Double-path target detection method for complex ground reflection environment

Abstract

The invention relates to the technical field of radar signal processing, and discloses a complex ground reflection environment-oriented dual-path target detection method, which comprises the following steps of: performing quantile truncation and discretization processing on original range profile data, mapping a processed signal to a D-channel discrete feature space and generating a D-dimensional discrete feature matrix; the D-dimensional discrete feature matrix is input into a dual-channel parallel processing architecture, a CFAR adaptive detection module is adopted in a first path to suppress strong clutters to generate a first response matrix, and a false attention mechanism is adopted in a second path to enhance weak targets to generate a second response matrix; performing spatial alignment and weighting processing on the first response matrix and the second response matrix, and performing cross-domain confidence fusion to output a comprehensive response matrix; and carrying out fixed threshold binarization processing on the comprehensive response matrix based on a global threshold to obtain a target detection result. The method has both local gain compensation and global adaptive suppression, and can realize more robust detection of a weak target under a complex background.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing technology, specifically a dual-path target detection method for complex ground reflection environments. Background Technology

[0002] In recent years, low-altitude unmanned aerial vehicles (UAVs) have rapidly gained popularity in fields such as security monitoring, disaster relief, and agricultural inspection due to their versatility and cost-effectiveness. Existing low-altitude UAVs generally employ radar target detection technology. The Constant False Alarm Rate (CFAR) algorithm is a classic target detection method in radar systems. It divides the area around the target range cell into training and protection cells, adaptively sets a detection threshold using the statistical characteristics of the training cells (such as arithmetic mean, median, or ordered statistics), and then compares the echo amplitude of the target cell with this threshold to determine the presence of a target. While existing CFAR algorithms are reliable in ideal, uniform clutter scenarios, they suffer from the following shortcomings when applied to complex low-altitude near-ground environments: reflections of radar signals from densely distributed buildings, vegetation, and metal structures, multipath interference, and dynamic clutter coupling effects are prevalent. 1. The superposition of reflection and dynamic clutter causes weak target signals to be masked. Traditional CFAR thresholds are difficult to compensate for the detection of weak echoes, resulting in a significant decrease in detection rate. 2. Interference or terrain edge reflections cause drastic fluctuations in the statistics of training units, and the threshold rises or falls unstablely, making it difficult to maintain the predetermined false alarm rate; 3. It relies on local statistical features, ignores information fusion across distances and channels, has poor environmental adaptability, and cannot dynamically optimize detection strategies under complex multipath conditions; 4. Parameters (such as window size, sorting position, etc.) are highly dependent on experience-based parameter tuning, lacking interpretability and scalability, which is not conducive to the subsequent integration of multi-source data fusion or deep recognition modules.

[0003] Therefore, existing CFAR algorithms struggle to simultaneously achieve high detection rates for weak targets and low false alarm rates for strong clutter in complex near-ground multipath reflection and dynamic clutter environments, posing a serious challenge to target detection technology in complex near-ground electromagnetic environments. Summary of the Invention

[0004] The purpose of this invention is to solve the problem that existing radar target detection technologies are unable to simultaneously achieve high detection of weak targets and low false alarms against strong clutter in complex multipath reflection and dynamic clutter environments near the ground. It provides a dual-path target detection method for complex ground reflection environments, which combines local gain compensation and global adaptive suppression to improve the reliability and robustness of low-altitude near-ground target detection and achieve more robust detection of weak targets in complex backgrounds.

[0005] The objective of this invention is mainly achieved through the following technical solutions: A dual-path target detection method for complex ground reflection environments includes the following steps: Step S1: Obtain the original range image data, perform quantile truncation and discretization on the original range image data through an adaptive discrete event model, map the processed signal to a D-channel discrete feature space and generate a D-dimensional discrete feature matrix; where D is the number of discrete channels. Step S2: Input the D-dimensional discrete feature matrix into the dual-path parallel processing architecture. The dual-path parallel processing architecture includes two paths. The first path uses the CFAR adaptive detection module to suppress strong clutter and generate the first response matrix. The second path uses a pseudo-attention mechanism to enhance weak targets and generate the second response matrix. Step S3: Perform spatial alignment and weighting on the first and second response matrices, and output the comprehensive response matrix through cross-domain confidence fusion. Step S4: Based on the global threshold, the comprehensive response matrix is ​​binarized using a fixed threshold to obtain the target detection result.

[0006] This invention first employs an adaptive discrete event model to truncate and discretize the echo amplitude using quantiles. Then, it utilizes a dual-path parallel processing architecture to process the echoes in parallel, adaptively suppressing strong clutter while simultaneously compensating for local gain in weak targets. Finally, the outputs from the two paths are weighted and fused, and a unified threshold decision is used to achieve target detection. The dual-path target detection algorithm employed in this invention effectively improves robustness and sensitivity, providing an interpretable and scalable engineering solution for near-ground detection by low-altitude unmanned aerial vehicles (UAVs).

[0007] Furthermore, the adaptive discrete event model includes a dynamic quantile calculation module, a nonlinear normalization module, and a discrete channel mapping module. The original distance image data obtained in step S1 first enters the dynamic quantile calculation module, which calculates the truncation threshold of the current frame in real time through ascending order and linear interpolation. The signal after threshold truncation enters the nonlinear normalization module for nonlinear mapping and normalization processing. The normalized data undergoes non-uniform quantization through the discrete channel mapping module to generate a D-dimensional discrete feature vector. The nonlinear normalization module adds multiple pulse interpolations before nonlinear mapping and compresses outliers exceeding the truncation threshold to the saturation region.

