value representing signal strength. This is achieved through the calculation formula... Achieve dynamic range compression; among which, It is a very small positive number used to prevent logarithmic operations on zero.

[0067] The calculated The values ​​and their corresponding unit coordinates are used to construct a two-dimensional amplitude matrix, which is labeled as the real-time radar matrix.

[0068] It should be noted that, to ensure the final generated real-time radar matrix is ​​a structurally complete two-dimensional data matrix that can be directly used for subsequent spatiotemporal combination and network processing, all elements of the matrix must be explicitly assigned values ​​during the construction process. Specifically, when initializing the two-dimensional matrix, all its elements are pre-set to a constant value, denoted as C, representing the background or a state with no significant signal. This constant value C can be set according to the system noise floor or the display dynamic range, for example, set to 0, or a negative decibel value (such as -120dB) much lower than the typical amplitude of the target. Subsequently, only the data obtained after amplitude detection and logarithmic compression is processed... The values ​​are filled into this matrix corresponding to the coordinates of the salient spectrum energy units. At the specified position, the remaining unfilled positions in the matrix remain at the initial constant value C. Through this initialization and selective filling operation, the obtained real-time radar matrix has a complete two-dimensional grid structure in terms of data structure. Its row and column indices strictly correspond to the range and Doppler dimensions, ensuring that it can be used as regular input data to be aligned and stacked with historical radar matrices to form the spatiotemporal three-dimensional data blocks required for subsequent steps.

[0069] For example, the air traffic control radar operates at a frequency of 1.3 GHz (corresponding to a wavelength λ of 0.2308 meters) and a pulse repetition frequency (PRF) of 1000 Hz. The transmitted signal is linear frequency modulated (LFM), with a pulse width of 40 µs, a bandwidth B of 1 MHz, and a range resolution ΔR of 150 meters. The maximum detection range corresponds to 2000 range cells. The coherent processing interval is selected as 128 pulses (corresponding to 0.128 seconds), the Doppler resolution is 7.81 Hz, and the generated two-dimensional range-Doppler spectrum has a dimension of 2000 (range cells) × 128 (Doppler cells). The antenna scanning angular velocity Ω is 24 rpm (i.e., 2.513 rad / s). Calculate the lower limit of clutter Doppler broadening caused by antenna scanning modulation. Correlation bandwidth of background power spectrum ,therefore, Number of subbands The number of spectral units contained in each subband The target spans a maximum of 2 units in the distance dimension. The maximum span on the Dopplerweis is 1 unit. ; The number of protection units was calculated. , Preset multiplier .

[0070] After processing, the coordinates are at and The location is marked as a candidate significant unit, and its original complex spectral values ​​are respectively... , The significant spectral energy is obtained by passing the continuity test of the distance dimension (difference value is 1); logarithmic compression is then performed to obtain... , .

[0071] Initialize a 2000-row (distance cell) × 128-column (Doppler cell) real matrix, filling all elements with a constant value representing the background of -120dB. Then, calculate all the required... Fill the values ​​into the corresponding positions in the matrix, for example: (500,39)=32.04dB, (501,39)=31.60dB. The final two-dimensional amplitude matrix is ​​the real-time radar matrix.

[0072] S202. Stack the real-time radar matrix and the historical radar matrix in time sequence to obtain a spatiotemporal three-dimensional data block.

[0073] In some implementations, a common starting point for interception is selected from the real-time radar matrix and two stored historical radar matrices. Using this point as a reference, M cells are continuously extracted along the distance dimension and N channels are continuously extracted along the orientation dimension on the three matrices, thus obtaining three spatiotemporally aligned complex submatrices: (Current time unit: t) (Previous time unit t-1) (The first two time units are t-2).

[0074] Perform a two-dimensional discrete Fourier transform on each complex submatrix to obtain its spectrum. Define its spectrum index. Where u represents the index of the distance-frequency dimension, and v represents the index of the Doppler dimension. Extracting the low-frequency region at the center of the spectrum. , ; through calculation formula Calculated region Principal phase components of all spectral units ;in, For set Chinese index pairs The total number, that is, the total number of spectral units in the low-frequency region; Indicates taking the complex argument; Indicates the spectrum matrix at the index The complex values ​​at the given points. The principal phase components are obtained by calculating the three submatrices separately. , , Through calculation formula and calculation formula The global phase rotation corresponding to the historical complex submatrix is ​​calculated; where, This is a phase winding operation. It is calculated using the formula... Perform reverse phase compensation on each element z in the historical complex submatrix, where j is the imaginary unit, thereby obtaining the phase-compensated historical submatrix. and Stack the three submatrices in time order to form a three-dimensional complex tensor. Its dimensions are This data block is a spatiotemporal three-dimensional data block, which provides phase-aligned coherent echo information of the local spatial domain in three consecutive frames.

