Machine learning-based 3D radar positioning anti-shielding interference method

Abstract

This invention discloses a machine learning-based 3D radar positioning anti-interference method, belonging to the field of radar signal processing and machine learning technology. The method includes: acquiring multiple frames of radar echo signals, extracting local point cloud fragments of the obscured target and the global point cloud of the unobscured reference target; inputting the local point cloud fragments and the global point cloud into a spatiotemporal attention graph neural network, calculating the implicit structural association weight between them through a cross-target attention module; transferring geometric features from the global point cloud to the local point cloud fragments based on this weight, generating a complete implicit neural field representation of the obscured target; and using this representation to perform raycasting decoding on the three-dimensional spatial position of the obscured target to obtain the anti-interference 3D radar positioning coordinates. This invention can effectively utilize the global structural information of the unobscured target to recover the complete geometry of the obscured target, improving the positioning accuracy and robustness of radar under obscuration interference.

Description

Technical Field

[0001] This invention relates to the fields of radar signal processing and machine learning technology, specifically to a machine learning-based 3D radar positioning anti-interference method. Background Technology

[0002] In 3D radar positioning applications, incomplete echo signals caused by target occlusion by non-metallic obstacles or dense scatterers are a common interference problem. Existing technologies typically employ methods such as incoherent superposition of multiple echo signals or Kalman filtering for anti-occlusion processing. These methods assume that the missing scattering points caused by occlusion are randomly and uniformly distributed, and fill in the missing information through time accumulation or motion prediction. However, in real-world occlusion scenarios, local point cloud fragments often exhibit irregular, non-uniform, and sparse distributions, and the missing parts are strongly correlated with the overall structure of the unoccluded target. Conventional statistical filtering and temporal smoothing methods are insufficient to effectively recover the details of the occluded parts from global geometric priors. The shortcomings of existing technologies include: a lack of explicit modeling of the structural similarity between the occluded target and the reference target, resulting in significant deviations between the completed point cloud and the true shape; and the inability of direct stitching or interpolation operations to utilize the spatial distribution patterns of high-dimensional geometric features, limiting positioning accuracy to the quality of local fragments. The problems to be solved by this invention are: how to use the global point cloud information of the unoccluded reference target to recover the complete three-dimensional geometric representation of the occluded target from local point cloud fragments through implicit structural association learning, and thereby achieve high-precision anti-occlusion localization; and how to incorporate the inter-frame consistency of time series into the feature transfer process to avoid localization jitter caused by single-frame noise. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of existing 3D radar in terms of low positioning accuracy under occlusion interference and inability to effectively utilize global structural information to recover local missing information. It provides a machine learning-based 3D radar positioning anti-occlusion interference method, which realizes geometric feature transfer and generates implicit neural field representation through cross-target attention mechanism, thereby achieving highly robust positioning of occluded targets.

[0004] The objective of this invention can be achieved through the following technical solutions:

[0005] This invention provides a machine learning-based 3D radar localization method to combat occlusion interference, aiming to solve the technical problem of accurately locating occluded targets in radar detection. The method extracts local point cloud fragments of the occluded target and the global point cloud of the unoccluded reference target from multi-frame radar echo signals. It then uses a cross-target attention module in a spatiotemporal attention graph neural network to calculate the implicit structural association weights between the two, thereby transferring geometric features from the global point cloud to the local point cloud fragments to generate a complete implicit neural field representation of the occluded target. Finally, ray casting decoding is used to obtain the occlusion-resistant 3D radar localization coordinates. Using this method, the complete geometric information of the unoccluded target can be effectively used to complete the missing structure of the occluded target, significantly improving the robustness and accuracy of radar localization in occluded scenarios.