[0008] Furthermore, step S1 specifically includes the following steps; Step S11: Obtain raw distance image data; Step S12: Let the original range image data of the current frame be denoted as... Let 𝑢𝑝𝑝𝑒𝑟_𝑝𝑒𝑟𝑐𝑒𝑛𝑡𝑖l𝑒 represent the division ratio set by the clipping; Arrange 𝐫 in ascending order to obtain a new ordered vector. And calculate the quantile index 𝑝 according to the following formula: ; Among them, 𝑟1, 𝑟2, … , 𝑟 𝑁 Represents the amplitude sequence of the original range image in the current frame, 𝑟 (1) , 𝑟 (2) , … , 𝑟 (𝑁) This represents the ordered vector sequence after the amplitude sequence of the original range image of the current frame is arranged in ascending order, where N represents the number of pixels in the range image of the current frame; Let 𝑖=⌊𝑝⌋, 𝑗=𝑖+1, the quantile truncation uses a piecewise function with a conditionally triggered structure. It is represented and defined as: Where 𝑖 represents the floor index, and 𝑗 represents the previous index of 𝑖. t Indicates the current frame number. This represents the element at index i+1 in the ordered vector sequence. This represents the element at index j+1 in the ordered vector sequence. This indicates the preset threshold used to control the activation of the dynamic adjustment mechanism; The energy mutation factor is defined as: in, Let the distance image energy be the value in frame t. Let the range image energy be the energy of the (t-1)th frame. The intensity of environmental abrupt changes is quantified by the energy ratio of adjacent frame distance images. If the threshold is exceeded, threshold relaxation is triggered. Each amplitude Cut off as , And normalize to obtain the amplitude sample 𝑐 𝑖 If no additional logarithmic or piecewise mapping is used, this normalization preserves the linear flip structure and compresses any out-of-limit amplitude to [the specified value]. ; Step S13: Normalize the amplitude samples Further discretization to 𝐷 amplitude channels; using a multi-pulse interpolation mapping method, the following calculations are performed: ; ; in, This indicates the starting index of the normalized magnitude in the discrete channel. Indicates the channel interpolation period; In all cases where ℓ=0, 𝑝𝑒𝑟𝑖𝑜d 𝑖 ,2𝑝𝑒𝑟𝑖𝑜d 𝑖 ,⋯∩[0,𝐷−1] sets the discrete excitation; ℓ represents the index of the discrete excitation in the channel dimension; Step S14: Starting from channel index 0, use 𝑝𝑒𝑟𝑖𝑜d 𝑖 Pulse signals are placed intermittently along the channel axis with a step size, and a D-dimensional discrete feature matrix 𝐗(𝑁,𝐷) is generated.

[0009] Furthermore, in step S2, when the first path uses the CFAR adaptive detection module to suppress strong clutter and generate the first response matrix, the process from input matrix to output statistics includes the following steps: On the distance axis, a guard unit and a training unit are defined for each target unit. The guard unit is excluded, and the training unit is used to calculate the relative energy and generate the flip factor. The energy value of each row in the training unit is flipped and normalized so that the weight of each training unit is inversely proportional to its energy intensity. The flipped weights are combined with the main diagonal channel retention to form a lateral suppression matrix, which is then multiplied by the input matrix and output as the matrix.

[0010] Furthermore, the step S2, in which the first path uses the CFAR adaptive detection module to suppress strong clutter and generate the first response matrix, specifically includes the following steps: Step S211: Input the discretized D-dimensional discrete feature matrix 𝐗(𝑁,𝐷) into the CFAR adaptive detection module, and use training units and guard units to estimate the background energy for the 𝑖-th distance unit; Let train_sz represent the radius of the training unit and guard_sz represent the radius of the guard unit. Remove the guard units centered at 𝑖 with a radius of guard_sz, and include the training units centered at 𝑖 with a radius of train_sz in the background estimation. For a frame containing n distance units, the neighborhood... Defined as: ,and ; Step S212: Calculate the normalized amplitude of the current frame, directly taking the average value of each distance unit in the channel dimension as its energy representative range average value ... ; Calculate the flip factor flip(u) of each distance unit u using the following formula: ; Each distance cell is assigned a flip value that is inversely proportional to its energy intensity; The neighborhood of distance unit 𝑖 𝑖 All distance units 𝑢∈𝛺 𝑖 Normalize them so that their sum is 1; The flip-normalization result is written into the lateral weight matrix 𝐖1, which is represented as: ; Setting it to 0 within the protection unit indicates that the target will not be interfered with; Step S213: Construct the main diagonal weights 𝐖 𝑝 diagonal weights R 𝑝 Represented as: Main diagonal weight 𝐖 𝑝 The value is non-zero at the main diagonal; When the discretized D-dimensional discrete feature matrix When performing weighting, the diagonal weights are set to 𝐖 𝑝 Combined with the lateral weight matrix 𝐖1, a linear weighting of the d-th column 𝐗[:,d] is obtained: By merging all channels, we obtain the first response matrix of the continuous response. , represented as: .

[0011] Furthermore, the second path in step S2 employs a pseudo-attention mechanism to enhance the generation of the second response matrix for weak targets, specifically including the following steps: Step S221: Define a one-dimensional neighborhood window with radius k between the d-th distance cell and the d-th channel of the discretized D-dimensional discrete feature matrix 𝐗(𝑁,𝐷). , represented as: Where u represents the u-th distance unit, and N represents the number of pixels of distance image in the current frame; Step S22, in a one-dimensional neighborhood window Take the average value within the range : ; Define the reciprocal attention factor , represented as: in, This represents a local one-dimensional neighborhood window centered on the i-th cell. j Indicates the cell index within the window. This represents the amplitude value of the j-th distance unit in the d-th channel; Step S223: Use the reciprocal attention factor and one-dimensional neighborhood window Average value within the range Multiplying the magnitudes yields the response of the pseudo-attention module, represented as: The second response matrix is ​​obtained by summing up all channels. , .

[0012] Furthermore, step S3 specifically includes the following steps: The first response matrix in the real number field Second response matrix Linear weighted fusion is performed to obtain the comprehensive response matrix of the continuous response. , represented as: Among them, 𝛼 mix For the fusion coefficient, 𝛼 mix ∈[0,1], N represents the number of pixels in the current frame relative to the image, 𝑖 is the 𝑖-th distance unit, 𝑖∈{1,…,𝑁}, d is the d-th channel, d∈{1,…,𝐷}.

[0013] Furthermore, step S4 specifically includes the following steps: Use a fixed global threshold With the comprehensive response matrix The comparison is performed, with targets exceeding the threshold marked as 1 and others as 0, resulting in a discrete detection matrix. fused ∈{0,1} 𝑁×𝐷 .

[0014] In recent years, numerous studies have attempted to introduce machine learning into radar target detection to compensate for the shortcomings of traditional CFAR (Content-Based Ranging) detection. One approach is to utilize neural networks to automatically learn the feature distributions of clutter and targets, thereby dynamically adjusting detection decisions. Another approach is to apply Transformer self-attention structures to fields such as UAV target detection and trajectory prediction, constructing spatiotemporal attention models to achieve end-to-end detection and intent-aware tracking of micro-UAVs, improving detection performance in dynamic multi-target scenarios. In general, these innovative methods combining deep learning, whether it's the improved CFAR fusion of statistical features and neural networks, or the end-to-end detection network incorporating attention mechanisms, have significantly enhanced the adaptability of the CFAR algorithm to low-altitude UAVs in complex environments. Deep learning detection methods, by learning features from data, possess adaptive capabilities to complex backgrounds, enabling direct target identification even when clutter is not completely filtered out. End-to-end learning of the combined features of clutter and targets further enhances their detection performance. In terms of performance, it shows a significant improvement over traditional CFAR. However, deep learning methods also have the following limitations: First, they require a large amount of training data containing various environments and target scenarios; second, the model interpretability is insufficient, making it difficult to directly control the false alarm rate; third, the computational complexity is usually higher than that of traditional CFAR, and even with lightweight or attention mechanisms, hardware acceleration is still required to meet real-time requirements; fourth, the model is highly dependent on the characteristics of the environment and target, and if external conditions change abruptly, retraining or adjustments are often required to maintain detection performance.