[0075] For example, the two-dimensional distance-Doppler spectrum has a dimension of 2000 (distance units) × 128 (Doppler units). Select the starting point for truncation. By extracting M=64 cells along the distance dimension and N=16 channels along the orientation dimension, three complex submatrices are obtained. , , Each matrix has a size of 64×16. Perform a two-dimensional FFT to obtain its spectrum. Define the low-frequency region size as P=5, Q=3, then Includes spectrum center The center has 5×3 spectral units, totaling Each unit is used. The average argument of the unit vector within this region is calculated to obtain... principal phase component Radius. Similarly, the calculation yields... principal phase component radian, principal phase component Radians. The global phase rotation is calculated. radian, Radians. Inverse phase compensation is performed on each element z in the historical complex submatrix to obtain the phase-compensated historical submatrix. and And stack them in chronological order to obtain spatiotemporal three-dimensional data blocks. This data block provides three consecutive frames (each with a 1ms time interval) of phase-aligned coherent echo data in a local region (distance cells 500 to 563 and Doppler cells 32 to 47).

[0076] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 3 As shown, the above S102 can be specifically implemented through the following S301, S302, S303 and S304, which are explained in detail below: S301. Perform three-dimensional convolution and nonlinear activation processing on the spatiotemporal three-dimensional data block to obtain the first feature map.

[0077] In some implementations, a spatiotemporal three-dimensional data block with a size of 3×M×N is used. Input model. Features are extracted from the input data block using a 3D convolutional layer. The kernel size of this convolutional layer is [size missing]. According to the calculation formula Perform initial feature extraction; among which, Represents a three-dimensional convolution operation. and For learnable weights and biases, This indicates a batch normalization operation. To modify the activation function of the linear unit, This is the first feature map. Its number of channels.

[0078] S302. Perform cross-dimensional adaptive weighting processing on the first feature map to obtain the second feature map.

[0079] Specifically, cross-dimensional adaptive weighting refers to calculating weight vectors along four dimensions: distance, orientation, time, and channel, and then calibrating the feature map after fusion.

[0080] In some implementations, according to the radar equations, the signal strength varies with distance R. Attenuation. Combined with the system's minimum detection range. Distance resolution and atmospheric attenuation factor, calculated by formula The actual distance corresponding to distance unit m is calculated; through the calculation formula Calculate the theoretical attenuation compensation coefficient for each distance cell m. Then, the distance dimension weight vector is obtained by normalizing through the maximum and minimum values. According to the antenna pattern function G(θ), the azimuth angle corresponding to each azimuth channel n is... Calculate its azimuth angle relative to the beam center. Gain ratio ,Will Perform maximum-minimum normalization to obtain the normalized azimuth dimension weight vector. For the input spatiotemporal three-dimensional data block Calculate the variance of all cell values ​​in each time slice i (i=1,2,3) along the distance and orientation dimensions. The three variance values ​​were then subjected to Softmax normalization to obtain the time-dimensional weight vector. Read the calibration gain vector of each physical receiving channel from the radar system's built-in configuration file. Through calculation formula The normalization coefficient for each physical channel was calculated. Linear interpolation is used for mapping: a function is defined on the physical channel index [0, N-1]. At the new uniform interval point Linear interpolation is performed on the above to obtain the channel dimension weight vector. .

[0081] The four weight vectors are fused into a four-dimensional weight tensor W, which is then multiplied element-wise with the first feature map for calibration. This calibration is performed using the calculated formula... The second feature map is obtained. This is the element-wise multiplication operation of a matrix.

[0082] S303. Perform spatial downsampling, three-dimensional convolution, and nonlinear activation on the second feature map to obtain the third feature map.

[0083] For example, a three-dimensional convolutional layer with a stride of (1,2,2) is used to... To perform spatial downsampling, the convolution kernel size can be (1,3,3), and the number of output channels is... After downsampling, batch normalization and ReLU activation are performed to obtain a size of... The third feature map ,in, , .

[0084] S304, multi-scale fusion and probability output processing, to obtain a binary segmentation map.

[0085] In some implementations, the third feature map is upsampled and concatenated with the first feature map along the channel dimension to obtain a fused feature map. A 3D transposed convolution is then used to... The feature map is obtained by upsampling back to M×N spatial dimensions. .Will Compared with the first stage By stitching along the channel dimension, a fused feature map is obtained. .