[0006] As a technical solution of the present invention, the specific process of extracting local point cloud fragments of the obscured target and global point cloud of the unobscured reference target from the multi-frame radar echo signals is as follows: First, the multi-frame radar echo signals are processed by range Doppler to generate a range Doppler spectrum; then, a first signal region with continuous Doppler frequency shift characteristics and a second signal region with discrete Doppler frequency shift characteristics are segmented from the range Doppler spectrum; the first signal region is marked as the global point cloud of the unobscured reference target, and the subset of point clouds in the second signal region that meets the preset spatial continuity condition is marked as local point cloud fragments of the obscured target. Preferably, when segmenting the first signal region with continuous Doppler frequency shift characteristics and the second signal region with discrete Doppler frequency shift characteristics from the range Doppler spectrum, a density-based spatial clustering algorithm is used to cluster and segment the connected components in the range Doppler spectrum. In this way, it is possible to accurately distinguish between complete targets with stable motion trajectories and targets whose signals are discontinuous due to obstruction, providing reliable reference target point cloud data for subsequent feature transfer.

[0007] As a technical solution of the present invention, when calculating the implicit structural association weights between the local point cloud fragments and the global point cloud through the cross-target attention module in the spatiotemporal attention graph neural network, the specific steps include: using each point in the local point cloud fragments as a query node and each point in the global point cloud as a key node, calculating the feature similarity matrix between the query node and the key node; performing time-dimensional sliding window aggregation on the feature similarity matrix to obtain a time-aggregated similarity matrix; and performing spatial-dimensional graph convolution propagation on the time-aggregated similarity matrix to obtain the implicit structural association weights. Preferably, the process of performing time-dimensional sliding window aggregation on the feature similarity matrix includes: obtaining the current feature similarity matrix between the local point cloud fragments and the global point cloud in the current frame; obtaining the historical feature similarity matrix of each frame within the previous K historical time windows, where K is a preset time window length; and calculating the time-aggregated similarity matrix according to the current feature similarity matrix and the historical feature similarity matrix, using the following formula:

[0008]

[0009] in, Let be the time aggregation similarity matrix at time t. Let be the current feature similarity matrix at time t. Let be the historical feature similarity matrix at time ti. The preset time decay weight coefficient, with a value range of (0,1), is used to control the fusion ratio of current frame information and historical frame information. By introducing information fusion in the time dimension, the geometric continuity in the target's temporal motion can be utilized to suppress the correlation deviation caused by single-frame noise, making the cross-target attention weight more stable and reliable.

[0010] As a technical solution of the present invention, calculating the feature similarity matrix between the query node and the key node includes: extracting a first local density feature and a first normal vector feature for each point in the local point cloud fragment; extracting a second local density feature and a second normal vector feature for each point in the global point cloud; concatenating the first local density feature and the first normal vector feature into a query feature vector; concatenating the second local density feature and the second normal vector feature into a key feature vector; calculating the dot product between the query feature vector and the key feature vector, and using the result of the dot product as an element in the feature similarity matrix. Preferably, when extracting the first local density feature and the first normal vector feature, principal component analysis is used to perform covariance matrix decomposition on the neighborhood point set of each point in the local point cloud fragment. By extracting local geometric features and calculating the dot product similarity, the structural correlation between the point cloud fragment and the global point cloud can be effectively captured, laying the foundation for subsequent geometric feature transfer.

[0011] As a technical solution of the present invention, the process of transferring geometric features from the global point cloud to the local point cloud fragments according to the implicit structure association weights to generate a complete implicit neural field representation of the occluded target includes: weighting and summing the coordinate features of the global point cloud according to the implicit structure association weights to obtain a transferred feature field; inputting the coordinates of the local point cloud fragments into a first multilayer perceptron to obtain local implicit features; adding the transferred feature field and the local implicit features element-wise to obtain fused implicit features; and inputting the fused implicit features into a second multilayer perceptron to output a complete implicit neural field representation of the occluded target. Preferably, the process of weighting and summing the coordinate features of the global point cloud according to the implicit structure association weights is performed according to the following formula:

[0012]

[0013] in, For querying points in the local point cloud fragments The migration feature field value at the location, The total number of points in the global point cloud. For the query point With the j-th point in the global point cloud Implicit structural association weights between them The coordinate feature encoding vector of the j-th point in the global point cloud. Let be the three-dimensional spatial coordinates of the j-th point in the global point cloud. Through weighted migration, the geometric features of the missing region of the occluded target can be completed from the corresponding structure of the reference target, ensuring the integrity and continuity of the generated implicit neural field.