[0015] Against this backdrop, this invention proposes a dual-path target detection method for complex ground reflection environments, used for detection... Detecting small targets in radar echoes. The proposed method aims to simultaneously inherit the advantages of the traditional CFAR algorithm in statistical threshold control. This invention leverages the advantages of deep learning attention mechanisms in extracting local salient features and achieves more robust detection of weak targets in complex backgrounds while significantly reducing reliance on massive training data. The core idea of ​​this invention is as follows: (1) Moving away from the frequency domain signal processing concept, the "local background estimation + adaptive threshold" concept of one-dimensional CFAR is expanded. Extending to the discrete channel dimension, a detection framework based on a two-dimensional event matrix is ​​formed. By performing a relaxation operation on the quantile truncation threshold for the amplitude of each distance unit and performing piecewise nonlinear interpolation mapping, it is possible to capture the non-uniform distribution of multipath reflections and clutter more flexibly in the two-dimensional plane. This method combines the robustness of CFAR-type threshold control with flexible response to sudden disturbances, and also inherits the interpretability of CFAR in controlling the false alarm rate.

[0016] (2) A dual-path parallel processing architecture is proposed. After discrete mapping of the radar echo, it enters two parallel paths. The first path robustly suppresses strong clutter through global statistical balancing and lateral suppression mechanisms; the second path compensates for weak targets with reciprocal gain through local energy statistics. Subsequently, the continuous response matrices output by the two paths are weighted and fused in the real domain. In this way, the first path can provide effective suppression of macro background clutter and stable false alarm rate control, while the second path B is specifically responsible for compensating and enhancing low-amplitude, easily missed targets. During the fusion stage, the contribution ratio of the two paths can be flexibly adjusted by the fusion coefficient, thereby achieving a balance between suppressing strong clutter and enhancing weak targets.

[0017] (3) The system can complete the comprehensive decision of all channels at once through a single global threshold after the fusion phase, retaining Instead of directly performing channel-level or pixel-by-pixel binarization, a complete continuous response matrix is ​​used. This means that once any channel exceeds the threshold, a target is considered detected. An early quantile truncation threshold relaxation operation is applied, proactively relaxing the suppression of large-amplitude signals while maintaining the advantages of local background estimation and adaptive thresholding. This avoids misclassifying potential targets or interference peaks as noise and weakening them. This "real-domain fusion first, then one-time threshold determination" approach not only simplifies the threshold setting process but also enhances the system's adaptability to complex backgrounds and weak targets.

[0018] In summary, the present invention has the following advantages compared with the prior art: (1) Existing radar target detection methods based on constant false alarm rate (CFAR) often struggle to simultaneously detect weak target energy and suppress strong background clutter in near-ground multipath reflection and dynamic clutter environments. On the one hand, weak target energy is easily masked by high-density clutter, leading to a decrease in detection rate; on the other hand, traditional CFAR methods that rely solely on statistical threshold adaptation are difficult to stably control the false alarm rate in clutter edge or sudden interference scenarios, and have poor adaptability to environmental changes, lacking the ability to compensate for local gain of target features. The present invention provides a dual-path target detection method for complex ground reflection environments, which can maintain the interpretability and low false alarm rate advantages of the CFAR method, and can also implement a detection algorithm for weak targets with local gain compensation, thereby improving the target detection performance and robustness in complex low-altitude near-ground environments.

[0019] (2) The dual-path parallel processing architecture of the present invention performs strong clutter adaptive suppression through the first path, and effectively suppresses high-energy background clutter by using dynamic weight update and lateral suppression, ensuring the stability of false alarms under complex sudden interference; the second path performs weak target gain compensation, and performs gain compensation for weak target points by calculating local energy reverse gain in real time, which significantly improves the detection capability of weak scatterers (such as pedestrians and low-reflection cones).

[0020] (3) The algorithm structure adopted in this invention is transparent, the fusion coefficient and global threshold can be adjusted online, it is compatible with backend multi-source data fusion or deep learning secondary recognition, it is easy to deploy in engineering and upgrade the system, and has good interpretability and scalability. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings: Figure 1 A flowchart of a specific embodiment of the present invention; Figure 2 This is a flowchart of an adaptive discrete event model processing according to a specific embodiment of the present invention; Figure 3 A bar chart showing the calculation time for 100 frames with different φ values; Figure 4 This is a flowchart illustrating the processing of the first path in a specific embodiment of the present invention; Figure 5 This is a schematic diagram of the second path in a specific embodiment of the present invention; Figure 6 This is a detection effect diagram on a single frame of data according to a specific embodiment of the present invention; Figure 7 This is a schematic diagram comparing the ROC curves of a specific embodiment of the present invention with those of a conventional method. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0023] Example: like Figure 1As shown, a dual-path target detection method for complex ground reflection environments includes the following steps: Step S1: Acquire raw range image data, perform quantile truncation and discretization processing on the raw range image data using an adaptive discrete event model, map the processed signal to a D-channel discrete feature space, and generate a D-dimensional discrete feature matrix; where D is the number of discrete channels; Step S2: Input the D-dimensional discrete feature matrix into a dual-path parallel processing architecture, wherein the dual-path parallel processing architecture includes two paths. The first path uses a CFAR adaptive detection module to suppress strong clutter and generate a first response matrix, and the second path uses a pseudo-attention mechanism to enhance weak targets and generate a second response matrix; Step S3: Perform spatial alignment and weighting processing on the first and second response matrices, and output a comprehensive response matrix through cross-domain confidence fusion; Step S4: Obtain the target detection result by performing fixed threshold binarization processing on the comprehensive response matrix based on a global threshold.