[0086] Then, the fused feature map is subjected to 3D convolution and nonlinear activation processing to obtain a single-channel feature map. The input is further integrated into a layer consisting of 3D convolution, batch normalization, and ReLU activation. Finally, a 1×1×1 3D convolutional layer with 1 output channel is used to process the data, resulting in a single-channel feature map. .

[0087] The single-channel feature map is processed by the Sigmoid activation function to obtain a probability map. Applying the Sigmoid function for nonlinear mapping The resulting probability graph In, each pixel value Indicates the spatial location The probability of the presence of a radar target.

[0088] Pixels with probability values ​​greater than a preset segmentation threshold are marked as target pixels, and non-target pixels are marked as background pixels, resulting in a binary segmentation map. The preset segmentation threshold is not a fixed value, but rather adaptively adjusted from the probability map using the Otsu algorithm. The Otsu algorithm determines the optimal global threshold by maximizing the inter-class variance between the foreground and background classes. After calculating the threshold, a binary segmentation map is generated according to the following rules. : In this context, pixels with a value of 1 are marked as target pixels, and pixels with a value of 0 are marked as background pixels.

[0089] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 4 As shown, the above S103 can be implemented through the following S401 and S402, which are explained in detail below: S401, Connected Component Analysis and Processing.

[0090] In some implementations, an 8-neighbor labeling algorithm is used to process the binary segmentation image through a two-pass scan. In the first scan, for each target pixel (pixel with a value of 1), label equivalence relations are assigned and recorded based on the labels of its upper-left, upper, upper-right, and left-side neighboring pixels. In the second scan, the equivalence relations are parsed, and all pixels belonging to the same connected region are assigned a unique label identifier, thereby identifying all independent, spatially continuous connected regions and recording basic attributes such as the pixel set and bounding rectangle boundary of each region.

[0091] For example, for a binary segmentation graph of size 2000 (distance dimension) × 128 (Doppler dimension), an 8-neighbor labeling algorithm is used for connected component analysis. After processing using the two-pass scanning method, 47 connected regions are obtained; the following are 6 representative independent connected regions and their precise attributes: Region 1: Consists of 215 target pixels, with a bounded rectangle containing distance cells [1010, 1025] and Doppler cells [60, 72]. This region is relatively large and compact, corresponding to a high-confidence real moving target.

[0092] Region 2: Consists of 98 target pixels, with a bounded rectangle of: distance unit [480, 489], Doppler unit [45, 53]. This region corresponds to a weaker or more distant candidate target.

[0093] Region 3: Consists of only 3 pixels, scattered across distance cells [255, 256] and Doppler cells [90, 91]. This is likely a noise point.

[0094] Region 4: Consists of 15 pixels, with a bounded rectangle of: distance cells [1500, 1502], Doppler cells [10, 14]. This could be tiny debris generated by ground clutter or intermittent interference.

[0095] Region 5: Consists of 55 pixels, with a bounded rectangle of range cells [1750, 1756] and Doppler cells [100, 105]. This region may contain weak targets or clutter blocks with some correlation.

[0096] Region 6: Consists of 30 pixels, elongated in shape, with a bounded rectangle containing: distance cells [50, 55] and Doppler cells [5, 12]. It may represent a target moving tangentially or a special form of interference.

[0097] S402, Multidimensional Feature Vector Extraction Processing.

[0098] In some implementations, for each labeled connected region, multidimensional features are simultaneously extracted from the spatiotemporal 3D data block that precisely corresponds to its spatial location. Specifically: The total number of pixels in the connected region is calculated as the area A. The aspect ratio is calculated based on the span W of its bounding rectangle in the Doppler dimension and the span H in the distance dimension. This yields the geometric properties.

[0099] Extract all amplitude data points from the corresponding data region. Calculate its mean with standard deviation And calculate the normalized information entropy. To quantify the disorder of the amplitude distribution, among which Let K be the probability that the amplitude falls into the j-th quantization interval, K be the total number of intervals, and N be the total number of data points corresponding to the connected region; then the statistical characteristics are obtained.

[0100] Extract corresponding data sub-blocks from three consecutive time slices, calculate the amplitude-weighted centroid coordinates at time k, and their distance from the centroid. Doppler center of mass The calculation is similar; where, The data amplitude at the i-th position in the k-th time slice. and The distance cell index and Doppler cell index corresponding to the i-th position are used. The displacement vector is calculated based on the centroid coordinate difference between adjacent time points, and the average radial velocity v is derived. The motion characteristics are obtained.

[0101] Estimating radial acceleration using radial velocities at consecutive time points ,in The pulse repetition interval is based on the maximum reasonable acceleration threshold. Through function The motion compliance features are calculated. Finally, the geometric features, statistical features, motion features, and motion compliance features are concatenated in a predetermined order to obtain a multidimensional feature vector z.