[0014] As a technical solution of the present invention, inputting the coordinates of the local point cloud fragments into a first multilayer perceptron to obtain local implicit features includes: performing position encoding on the coordinates of the local point cloud fragments to obtain a high-frequency position encoding vector; inputting the high-frequency position encoding vector into the first linear layer of the first multilayer perceptron to obtain a first intermediate feature; inputting the first intermediate feature into the first nonlinear activation layer of the first multilayer perceptron to obtain a first activation feature; and inputting the first activation feature into the second linear layer of the first multilayer perceptron to obtain local implicit features. Introducing high-frequency information through position encoding can improve the network's ability to model spatial details, allowing the local implicit features to more precisely describe the observed geometry.

[0015] As a technical solution of the present invention, the method of using the complete implicit neural field representation to perform ray projection decoding on the three-dimensional spatial position of the occluded target to obtain the anti-occlusion 3D radar positioning coordinates includes: taking the geometric center of the local point cloud fragment of the occluded target as the origin, emitting multiple virtual rays in multiple preset directions in space; collecting multiple sampling points at equal intervals along each virtual ray, inputting the spatial coordinates of each sampling point into the complete implicit neural field representation to obtain the occupancy probability value of each sampling point; integrating the occupancy probability value along each virtual ray to obtain the edge probability distribution of each virtual ray; and using the spatial coordinates corresponding to the peak position of the edge probability distribution as the anti-occupancy 3D radar positioning coordinates. Preferably, the multiple preset directions are obtained by uniformly sampling on a unit sphere using the Fibonacci grid algorithm to ensure the uniformity of the ray direction distribution and avoid positioning deviations due to missing directions. As a further preferred embodiment, the step of acquiring multiple sampling points at equal intervals along each virtual ray and inputting the spatial coordinates of each sampling point into the complete implicit neural field representation to obtain the occupancy probability value of each sampling point includes: determining the coordinates of the starting and ending sampling points of each virtual ray; calculating the total number of sampling points based on the coordinates of the starting and ending sampling points and a preset sampling step size; generating the spatial coordinates of each sampling point based on the total number of sampling points and the preset sampling step size; and using the spatial coordinates of each sampling point sequentially as the query coordinates of the complete implicit neural field representation to obtain the occupancy probability value output by the complete implicit neural field representation. Through ray projection decoding, the three-dimensional position of the occluded target can be accurately recovered from the continuous implicit field, effectively overcoming the positioning error caused by discrete point clouds in traditional methods.

[0016] As a technical solution of the present invention, after obtaining the anti-occlusion 3D radar positioning coordinates, the method further includes a step of performing temporal consistency self-supervised learning on the spatiotemporal attention graph neural network: obtaining a first timestamp of the anti-occlusion 3D radar positioning coordinates and a second timestamp of the local point cloud fragments of the occluded target; constructing a temporal consistency loss function based on the time difference between the first timestamp and the second timestamp; and updating the network parameters of the spatiotemporal attention graph neural network through backpropagation based on the temporal consistency loss function. Preferably, constructing the temporal consistency loss function includes: extracting a first displacement vector of the anti-occlusion 3D radar positioning coordinates between adjacent first timestamps; extracting a second displacement vector of the local point cloud fragments of the occluded target between adjacent second timestamps; calculating the Euclidean distance difference between the first displacement vector and the second displacement vector; and using the square of the Euclidean distance difference as the value of the temporal consistency loss function. By introducing temporal consistency constraints, the positioning results of consecutive frames output by the network can be made consistent with the motion pattern of the input point cloud fragments, further improving the temporal stability and anti-interference capability of the positioning.