[0024] In the first stage of this embodiment, an adaptive discrete event model is used to perform quantile pruning and amplitude flipping on the radar range image data, mapping the signal to the discrete feature space of the 𝐷 channel. This preserves the target-sensitive components while achieving dimensionality reduction and amplitude differentiation. The second stage deploys a dual-path parallel processing architecture: the first path constructs a dynamic clutter baseline and suppresses strong scattering background through an adaptive lateral suppression mechanism; the second path uses a pseudo-attention mechanism to enhance the energy of weak targets and generates an attention mask in the local neighborhood to improve the signal-to-noise ratio. The third stage performs spatial alignment and weighting processing on the dual-path outputs through cross-domain confidence fusion. The fourth stage designs a global threshold detection method. Benefiting from the parallel processing and appropriate environmental awareness control in the previous stage, this stage further uses a unified threshold for one-time judgment of the fused response, maintaining good target detection stability.

[0025] The adaptive discrete event model in this embodiment includes a dynamic quantile calculation module, a nonlinear normalization module, and a discrete channel mapping module. The raw distance image data obtained in step S1 first enters the dynamic quantile calculation module. This module calculates the truncation threshold of the current frame in real time through ascending order and linear interpolation. The signal after threshold truncation enters the nonlinear normalization module for nonlinear mapping and normalization processing, where outliers exceeding the truncation threshold are compressed into the saturation region. Finally, while considering both detection performance and computational resources, weak amplitude information is further amplified while retaining strong target features. Multiple pulse interpolations are added before the nonlinear mapping, and the normalized data undergoes non-uniform quantization through the discrete channel mapping module to generate an 8-dimensional discrete feature vector. In target detection scenarios, the amplitude distribution of a one-dimensional distance image typically exhibits significant instability. To utilize amplitude and information in subsequent target detection, this embodiment uses an adaptive discrete event model to map continuous amplitudes to multiple discrete channels and performs adaptive discretization and discretization channel allocation for the one-dimensional distance data.

[0026] Step S1 in this embodiment specifically includes the following steps; Step S11: Obtain the raw distance image data.

[0027] Step S12: Let the original range image data of the current frame be denoted as... Let 𝑢𝑝𝑝𝑒𝑟_𝑝𝑒𝑟𝑐𝑒𝑛𝑡𝑖l𝑒 represent the division ratio set by the clipping; Arrange 𝐫 in ascending order to obtain a new ordered vector. And calculate the quantile index 𝑝 according to the following formula: ; Among them, 𝑟1, 𝑟2, … , 𝑟 𝑁 Represents the amplitude sequence of the original range image in the current frame, 𝑟 (1) , 𝑟 (2) , … , 𝑟 (𝑁) This represents the ordered vector sequence after the amplitude sequence of the original range image of the current frame is arranged in ascending order, and N represents the number of pixels (number of range units) of the range image of the current frame. To improve the robustness of quantile linear interpolation under abrupt disturbances, this embodiment integrates a dynamic coupling mechanism and a boundary protection strategy. Let 𝑖=⌊𝑝⌋, 𝑗=𝑖+1. The quantile truncation employs a conditionally triggered piecewise function, defined as: When n is not an integer, the above formula can achieve smooth positioning of the quantile by interpolating the sorting values ​​of the (1+1)th and (2+1)th quantiles.

[0028] Where 𝑖 represents the floor index, and 𝑗 represents the previous index of 𝑖. t Indicates the current frame number. This represents the element at index i+1 in the ordered vector sequence. This represents the element at index j+1 in the ordered vector sequence. This indicates the preset threshold used to control the activation of the dynamic adjustment mechanism; The energy mutation factor is defined as: in, Let the distance image energy be the value in frame t. Let the range image energy be the energy of the (t-1)th frame. The intensity of environmental abrupt changes is quantified by the energy ratio of adjacent frame distance images. If the threshold is exceeded, threshold relaxation will be triggered.

[0029] It quantifies the intensity of environmental abrupt changes by using the energy ratio of adjacent frame distance images. When 𝜂(𝑡)>𝛾, the system is determined to be in a strong interference state and dynamic adjustment is triggered: the original interpolation result is multiplied by 𝜂(𝑡) to achieve threshold relaxation, and then the output value range is constrained by min(1.0,⋅), which both suppresses signal distortion and avoids numerical overflow. Its physical essence is to sense environmental anomalies through energy abrupt change gradients and establish a nonlinear coupling relationship between interpolation parameters and interference intensity.

[0030] Each amplitude Cut off as , And normalize to obtain the amplitude sample 𝑐 𝑖 If no additional logarithmic or piecewise mapping is used, this normalization preserves the linear flip structure and compresses any out-of-limit amplitude to [the specified value]. This process, while preserving most of the relative amplitude information, effectively suppresses the impact of extreme high values ​​on the overall normalization, thereby achieving adaptive alignment of amplitude ranges between different frames.

[0031] Step S13: Normalize the amplitude samples Further discretization to 𝐷 amplitude channels; to further highlight the weak amplitude range, this embodiment uses a multi-pulse interpolation mapping method to calculate: ; ; in, This indicates the starting index of the normalized magnitude in the discrete channel. Indicates the channel interpolation period; In all cases where ℓ=0, 𝑝𝑒𝑟𝑖𝑜d 𝑖 ,2𝑝𝑒𝑟𝑖𝑜d 𝑖 Discrete excitations are set at [0, ∩[0, ∅−1]; ℓ represents the index of the discrete excitation in the channel dimension. Therefore, compared to setting only 1 at a single point ∅, the activation of weak targets can be significantly improved with periodic pulses in the column dimension. These column positions are periodically set to 1. That is, if ∅... 𝑖 A large value implies a weak amplitude, therefore 𝑝𝑒𝑟𝑖𝑜d 𝑖 Smaller amplitude cells produce denser pulse excitations along the channel axis; conversely, cells with stronger amplitudes produce sparser pulses along the column dimension to avoid over-consuming channel resources.

[0032] like Figure 2As shown, this embodiment starts from the original range image input, first calculates the inter-frame energy change factor 𝜂(𝑡) to sense environmental abrupt changes, and dynamically generates a truncation threshold clip_val(t) based on quantile interpolation; then, threshold truncation processing is performed, and the branch selects a normalization path of nonlinear amplification (secondary processing) or linear preservation (no additional modification) based on the judgment result 𝜂(𝑡)>𝛾; then, through multi-pulse interpolation mapping, the normalized amplitude 𝑐 is... 𝑖 Convert to channel activation cycle 𝑝𝑒𝑟𝑖𝑜d 𝑖 The process generates a D-dimensional discrete feature matrix; finally, it combines the three-dimensional parametric surface analysis to analyze the impact of the number of channels D on the detection probability and computational complexity, and outputs a heat map of the result matrix, thus realizing intelligent characterization of weak targets with high activation density and strong clutter with low channel occupancy.