[0102] It should be noted that the maximum reasonable acceleration threshold is a key parameter based on prior physical knowledge and the specific task scenario. Its parameter setting is determined by the physical dynamic limits of the target moving entity. For example, for civil airliners, their typical maximum acceleration capability can be based on the aircraft performance manual or industry standards (such as RTCADO-185B).

[0103] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 5 As shown, the above S104 can be specifically implemented through the following S501 to S503, which are explained in detail below: S501, Multidimensional Feature Encoding and Deep Feature Representation.

[0104] Multidimensional feature encoding refers to the process of nonlinearly transforming and abstracting multiple feature sub-vectors describing different attributes of a target through independent neural network modules. Deep feature representation refers to the feature vectors generated after the original features are mapped through the above encoding network, which contain higher-level semantic information.

[0105] In some implementations, the multidimensional feature vector z is split into four sub-vectors, which are geometric feature sub-vectors. Statistical feature vectors Motion feature vectors and motion compliance feature subvector Each feature vector is fed into a separate fully connected encoding sub-network. Each sub-network has the same two-layer structure but does not share parameters. (The geometric feature vectors are then used as inputs.) Taking the processing of [the data] as an example, the process is as follows: [The process involves] calculating... The input is mapped to a 64-dimensional hidden space; where, The weight matrix is ​​obtained by training the model. For bias vectors, Presentation layer normalization operation, This is the activation function. It is calculated using the formula... The features are further compressed to a unified deep representation dimension. The statistical feature sub-vectors, motion characteristic sub-vectors, and motion compliance feature sub-vectors are then passed through corresponding encoding sub-networks to obtain... , and .

[0106] For example, the encoding process and result of a 125-dimensional original multidimensional feature vector extracted from region 1 can be specifically described as follows: The geometric eigenvector of this vector Including pixel area and aspect ratio, after processing by its dedicated encoding sub-network, a 32-dimensional deep feature representation is obtained. In this representation, the activation values ​​for the 7th and 19th dimensions, corresponding to shape compactness and large size semantics, are 0.85 and 0.92, respectively; this encodes the region as an abstract concept of a large, well-defined entity.

[0107] Multidimensional features of the same region are encoded in parallel: their statistical feature subvectors generated In the middle, the value of the third dimension is 0.91; its motion feature subvector generated In the model, the 12th dimension representing stable radial motion has a value of 0.88; its compliance feature subvector... generated In the model, the reasonable value for the first dimension representing acceleration is 0.95. The outputs of the four sub-networks together form a comprehensive, high-level abstract description of the real target.

[0108] In contrast, for the feature vector extracted from region 3, its compliance feature subvector... generated In the data, the first dimension representing acceleration has a reasonable value of 0.08. As for the feature vector of region 5, the deep feature representations obtained after encoding... In the figure, the reasonable value of the first dimension representing acceleration is 0.60.

[0109] This step transforms the original features into a common 32-dimensional deep semantic space through parallel nonlinear mapping. In this space, the feature value distributions of candidate regions of different categories differ significantly, providing highly discriminative input for subsequent processing.

[0110] S502, cross-attention weighted fusion.

[0111] In some implementations, , , and spliced ​​into a matrix ; Generate queries through learnable linear transformations ,key Sum matrix; , and The weight parameters are calculated using the formula Calculate the attention score and normalize it to obtain the attention weight matrix. ; is the dimension of the linear projection.

[0112] matrix The element in the i-th row and j-th column This represents the degree of attention given to the j-th feature by the i-th feature. Taking the weight value of each column of this matrix yields a four-dimensional attention weight vector. ,satisfy Based on this weight vector, the original deep feature representations are weighted and fused to obtain the aggregated feature representation. .

[0113] For example, for region 1, the cross-attention layer calculates the weight vector as follows: This indicates that motion characteristics are important when determining the region. and geometric features They are given higher importance and ultimately aggregate features It is a weighted combination of these features.

[0114] For region 3, its weight vector is Compliance characteristics and statistical characteristics The main weights were obtained, which led to the fusion result. It mainly amplifies the characteristics of its irrational motion and chaotic energy distribution.

[0115] For region 5, its weight vector is This shows that the model balances and explores multiple features.

[0116] Through this step, the model can adaptively and dynamically adjust the contributions of different feature dimensions according to the specific circumstances of each candidate target, thereby generating an aggregated representation that focuses on the most relevant information for the final decision.

[0117] S503, Classification Scoring and Probability Generation.

[0118] In this context, classification scoring refers to mapping the aggregated feature representation to a raw scalar score through a classifier. This score represents the raw confidence that the input feature belongs to the target category.