[0017] The beneficial effects of this invention are:

[0018] This approach utilizes a cross-target attention module within a spatiotemporal attention graph neural network to calculate implicit structural association weights between local point cloud fragments and the global point cloud. This module uses each point in the local point cloud fragment as a query node and each point in the global point cloud as a key node. After calculating the feature similarity matrix, it employs temporal sliding window aggregation and spatial graph convolution propagation to obtain the association weights. Unlike conventional nearest neighbor matching or global feature stitching methods, this attention mechanism adaptively mines non-local structural correspondences between local fragments and the global point cloud. Even with extremely sparse local fragments, it can find the most relevant geometric substructures in the global point cloud based on the similarity of geometric features (such as local density and normal vectors). This design eliminates the reliance on pre-defined rigid transformations or local registration for geometric feature transfer, instead relying on data-driven implicit associations. This allows for accurate acquisition of the reference target's geometric information even when occluded areas lack clear matching cues, significantly improving the fidelity of the completion. Furthermore, the temporal sliding window aggregation performs weighted fusion of the similarity matrices of the current frame and historical frames, introducing temporal smoothing constraints to prevent weight fluctuations caused by single-frame noise or sudden occlusion changes, ensuring the stability of feature transfer and thus enabling the subsequently generated implicit neural field representation to have inter-frame consistency. The complete implicit neural field representation is used to perform ray casting decoding on the occluded target's 3D spatial location. This step emits virtual rays from the geometric center of the local point cloud fragments as the origin to multiple preset directions, samples points at equal intervals along the rays, and queries the occupancy probability in the implicit field. The edge probability distribution of each ray is obtained through integration, and the peak position is used as the positioning coordinate. Unlike traditional methods that directly cluster or calculate the centroid of the completed point cloud, the implicit neural field encodes the target's 3D shape as a continuous function, capable of representing geometric details of arbitrary fineness. Moreover, the ray casting decoding process is naturally robust to point cloud density and uneven distribution. Even if the completed implicit field is ambiguous in some areas due to the sparseness of the input fragments, the marginal probability distribution obtained by ray integration can still suppress local noise through the combined effect of multi-directional voting, ensuring that the peak position accurately reflects the true spatial boundary of the target. Furthermore, multiple preset directions are uniformly sampled using a Fibonacci grid algorithm to ensure that the ray covers the entire spatial orientation, avoiding positioning errors caused by directional bias. This decoding method transforms the positioning problem into probability density estimation on the ray, effectively decoupling shape restoration and position estimation. Even if there are local deformations in the completed shape, global integration can still provide a stable position output, thus achieving high-precision and robust 3D radar positioning in complex occlusion scenarios. Attached Figure Description

[0019] The invention will now be further described with reference to the accompanying drawings.

[0020] Figure 1 This is a flowchart illustrating the workflow of a machine learning-based 3D radar positioning anti-interference method according to the present invention.

[0021] Figure 2 This is a flowchart of the construction of the implicit structure association weight matrix of point clouds;

[0022] Figure 3 This is a flowchart of point cloud anti-occlusion 3D radar localization based on implicit neural fields. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] See Figure 1 This invention provides a machine learning-based 3D radar positioning anti-occlusion interference method. The method includes: acquiring multiple frames of radar echo signals; extracting local point cloud fragments of an occluded target and a global point cloud of an unoccluded reference target from the multiple frames of radar echo signals; inputting the local point cloud fragments and the global point cloud into a spatiotemporal attention graph neural network; calculating the implicit structural association weights between the local point cloud fragments and the global point cloud through a cross-target attention module in the spatiotemporal attention graph neural network; transferring geometric features from the global point cloud to the local point cloud fragments based on the implicit structural association weights to generate a complete implicit neural field representation of the occluded target; and using the complete implicit neural field representation to perform raycasting decoding on the three-dimensional spatial position of the occluded target to obtain the anti-occlusion 3D radar positioning coordinates.

[0025] Example 1: In a specific implementation, extracting local point cloud fragments of the obscured target and global point cloud of the unobscured reference target from the multi-frame radar echo signal includes the following steps.