[0033] Step S14: Starting from channel index 0, use 𝑝𝑒𝑟𝑖𝑜d 𝑖 Pulse signals are placed intermittently along the channel axis (from 0 to φ−1) with a step size (i.e., 1 is filled in the corresponding column position), generating a D-dimensional discrete feature matrix φ(φ, φ). Thus, the pulse distribution in the same row (corresponding to the same distance unit) of the entire φ(φ, φ) matrix reflects the difference in the magnitude of the original amplitude. The discrete excitation process in this embodiment is as follows: starting from channel index 0, ... Using a step size, 1 pulse values ​​are periodically inserted into the D-dimensional feature matrix to improve the visibility of low-amplitude targets.

[0034] When designing the R value for the feature space dimension, a balance needs to be found among multiple factors, including hardware computing power, background noise, and target motion characteristics. If the R value is too low, the amplitude quantization is too coarse, easily leading to blurred target edges and coupling between weak scatterers and clutter. If the R value is too high, the computational and storage requirements in the high-dimensional feature space increase significantly, and the captured noise microstructures can cause problems such as incorrect locking. To characterize this balance, the system's storage and computational overhead can be approximated using the following form: in, This indicates the amount of memory required for the algorithm to run, usually in units such as bytes or kilobytes (KB), which quantifies the hardware storage resource requirements of the detection algorithm. The computational complexity of an algorithm measures the time and computational resources required to execute it. Higher complexity means a greater demand on the processor's computing power, and a longer processing time for a single frame of data. The computation time of the algorithm is linearly proportional to the data size N×D. 𝑁 represents the length of the distance image, and 𝜶 is the average storage cost per element in a single channel. As 𝐷 increases, there is a non-linear increase or decrease in signal resolution and computational load, which should be weighed according to the specific requirements of the scenario.

[0035] As can be seen from Table 1, at a fixed false alarm rate of 𝑷 fa At a value of 0.01, different values ​​of 𝐷 will result in drastically different detection probabilities 𝑷. d For example, when 𝐷 = 15, the detection probability can reach approximately 0.868, while when 𝐷 is further increased (such as to 127 or 256), 𝑷 d Instead, it dropped below 0.1. This indicates that while an excessively high number of discrete channels can refine the amplitude characterization, it can also introduce excessive sparsity and noise interference, thereby limiting the overall detection performance.

[0036] like Figure 3 As shown, for the same 100 frames of data, the computation time for different R values ​​increases rapidly with R: from only a few seconds when R=3, it gradually climbs to about two hundred seconds when R=127, and finally exceeds 300 seconds when R=256. This indicates that when pursuing more refined amplitude characterization (larger R values), the computational cost of the algorithm increases significantly. Therefore, in the application of this embodiment, a trade-off should be made between target detection rate and real-time performance, and an appropriate R value should be selected to achieve better overall performance.

[0037] To better balance global statistical robustness and local gain adaptation, the λ matrix obtained after nonlinear interpolation is input into a parallel dual-path feature extraction. The first path employs a dynamic weighting strategy to achieve global adaptive suppression of background clutter, outputting a continuous response matrix λ. cfar The second path enhances the saliency of weak targets through local neighborhood energy compensation, generating a response matrix 𝑉 local The results from the two paths are weighted and fused to generate the comprehensive response 𝑉 final Finally, the detection results are output through binarization processing with a fixed threshold.

[0038] like Figure 4As shown, in step S2 of this embodiment, when the first path uses the CFAR adaptive detection module to suppress strong clutter and generate the first response matrix, the process from input matrix to output statistics includes the following steps: Guard units and training units are defined for each target unit on the distance axis. Guard units are excluded to avoid background contamination. Training units are used to calculate relative energy and generate flip factors. The energy values ​​of each row in the training units are flipped and normalized so that the weight of each training unit is inversely proportional to its energy intensity (the stronger the energy, the lower the weight; the weaker the energy, the higher the weight). The flipped weights are combined with the main diagonal channel retention (Diagonals = 1) to form a lateral suppression matrix, which is then multiplied by the input matrix, thereby simultaneously suppressing strong background energy and preserving the original amplitude of the target unit. Finally, this matrix is ​​output as the detection response of the first path, used for subsequent comparison with a threshold or fusion with other detection results. The entire process ensures effective suppression of high-energy regions and adaptive enhancement of weak target units without the need for cross-frame information or pre-training parameters.

[0039] In this embodiment, step S2, where the first path uses the CFAR adaptive detection module to suppress strong clutter and generate the first response matrix, specifically includes the following steps: Step S211: Input the discretized D-dimensional discrete feature matrix 𝐗(𝑁,𝐷) into the CFAR adaptive detection module, and use training units and guard units to estimate the background energy for the 𝑖-th distance unit; Let train_sz represent the radius of the training unit and guard_sz represent the radius of the guard unit. Guard units centered at x with a radius of guard_sz are removed, and training units centered at x with a radius of train_sz are included in the background estimation. This process generates a series of convolutional kernels or connection patterns to facilitate subsequent local statistical analysis of the data.

[0040] For a frame containing n distance units, the neighborhood... Defined as: ,and ; 𝛺 𝑖 That is, the corresponding reference area that can be used to estimate the background is excluded. This avoids the contamination of noise estimation by the "protected area" near the target, while allowing enough "training cells" to be collected at a distance to calculate the average background energy.

[0041] Step S212: To achieve adaptive behavior within the same frame, the dynamic weight update strategy adaptively adjusts the lateral weights according to the amplitude distribution of the frame, appropriately suppressing regions with excessively high energy and providing more compensation to regions with insufficient energy. The normalized amplitude of the current frame is statistically analyzed, and the average value of each distance unit in the channel dimension is directly taken as its energy representative range average value ... ; Calculate the flip factor flip(u) of each distance unit u using the following formula: ; Each distance cell is assigned a flip value that is inversely proportional to its energy intensity. Specifically, different flip mapping values ​​are assigned to high-energy and low-energy regions. The flip value is relatively low in high-energy regions and relatively high in low-energy regions, thus playing a "suppressing the strong and supporting the weak" adjustment role in local estimation.

[0042] The neighborhood of distance unit 𝑖 𝑖 All distance units 𝑢∈𝛺 𝑖 Normalize them so that their sum is 1; The flip-normalization result is written into the lateral weight matrix 𝐖1, which is represented as: ; Setting the weight to 0 within the protection unit indicates that the target will not be interfered with, thus forming the main idea of ​​"dynamic weight update".