[0119] In some implementations, aggregated features are represented The classification score is obtained by linear transformation through a fully connected classification layer. ; For the weight vector, This is a bias term. The final probability value is generated by performing nonlinear compression and normalization using the Sigmoid function. This probability value This is the final probability estimate of whether the currently processed connected region belongs to the real target region.

[0120] For example, the aggregated feature representations of regions 1, 3, and 5 are calculated to obtain the following results: , , ; corresponding probability value , , .

[0121] This step ultimately outputs a specific probability value for each connected region, completing the final transformation from features to classification confidence. High probability values ​​(such as 0.913) indicate high-confidence object detection, while low probability values ​​(such as 0.069) correspond to suppressed false alarms.

[0122] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 6 As shown, the above S106 can be implemented through the following S601 and S602, which are explained in detail below: S601, Coordinate Mapping.

[0123] In some implementations, for each target in the final target set, its corresponding connected region is obtained. The arithmetic mean of the row indices (corresponding distance cells) of all target pixels within that connected region in the binary segmentation map is then calculated. Arithmetic mean of column indexes (corresponding to azimuth / Doppler units) , this The center pixel coordinates of the target in the image domain are determined. Then, physical coordinate transformation is performed based on the inherent spatial calibration parameters of the radar system. This is achieved through calculation... The actual distance R to the target is calculated; where, This is the distance from the starting offset. This represents the system's distance resolution. Calculated using the formula... The actual azimuth angle θ of the target is calculated; where, The starting angle of the azimuth. This represents the azimuth resolution of the system.

[0124] S602, Data Encapsulation.

[0125] In some implementations, the calculated physical coordinates of each target, the calculated classification probability value p, and the timestamp (UTC time, accurate to milliseconds) of the current frame radar data are encapsulated into a structured data record. These structured records of all targets are then summarized sequentially (e.g., by distance from nearest to farthest) to generate the final list of target point information.

[0126] For example, the calibration parameters of a radar system are: range initial offset. Meters, distance resolution ΔR = 30 meters / unit; azimuth starting angle degrees, azimuth resolution Degree / unit. The timestamp of the current frame data is 2026-01-15-14:30:01.025-UTC. The center pixel coordinates of the connected region of target A corresponding to region 1. , Classification probability value The actual distance is calculated through coordinate mapping. meters, azimuth Degree; the encapsulated structured data record is as follows: The center pixel coordinates of the connected region of target B corresponding to region 5. , Classification probability value The actual distance is calculated through coordinate mapping. meters, azimuth Degree; the encapsulated structured data record is as follows: .

[0127] In one possible implementation, combining Figure 1 ,like Figure 7 As shown, following S106, the end-to-end air traffic control primary radar target detection method provided in this application embodiment includes the following S701 to S703: S701. Obtaining the training dataset.

[0128] The training dataset refers to the data set used to train and optimize the target segmentation model and the target classification model. It consists of a large number of input samples (spatiotemporal 3D data blocks) and corresponding output supervision labels (ground truth labels of target locations), and is the foundation for the model to learn its target detection capabilities.

[0129] In some implementations, based on ray tracing and radar equations, radar echoes are simulated under different environmental clutter, target radar cross-sections, target motion trajectories, and noise levels. Simultaneously, precise pixel-level masks of target positions are generated as ground truth labels for the target positions. Raw echo data received by the primary air traffic control radar during actual operation is collected. Ordered statistical constant false alarm rate (CFAR) detection is applied to this data. Detection results with extremely high confidence (such as stably occurring points appearing in multiple consecutive frames) are used as pseudo-ground values ​​and labeled to obtain ground truth labels for the target positions, used to characterize clutter and interference characteristics in the real environment. The simulated and real raw echo data are preprocessed using the same preprocessing procedures as real-time processing to obtain standard spatiotemporal 3D data blocks. Simultaneously, their corresponding ground truth labels are aligned and normalized. Finally, the spatiotemporal 3D data blocks are paired one-to-one with their ground truth labels to form a complete training dataset.

[0130] For example, simulated radar data was generated using the RadarSimPy tool, simulating the echo of X-band radar in a scenario including building clutter and multiple civilian aircraft targets, generating 10,000 data-label pairs. Real data was obtained from three months of historical records of an ASR-10 primary radar at an airport, resulting in 2,000 valid data-label pairs after filtering and pseudo-ground value labeling. The two sets of data were mixed and randomly shuffled, then divided into training, validation, and test sets in an 8:1:1 ratio. The training set contained 9,600 samples, forming the basis for model training.

[0131] S702, Training of the target segmentation model.