[0026] Range Doppler processing is performed on the multi-frame radar echo signals to generate a range Doppler spectrum. The range Doppler processing includes pulse compression of each frame of radar echo signal to obtain range dimension information, and then Fourier transform along the slow time dimension to obtain Doppler dimension information, finally forming a two-dimensional spectrum with the range cell as the horizontal axis and the Doppler frequency as the vertical axis.

[0027] In some embodiments, a first signal region with continuous Doppler frequency shift characteristics and a second signal region with discrete Doppler frequency shift characteristics are segmented from the range-Doppler spectrum. The continuous Doppler frequency shift characteristics refer to the target exhibiting significant energy responses at multiple consecutive Doppler frequency units, manifested as connected regions extending continuously along the Doppler direction in the spectrum; the discrete Doppler frequency shift characteristics refer to the target exhibiting energy responses only at discontinuous Doppler frequency units, manifested as multiple isolated or discontinuous connected regions in the spectrum. During segmentation, a density-based spatial clustering algorithm (DBSCAN) is used to cluster the connected regions in the range-Doppler spectrum. Specifically, each pixel in the spectrum with energy exceeding a preset threshold is considered a point object. The two-dimensional coordinates of the pixel (range unit number, Doppler unit number) are used as the positional feature. The neighborhood radius parameter of the DBSCAN algorithm is set to 3 pixels, and the minimum number of neighboring points is 5. All high-energy pixels in the spectrum are clustered. Each cluster obtained after clustering corresponds to a connected region. In all clusters, if the Doppler coordinates of pixels within a cluster are distributed across multiple consecutive Doppler units without any breaks, then the cluster is marked as a first signal region with continuous Doppler frequency shift characteristics; if the Doppler coordinates of pixels within a cluster are distributed with breaks or exist only on isolated Doppler units, then the cluster is marked as a second signal region with discrete Doppler frequency shift characteristics.

[0028] In some embodiments, the first signal region is marked as a global point cloud of an unobstructed reference target. Specifically, the three-dimensional spatial coordinates corresponding to each pixel in the first signal region are extracted. The three-dimensional spatial coordinates are obtained by coordinate transformation from the radar ranging equation, azimuth angle, and elevation angle. The set of three-dimensional spatial coordinates of all pixels constitutes the global point cloud.

[0029] In some embodiments, a subset of point clouds in the second signal region that satisfies a preset spatial continuity condition is marked as a local point cloud fragment of the occluded target. The preset spatial continuity condition refers to the following: in the three-dimensional spatial coordinate point cloud extracted from the second signal region, the Euclidean distance between any two adjacent points is less than a preset distance threshold (e.g., 0.5 meters), and the point cloud distribution in three-dimensional space forms continuous segments rather than isolated points. Specifically, this includes: performing spatial neighborhood clustering on the three-dimensional spatial coordinates corresponding to all pixels in the second signal region, using the DBSCAN algorithm with an Euclidean distance threshold of 0.5 meters and a minimum clustering point count of 5. Each clustering result is a subset of point clouds that satisfies the spatial continuity condition. Each such subset of point clouds is marked as a local point cloud fragment of an occluded target. If there are multiple clustering results, they are respectively used as local point cloud fragments of different occluded targets.

[0030] Example 2: In a specific implementation, the implicit structural association weights between the local point cloud fragments and the global point cloud are calculated through the cross-target attention module in the spatiotemporal attention graph neural network, including the following steps.