[0043] Step S213: Construct the main diagonal weights 𝐖 𝑝 diagonal weights R 𝑝 Represented as: Main diagonal weight 𝐖 𝑝 The aim is to directly preserve the energy contribution of the current range cell itself, so as not to excessively weaken potential target signals during background statistics or suppression. Main diagonal weights 𝐖 𝑝 It can be viewed as a sparse matrix with non-zero values ​​on the main diagonal.

[0044] The dynamic weight matrix 𝐖1 is updated based on local statistics of the amplitude distribution of the current frame. These flipped values ​​are then normalized to obtain new lateral suppression weights, which are then written into 𝐖1. Since this process only depends on the amplitude distribution of the current frame and does not require cross-frame history or pre-trained parameters, it can "suppress strong targets and support weak targets" in real time in each frame, improving the preservation of weak targets and reducing the adverse effects of strong clutter regions on background estimation.

[0045] When the discretized D-dimensional discrete feature matrix When performing weighted average, 𝐖 𝑝 This ensures that the nth unit has a direct "through" contribution to the output statistics. The main diagonal weights are then set to... 𝑝 Combined with the lateral weight matrix 𝐖1, a linear weighting of column 𝐗[: , d] is performed to obtain: By merging all channels, we obtain the first response matrix of the continuous response. , represented as: .

[0046] Figure 5 This demonstrates the process by which a pseudo-attention mechanism statistically analyzes the local energy of each row *d* in the input matrix and generates an inverse gain. The dark area on the left of the figure represents a local window *d* along the row direction, where the values ​​of each row in the same column (d) are averaged to obtain a local mean. Subsequently, an anti-attention factor is derived based on these means and used for element-wise multiplication of the original amplitude, thereby highlighting useful scattering points and suppressing strong energy background in the output stage. Specifically, if the local mean is small (weak target energy), the inverse factor is larger, which can be preserved or enhanced in the output matrix; if the local mean is large (strong clutter energy), the factor is smaller, thereby suppressing the amplitude of high-energy rows. This process is repeated for each cell, selecting only ± *d* rows as the local statistical range along the row axis, and performing weighted calculations for each column (d), ultimately generating a pseudo-attention output matrix to highlight useful scattering points and suppress strong energy background.

[0047] Unlike common deep learning attention mechanisms that rely on a large number of learnable parameters, this embodiment proposes a lightweight "pseudo-attention" module for local energy compensation without relying on large-scale training data. Step S2 of this embodiment, where the second path uses a pseudo-attention mechanism to enhance the generation of the second response matrix for weak targets, specifically includes the following steps: Step S221: Define a one-dimensional neighborhood window with radius k between the d-th distance cell and the d-th channel of the discretized D-dimensional discrete feature matrix 𝐗(𝑁,𝐷). To capture the energy distribution at this location within a local area, it can be represented as: Here, u represents the u-th distance cell; the window here only includes the nearest few cells, which reflects the localized calculation method and avoids the high complexity caused by global dependencies.

[0048] Step S22, in a one-dimensional neighborhood window Take the average value within the range : ; like A smaller value indicates weaker energy at this location, potentially suggesting the presence of a weak target or low background noise; if A larger value indicates that the energy in the neighborhood of this location is already sufficiently strong. This allows for greater gains at weak energy locations and the maintenance or moderate reduction of strong energy locations during the inference phase.

[0049] Define the reciprocal attention factor , represented as: in, This represents a local one-dimensional neighborhood window centered on the i-th cell. j Indicates the cell index within the window. This represents the amplitude value of the j-th distance unit on the d-th channel.

[0050] when The smaller, The larger the local mean, the more "compensation" is applied to weak energy units; conversely, the larger the local mean, the more "compensation" is applied to weak energy units. The smaller the value, the milder the suppression of strong scattering points. In this embodiment, the lower the local average energy (potentially weak target), the better. The larger the value, the stronger the effect; the higher the energy (potential for strong clutter). The smaller the value, the more it has an inhibitory effect.

[0051] The adaptive factor requires no training parameters; it is calculated online as the frame of data is processed.

[0052] Step S223: Use the reciprocal attention factor and one-dimensional neighborhood window Average value within the range Multiplying the magnitudes yields the response of the pseudo-attention module, represented as: The second response matrix is ​​obtained by summing up all channels. , This method also calculates within the same frame, without requiring cross-frame inheritance or the ability to learn large matrices, making it both concise and capable of real-time compensation for weak signals.

[0053] After obtaining the consecutive outputs of the first and second paths, immediately performing binary decision-making often results in the loss of a significant amount of rich information, such as target strength and detection confidence. To address this issue, this embodiment retains these two real-domain output responses during the output stage, denoted as follows: , .

[0054] Since the two paths do not describe the same physical meaning—the first path focuses on suppressing strong clutter, while the second path focuses on gaining weak targets—retaining both continuous values ​​provides more flexibility in subsequent processing. To synthesize the outputs of the two paths, this embodiment performs a linear weighted fusion in the real domain to obtain the final continuous response matrix 𝑉 final .

[0055] Step S3 in this embodiment specifically includes the following steps: processing the first response matrix in the real number field. Second response matrix Linear weighted fusion is performed to obtain the comprehensive response matrix of the continuous response. , represented as: Among them, 𝛼 mix For the fusion coefficient, 𝛼 mix ∈[0,1], 𝑖 is the 𝑖-th distance unit, 𝑖∈{1,…,𝑁}, d is the d-th channel, d∈{1,…,𝐷}.

[0056] When mix When ≈ 1, the result is closer to the detection statistics of the first path; when 𝛼 mix When the value is approximately 0, the system becomes more reliant on the second path. This can be achieved by adjusting a single control parameter φ. mix This allows us to find a suitable balance between suppressing strong clutter and compensating for weak targets.

[0057] After completing the fusion results final After calculation, to meet the performance constraints of the radar system on detection probability and false alarm rate, a fixed global threshold is introduced. final This generates the final binarized detection result. Step S4 in this embodiment specifically includes the following steps: using a fixed global threshold... With the comprehensive response matrix The comparison is performed, with targets exceeding the threshold marked as 1 and others as 0, resulting in a discrete detection matrix. fused ∈{0,1} 𝑁×𝐷 .

[0058] Unlike traditional layer-by-layer thresholding, this embodiment adopts the approach of "preserving continuous response first and then setting a unified threshold". This approach not only allows for flexible adjustment of false alarm rate and false negative rate, but also preserves amplitude information as much as possible during visualization and subsequent expansion (such as multi-source fusion and deep learning secondary recognition), thereby improving the system's adaptability and interpretability.