[0132] In some implementations, a batch of spatiotemporal 3D data blocks is taken from the training dataset and input into the target segmentation model to obtain a predicted segmentation probability map. This is then calculated using... Computational spatial adaptive weighted focus loss; where, It is a distance-dependent unit The increased weight coefficients linearly increase the loss weights with increasing distance, forcing the model to pay attention to distant regions with severe signal attenuation during training. It is the posterior probability predicted by the model that the pixel is the target. This is used to reduce the loss contribution of easily classified samples, allowing training to focus more on difficult-to-classify samples. This is used to balance the impact of an imbalance in the number of positive and negative samples. Through the calculation formula... Calculate the topology consistency loss ;in, and These represent the number of connected regions in the prediction graph and the ground truth graph, respectively. This represents the number of connected regions in the truth graph. and These represent the pixel areas of the c-th connected region after sorting by area in the predicted and ground truth maps, respectively. This loss function aims to constrain the prediction results from an overall structural perspective, avoiding the generation of too many fragmented noise regions or excessive merging of independent targets. Final Loss ,in, These are the weighting coefficients used to balance the two losses.

[0133] Using the calculated total loss The Adam adaptive moment estimation algorithm is used to update the trainable parameters of all 3D convolutional layers and cross-dimensional weighted layers in the target segmentation model through direction propagation. This process is repeated iteratively until the validation set loss no longer decreases over several consecutive preset training epochs, and the average intersection-over-union ratio (IoU) of the validation set tends to stabilize. At this point, the model is considered to have converged, and training is stopped. This criterion ensures that the model has generalized ability on unseen data, rather than overfitting to the training data.

[0134] For example, in one training session, the initial learning rate is set to... Batch size is 8, focus loss parameter , Topological loss weights After 50 training cycles, early stopping was triggered because the validation set loss did not reach a new low for 10 consecutive cycles. At this point, the model's average intersection-union ratio on the validation set stabilized at 0.82, the segmentation results showed good integrity in both short and long ranges, and the false fragment regions in the prediction map were significantly reduced.

[0135] S703, Training of the target classification model.

[0136] In some implementations, all spatiotemporal 3D data blocks in the training dataset are forward-inferred using a trained target segmentation model to obtain a binary segmentation map. Connectivity analysis and multi-dimensional feature vector extraction are then performed to obtain a multi-dimensional feature vector corresponding to each connected region. For each connected region, the maximum intersection-union ratio (CIU) with all target regions in the corresponding ground truth labels is calculated. If this value is greater than a preset threshold, the feature vector of that connected region is labeled as a positive sample (label 1); otherwise, it is labeled as a negative sample (label 0). This process generates a large number of labeled feature vector samples.

[0137] The labeled feature vector sample set is input into the target classification model. The model outputs the predicted probability through forward propagation and uses the binary cross-entropy loss function. Calculate the prediction error; where, For sample labels, To predict probabilities for the model, The batch size is specified. Backpropagation and optimizer iterations are performed using Adam to update model parameters. After each training epoch, the model's classification F1 score on an independent validation set is calculated. If the F1 score does not improve within 10 consecutive epochs, the model performance is considered converged, training is stopped, and the model parameters at the best performance on the validation set are saved, thus obtaining the final target classification model.

[0138] For example, using 9600 training samples from the training dataset, the segmentation model generates approximately 45,000 connected component feature vectors, of which approximately 12% are labeled as positive samples and 88% as negative samples. A classification model is then trained using this dataset, with a learning rate set to... The batch size was 64. The training process automatically terminated after the F1 score on the validation set did not exceed the peak value of 0.943 for 10 consecutive epochs. The final target classification model achieved a classification accuracy of 99.2% and a recall of 95.7% on the independent test feature set, indicating that it can effectively distinguish between real targets and false alarms.

[0139] Based on the above technical solution, the layer-by-layer filtering of continuous frame echoes using the 3D convolutional kernel in this scheme effectively mines the temporal coherence of the target, enabling weak target signals to accumulate and be significantly enhanced in the time dimension. The cross-dimensional adaptive weighting module explicitly embeds physical prior knowledge such as the range attenuation weights calculated from the radar equations and the antenna pattern gain weights into the network, giving the model an adaptive background suppression capability that matches the characteristics of the radar system. Combining four-dimensional handcrafted features and a cross-attention mechanism, the classification decision process can dynamically evaluate the contribution of each dimension of features (such as the key role of motion compliance features in detecting abnormal maneuvers), forming a clear and traceable detection verification decision path. This improves the probability of detecting low signal-to-noise ratio targets while reliably suppressing false alarms under complex non-uniform clutter, significantly improving the overall system's detection robustness and result interpretability.