[0059] It is important to emphasize that, although the threshold here is... finalWhile the amplitude is fixed, the dynamic cropping and α factor relaxation described earlier adaptively widen or tighten the upper bound of the amplitude during each frame's input stage. This preserves the relative differences between frames at different energy levels and prevents the amplitude in strong-energy frames from being compressed to 1.0 or the data in weak-energy frames from approaching 0. Thus, even when using a fixed threshold α... final It can also adapt well to inter-frame energy differences and maintain relatively stable detection performance in scenarios with strong clutter or sudden interference.

[0060] In multi-frame scenarios, to evaluate the detection performance of various fixed thresholds across different frames, it is necessary to perform optimal matching between the binary detection results and the ground truth target annotations. Specifically, if a detection result of 1 is obtained in any channel (d) in frame t, it is considered that a target has been detected in that distance cell t, forming a set. in, Let i represent the set of all range indices detected as "targets" in frame f, where i is the range cell index (a point in the radar ranging data) and d is the channel index, representing a certain dimension of the discrete channel. The value is the binarized result of the fused response matrix. If it exceeds the threshold T... final If a target is detected, the point is recorded as 1. Whenever any channel d of distance unit i is determined to be "target detected" (i.e., a value of 1) in the f-th frame, this i is recorded as the detection result.

[0061] Correspondingly, the set of real target annotations is denoted as To objectively assess the detection results, a distance matrix is ​​defined. in, This represents the distance difference (error) between the p-th real target and the q-th detection result. This represents the position (distance cell index) of the p-th real target. This represents the position of the q-th detected target. The distance error between all real targets and the detected targets is calculated, constructing a "cost matrix" for matching. This forms the basis for the subsequent application of the Hungarian Algorithm for optimal target allocation. If the error exceeds a certain tolerance threshold, It will be set to a very large value, representing "this cannot match the target".

[0062] If an element exceeds a pre-defined distance tolerance, it is assigned a maximum value to prevent unreasonable matching. Then, the Hungarian algorithm is applied to this matrix to find conflict-free target-detection pairs under the principle of global optimality, and the number of successful matches and unmatched detections is counted. After accumulating all frames, the total number of successful matches is denoted as 𝐵, the number of unmatched detections as FA (False Alarm), and the total number of real targets as 𝐺. The total number of pixels is determined by the product of the frame number 𝐹 and the distance image length 𝑁 per frame. The false alarm rate can then be defined. , With 𝑻 final and / or fusion coefficient 𝜶 mix The adjustment will change the number of detected targets and the number of false alarms, thus affecting (𝑷 fa , 𝑷 d Different trade-off curves are formed on the plane; by selecting an appropriate threshold, it is possible to flexibly switch between suppressing false alarms and capturing weak targets. Overall, this embodiment, in conjunction with previous work, can ensure visualization and scalability in the real number domain, effectively control the false alarm rate under a fixed threshold, and accurately evaluate the detection performance in multi-target or multi-frame scenarios through the Hungarian algorithm, providing a transparent and convenient performance control method for subsequent real-time applications and system optimization.

[0063] Figure 6 The image shown is a detection result of the proposed method in this embodiment on a single frame of data. The black solid line represents the original one-dimensional distance image. The first dashed line represents the final response after fusion, and the second dashed line represents the set threshold. final = 0.5, and the dots mark the locations of scattered points on the original signal that exceed the threshold. It can be observed that after multi-pulse interpolation combined with adaptive lateral suppression and local attention fusion, the final response is significantly higher than the background in the target region (such as at distances of approximately 40, 70, 100, and 200 from bin), while it is significantly suppressed in low-amplitude or background areas, forming a good detection decision.

[0064] To further verify the detection performance of the proposed method under different background noise and target dispersion conditions, a comparative evaluation was conducted between traditional CFAR (including OS-CFAR, CA-CFAR, and GO-CFAR) and the AD (Advanced Detector) proposed in this embodiment. Specifically, multiple sets of threshold parameters were scanned for each detector, and the detection rate and false alarm rate were recorded on multiple frames of echo data. The overall adaptability of the algorithm in low false alarm rate and high clutter environments was visually demonstrated using ROC curves.

[0065] Figure 7 shows a comparison of the ROC curves of this embodiment and the traditional method under the same test conditions, further comparing the performance of the algorithm and the traditional CFAR algorithm. It is clear from the figure that, at the same false alarm rate, this embodiment consistently maintains a higher detection probability than the traditional algorithm, and its ROC curve is closer to the ideal point in the upper left corner. This is especially true in the low false alarm rate region (𝑃 fa With a detection rate <0.01, the detection rate advantage of this embodiment is more significant, a characteristic that is of great importance for reducing the risk of false alarms in practical applications. Quantitative analysis shows that in [the specific value is missing in the original text]. fa When the threshold is 0.004, the detection rate of this embodiment is improved by approximately 20 percentage points compared to traditional methods. This result fully demonstrates the effectiveness of the proposed two-dimensional adaptive threshold and pseudo-attention mechanism in improving detection performance, especially in handling weak targets in complex ground reflection environments. In summary, the experimental results strongly confirm the superior performance and application potential of this embodiment in low-altitude complex environment target detection tasks.

[0066] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dual-path target detection method for complex ground reflection environments, characterized in that, The method comprises the following steps: Step S1, obtaining original range image data, performing quantile truncation and discretization processing on the original range image data through an adaptive discrete event model, mapping the processed signal to a D-channel discrete feature space, and generating a D-dimensional discrete feature matrix; wherein D is the number of discrete channels; Step S2, inputting the D-dimensional discrete feature matrix into a double-path parallel processing architecture, wherein the double-path parallel processing architecture comprises two paths, the first path adopts a CFAR adaptive detection module to suppress strong clutter to generate a first response matrix, and the second path adopts a pseudo attention mechanism to enhance weak targets to generate a second response matrix; Step S3, performing spatial alignment and weighting processing on the first response matrix and the second response matrix, and outputting a comprehensive response matrix through cross-domain confidence fusion; Step S4, performing fixed threshold binaryzation processing on the comprehensive response matrix based on a global threshold to obtain a target detection result.