[0140] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0141] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and modifications.

Claims

1. A method for end-to-end air traffic control primary radar target detection, characterized in that, include: The real-time radar data is preprocessed to obtain a real-time radar matrix, which is then combined with the historical radar matrices corresponding to the two consecutive time units to obtain a spatiotemporal three-dimensional data block. A spatiotemporal 3D data block is input into a target segmentation model to obtain a binary segmentation map; the spatiotemporal 3D data block and the binary segmentation map are input into a target classification model to obtain several connected regions and extract multidimensional feature vectors; the multidimensional feature vectors are processed by a classification network to obtain the probability value of each connected region being the target region; the target segmentation model and the target classification model are trained by a 3D convolutional neural network and a multidimensional feature encoding neural network, respectively. Connected regions with probability values ​​higher than a preset probability threshold are selected and marked as the initial target set; non-maximum suppression is applied to the initial target set to obtain the final target set; The center pixel position of the final target set is mapped to distance-azimuth coordinates to obtain a list of target point information.

2. The end-to-end air traffic control primary radar target detection method according to claim 1, characterized in that, The methods for acquiring the spatiotemporal three-dimensional data blocks include: From the real-time radar matrix and the two historical radar matrices, M identical range cells are continuously extracted along the range dimension, and N identical azimuth channels are continuously extracted along the azimuth dimension to obtain complex submatrices; Two-dimensional Fourier transforms are performed on the complex submatrices respectively to extract the low-frequency region of the corresponding spectrum and obtain the corresponding principal phase component by calculating the statistical average of the phase angles; The global phase rotation of the complex submatrix of the historical radar matrix is ​​obtained by calculating the difference between the principal phase component of the complex submatrix corresponding to the historical radar matrix and the principal phase component of the complex submatrix corresponding to the real-time radar matrix. Based on the global phase rotation, the complex submatrix of the historical radar matrix is ​​subjected to reverse phase compensation, and stacked with the complex submatrix corresponding to the real-time radar matrix in chronological order to obtain a spatiotemporal three-dimensional data block with a size of 3×M×N.

3. The end-to-end air traffic control primary radar target detection method according to claim 1, characterized in that, The method for obtaining the binary segmentation map includes: The spatiotemporal three-dimensional data block is subjected to three-dimensional convolution and nonlinear activation processing to obtain the first feature map; The first feature map is subjected to cross-dimensional adaptive weighting to obtain the second feature map; The second feature map is subjected to spatial downsampling, three-dimensional convolution, and non-linear activation to obtain the third feature map; The third feature map is upsampled and then connected to the first feature map along the channel dimension to obtain a fused feature map. The fused feature map is subjected to 3D convolution and nonlinear activation processing to obtain a single-channel feature map; The single-channel feature map is processed by the Sigmoid activation function to obtain a probability map; each pixel value in the probability map represents the probability that a target exists at that location; Pixels with a probability value greater than a preset segmentation threshold are marked as target pixels, and non-target pixels are marked as background pixels, thus obtaining a binary segmentation map.

4. The end-to-end air traffic control primary radar target detection method according to claim 3, characterized in that, The methods for obtaining the second feature map include: Calculate the signal attenuation curve based on the inherent system parameters of the radar, generate the compensation coefficient corresponding to each range cell of the first feature map, and mark it as a range dimension weight vector; The radar antenna pattern data is obtained based on the inherent antenna design parameters of the radar, and the gain ratio of each directional channel relative to the beam center is calculated to obtain the azimuth weight vector. Based on spatiotemporal three-dimensional data blocks, the time series statistical variance of each distance orientation unit is calculated, and nonlinear mapping processing is performed to obtain normalized compensation coefficients, which are marked as time dimension weight vectors. Based on the inherent hardware parameters of the radar receiving channel, the calibration gain value of each channel is obtained and the gain normalization coefficient of each receiving channel is calculated to generate a channel-dimensional weight vector. The distance dimension weight vector, orientation dimension weight vector, time dimension weight vector, and channel dimension weight vector are fused and then multiplied element-wise on the first feature map to obtain the second feature map.

5. The end-to-end air traffic control primary radar target detection method according to claim 1, characterized in that, The methods for obtaining the probability values ​​of each connected region being the target region include: The binary segmentation map and the spatiotemporal 3D data block are input into the target classification model; Through connected component analysis, several connected regions are obtained, and multidimensional feature vectors are extracted; the multidimensional feature vectors include: geometric features, statistical features, motion features, and motion compliance features; The multidimensional feature vector is split into four feature sub-vectors and input into four independent fully connected coding sub-networks for linear mapping to obtain four corresponding deep feature representations. The four deep feature representations are concatenated and input into the cross-attention layer. The attention weights between the four deep feature representations are calculated to obtain a four-dimensional weight vector. The four feature sub-vectors are then weighted and fused based on the weight vector to obtain the aggregated feature representation; The aggregated feature representation is mapped to a classification score through a fully connected layer. The classification score is then nonlinearly compressed and normalized using the Sigmoid function to obtain the probability value that the connected region belongs to the target region.

6. The end-to-end air traffic control primary radar target detection method according to claim 5, characterized in that, The methods for obtaining the multidimensional feature vectors include: Based on the pixel position of each connected region in the binary segmentation map, the number of pixels in the connected region and the aspect ratio of the bounding rectangle are calculated and marked as geometric features; Extract the values ​​of all matrix elements within the corresponding spatial region from the spatiotemporal three-dimensional data block, and calculate the mean and variance of the matrix elements as statistical features; Matrix elements of the corresponding spatial regions are extracted from three consecutive time slices of the spatiotemporal three-dimensional data block. The displacement is obtained by calculating the centroid coordinates of each slice and the difference between the centroid coordinates of adjacent time slices, and is marked as motion features. The radial acceleration of the target is calculated using second-order difference based on the displacement and compared with a preset target acceleration threshold to obtain a compliance flag bit marked as a motion compliance feature; The feature values ​​contained in the geometric features, statistical features, motion features, and motion compliance features are concatenated to obtain the multidimensional feature vector corresponding to the connected region.

7. The end-to-end air traffic control primary radar target detection method according to claim 1, characterized in that, The methods for obtaining the target point information list include: Calculate the mean row index and mean column index of all pixels in each connected region of the final target set in the binary segmentation image; and convert them into the target range value and azimuth value respectively according to the inherent spatial calibration parameters of the radar system. The distance, azimuth, probability, and radar data timestamp of each target are encapsulated into structured data records, and the results are aggregated to produce a list of target point information.

8. The end-to-end air traffic control primary radar target detection method according to claim 1, characterized in that, The step of preprocessing the real-time acquired radar data to obtain the real-time radar matrix includes: The real-time acquired radar echo signals are pulse compressed to obtain a one-dimensional range image sequence; The one-dimensional range image sequence is windowed and coherently accumulated, and then subjected to fast Fourier transform along the azimuth dimension to obtain a two-dimensional range-Doppler spectrum. The Doppler spectrum of each distance unit is divided into several sub-bands according to a preset Doppler frequency interval. Based on the statistical distribution of the spectral amplitude in each sub-band, the background level is obtained through ordered statistical processing. Spectral units with amplitudes higher than the background level by a preset multiple are marked as candidate significant units. Non-continuous candidate significant units in the distance dimension are removed to obtain significant spectral energy. The amplitude of the significant spectral energy is detected and logarithmically compressed to obtain an amplitude data matrix, which is labeled as the real-time radar matrix.

9. The end-to-end air traffic control primary radar target detection method according to claim 1, characterized in that, The methods for obtaining the target segmentation model and the target classification model include: Construct a training dataset; the training dataset includes: spatiotemporal three-dimensional data blocks corresponding to simulated and real radar data and ground truth labels for target locations; Using the spatiotemporal three-dimensional data block as input and the target location ground truth label as supervision, a predicted segmentation map is obtained through forward propagation calculation; The spatial adaptive weighted focus loss and topological consistency loss are calculated based on the predicted segmentation map and then summed in weight to obtain the final error value. Based on the final error value, the parameters are optimized through the backpropagation algorithm until the model converges. The training dataset is processed by the target segmentation model to generate a binary segmentation map and perform connected component analysis to obtain the multidimensional feature vector of each connected region. Calculate the overlap between each connected region and the spatial range defined by the corresponding target location ground value, and label connected regions that exceed the preset labeling threshold as positive samples, and label the rest as negative samples; Using the multidimensional feature vector as input, and the labeled positive and negative samples as supervision labels, the class prediction value is calculated through forward propagation; the prediction error is calculated through the cross-entropy loss function; and the network parameters are iteratively optimized through the backpropagation algorithm until the model converges, thus obtaining the target classification model.

10. The end-to-end air traffic control primary radar target detection method according to claim 9, characterized in that, The methods for obtaining the training dataset include: Simulated radar data is generated by a radar data simulator and the target position true value label is output synchronously. The real data is received by the radar antenna, and the target position detection result is obtained by ordered statistical constant false alarm detection, and marked as the target position true value label; Simulated radar data and real radar data are processed into spatiotemporal three-dimensional data blocks, and paired with corresponding ground truth labels to form the training dataset.

Images