2. The method of claim 1, wherein, The adaptive discrete event model comprises a dynamic quantile calculation module, a nonlinear normalization module, and a discrete channel mapping module. The original range image data obtained in step S1 is first input into the dynamic quantile calculation module. The module calculates the truncation threshold of the current frame in real time through ascending arrangement and linear interpolation. The signal truncated by the threshold is input into the nonlinear normalization module for nonlinear mapping and normalization processing. The normalized data is subjected to non-uniform quantization through the discrete channel mapping module to generate a D-dimensional discrete feature vector. The nonlinear normalization module adds multiple pulse interpolations before nonlinear mapping, and compresses the outliers exceeding the truncation threshold to the saturation region.

3. The method of claim 2, wherein, Step S1 specifically comprises the following steps: Step S11, obtaining original range image data; Step S12, let the current frame original distance image data be denoted as Let up p e r _ p e r c e n t i l e represent the clipping set proportion of quantiles. Sort r in ascending order to get a new ordered vector And calculate the quantile index p as follows: ； wherein, r1, r2, …, r 𝑁 denotes the amplitude sequence of the current frame original range image, r (1) , r (2) , …, r (𝑁) denotes the ordered vector sequence after arranging the amplitude sequence of the current frame original range image in ascending order, and N denotes the pixel number of the current frame range image; Let i = ⌊p⌋, j = i + 1, the quantile truncation is a piecewise function with conditional trigger structure denotes, is defined as: wherein, i represents a lower index, j represents a previous index of i, t represents a current frame number, represents an element with index i+1 in the ordered vector sequence, represents an element with index j+1 in the ordered vector sequence, represents a preset threshold value used to control starting of the dynamic adjustment mechanism; represents an energy mutation factor, defined as: wherein, is the tth frame distance image energy, is the t-1th frame distance image energy, The environmental abrupt change intensity is quantified by the ratio of adjacent frame distance image energies, if exceeds a set threshold, the truncation threshold relaxation is triggered. Each amplitude is truncated to , and normalized to an amplitude sample 𝑖 ; if there is no additional logarithmic or piecewise mapping, this normalization preserves the linear inversion structure and compresses any out-of-range amplitudes to ; Step S13, normalizing the amplitude samples Further discrete to D amplitude channels; using multi-pulse interpolation mapping method, calculate: ； ； wherein, denotes the starting index of the normalized amplitude in the discrete channel, denotes the channel interpolation period; Set discrete excitation at all ℓ = 0,perio 𝑖 ,2perio 𝑖 ,⋯∩[0,𝐷−1]; ℓ denotes the index of the discrete excitation in the channel dimension; Step S14, start from channel index 0, place pulse signals at intervals along the channel axis at a period of 𝑖 The step length is used to place pulse signals at intervals along the channel axis, and a D-dimensional discrete feature matrix X(N, D) is generated.

4. The method of claim 1, wherein, In the first path of step S2, when the CFAR adaptive detection module is used to suppress strong clutter to generate the first response matrix, the following steps are included from the input matrix to the output statistic: Define a protection unit and a training unit for each target unit on the distance axis, exclude the protection unit, and use the training unit to calculate the relative energy and generate a flip factor; Flip the energy value of each row in the training unit and perform normalization, so that the weight of each training unit is inversely proportional to its energy intensity; Combine the flipped weight with the main diagonal channel reservation to form a lateral suppression matrix, multiply the input matrix by the matrix, and output the matrix.

5. The method of claim 4, wherein, Step S2 specifically comprises the following steps: Step S211, inputting the D-dimensional discrete feature matrix X(N, D) processed through discretization into the CFAR adaptive detection module, and using the training unit and the protection unit to estimate the background energy for the ith distance unit; Let train_sz represent the radius length of the training unit, and guard_sz represent the radius length of the protection unit. Exclude the protection unit with guard_sz as the radius centered on i, and include the training unit with train_sz as the radius centered on i into the background estimation; For a frame with a total of N distance units, the neighborhood is defined as: , and ; Step S212, statistics are made on the normalized amplitudes of the current frame, and the average value of each distance unit u in the channel dimension is directly taken as the energy representative range average value norm_range_mean, and the calculation formula is: ； The flip factor flip(u) of each distance unit u is calculated as follows: ； A flip value inversely proportional to the energy intensity of each distance unit is assigned; normalize all distance units u e 𝛺 𝑖 in the field 𝛺 of distance unit i 𝑖 so that their sum is 1; The flip-normalization result is written into the lateral weight matrix W1, and the lateral weight matrix W1 is expressed as: ； It is set to 0 within the protection unit range, indicating no intervention to the target; Step S213, constructing the main diagonal weight matrix W 𝑝 , the main diagonal weight matrix W 𝑝 is expressed as: Main diagonal weight W 𝑝 Non-zero at the main diagonal. When weighting the D-dimensional discrete feature matrix 𝑝 The main diagonal weight matrix W is combined with the lateral weight matrix W1 to linearly weight the d-th column X[:,d] as follows:​ combining all the channels, i.e. the first response matrix of the continuous response is represented as: 。 6. The method of claim 1, wherein, The second path in the step S2 adopts a pseudo attention mechanism to enhance the weak target to generate a second response matrix, which specifically includes the following steps: Step S221, define a one-dimensional neighborhood window with a radius of k for the i-th distance unit of the D-dimensional discrete feature matrix X(N, D) and the d-th channel is expressed as: Wherein, u represents the u-th distance unit, and N represents the number of pixels of the distance image of the current frame; Step S22, averaging in one-dimensional neighborhood window :​ ； Defining an inverse attention factor is denoted as: wherein, denotes a local one-dimensional neighborhood window centered at the i-th cell, j denotes the index of a cell within the window, denotes the amplitude value of the j-th distance cell on the d-th channel; Step S223, using an inverse attention factor and a one-dimensional neighborhood window range-wise average multiplying the reduced amplitude, obtaining the response of the pseudo-attention module, denoted as: Summing all the channels gives the second response matrix , .

7. The method of claim 1, wherein, The step S3 specifically includes the following steps: linearly weighted fusion of the first response matrix and the second response matrix in the real number domain, to obtain a comprehensive response matrix of a continuous response , expressed as: wherein, a mix is a fusion coefficient, a mix ∈ [0, 1], N represents the number of pixels of the current frame, i is the i-th distance unit, i ∈ {1, …, N}, d is the d-th channel, d ∈ {1, …, D}.

8. The method according to any one of claims 1-7, wherein, The step S4 specifically includes the following steps: The step S4 specifically includes the following steps: with a fixed global threshold with the integrated response matrix Comparing, marking the targets exceeding the threshold as 1 and otherwise as 0, results in a discrete detection matrix S fused ∈{0,1} 𝑁×𝐷 .

Images