[0031] See Figure 2 Using each point in the local point cloud fragment as a query node and each point in the global point cloud as a key node, a feature similarity matrix is ​​calculated between the query node and the key node. Specifically, the first local density feature and the first normal vector feature are extracted for each point in the local point cloud fragment; the second local density feature and the second normal vector feature are extracted for each point in the global point cloud. Both the first and second local density features are obtained by counting the number of points within the neighborhood radius of each point, with the neighborhood radius set to 0.3 meters. Both the first and second normal vector features are obtained by principal component analysis to decompose the covariance matrix of the neighborhood point set of each point. Specifically, for each point, the 20 nearest neighbors are selected to form a neighborhood point set, the covariance matrix of the neighborhood point set is calculated, the covariance matrix is ​​decomposed into eigenvalues, and the eigenvector corresponding to the smallest eigenvalue is taken as the normal vector feature of that point. The first local density feature and the first normal vector feature are concatenated into a 4-dimensional vector, which serves as the query feature vector for that point. The second local density feature and the second normal vector feature are concatenated into a 4-dimensional vector, which serves as the key feature vector for that point. Then, the dot product between each query feature vector and each key feature vector is calculated, and all the dot product results are arranged into a matrix. The rows of the matrix correspond to each query node in the local point cloud fragment, and the columns correspond to each key node in the global point cloud. This matrix is ​​the feature similarity matrix.

[0032] In some embodiments, the feature similarity matrix is ​​aggregated using a sliding window over a time dimension to obtain a time-aggregated similarity matrix. Specifically, the current feature similarity matrix between the local point cloud fragment and the global point cloud in the current frame is obtained, denoted as... Obtain the historical feature similarity matrix for each frame within the previous K-frame historical time window, where K is the preset time window length, set to 3. The historical feature similarity matrix... , , Corresponding to the first Time, Number Time, Number The feature similarity matrix at each time point. Based on the current feature similarity matrix and the historical feature similarity matrix, the time-aggregated similarity matrix is ​​calculated according to the following formula:

[0033]

[0034] in, Let be the time aggregation similarity matrix at time t. Let be the current feature similarity matrix at time t. Let be the historical feature similarity matrix at time ti. The preset time decay weight coefficient has a value range of (0,1) and is used to control the fusion ratio of current frame information and historical frame information. The specific value is set to 0.7, which is determined based on the matching relationship between the radar frame rate and the target's movement speed: when the radar frame rate is 20Hz and the average target movement speed is less than 10m / s, the change in point cloud structure between adjacent frames is small. Setting it to 0.7 allows the current frame information to dominate while retaining the smoothing effect of historical information. K is set to 3, which is an empirically selected value that can maintain the temporal context while avoiding the cumulative error introduced by too many historical frames.

[0035] In some embodiments, the temporal aggregation similarity matrix is ​​subjected to spatial-dimensional graph convolutional propagation to obtain implicit structural association weights. The graph convolutional propagation is implemented using a single-layer graph convolutional network, whose input is the temporal aggregation similarity matrix. The output is an implicit structure correlation weight matrix. Specifically, a graph structure is constructed where each node in the graph is a point in the local point cloud fragment connected to each point in the global point cloud, and the edges of the graph are connected by non-zero elements in the temporal aggregation similarity matrix. The graph convolution propagation operation is as follows: the temporal aggregation similarity matrix is ​​normalized to obtain a normalized adjacency matrix. Then, through a learnable graph convolution weight matrix Perform a linear transformation, then apply the ReLU activation function to obtain the implicit structure association weight matrix, calculated as follows:

[0036]

[0037] in, , It is a diagonal matrix of degree. The learnable graph convolution weight matrix The dimension is ,in This is the sum of the number of points in the local point cloud fragments and the number of points in the global point cloud. During training, Updated via backpropagation. The implicit structure-related weight matrix... In, elements This represents the association weight between the q-th query point in the local point cloud fragment and the j-th key node in the global point cloud.

[0038] Example 3: In a specific implementation, the process of migrating geometric features from the global point cloud to the local point cloud fragments based on the implicit structure association weights to generate a complete implicit neural field representation of the occluded target includes the following steps.

[0039] Based on the implicit structural association weights, the coordinate features of the global point cloud are weighted and summed to obtain the transfer feature field. Specifically, the three-dimensional spatial coordinate vector of each point in the global point cloud is obtained. Let the global point cloud contain a total of... There are points, and the three-dimensional spatial coordinates of each point are denoted as . ,in This is the index of a point in the global point cloud, with a value of Obtain the weight distribution vector of the implicit structure association weights corresponding to each query point in the local point cloud fragment. Assume there are a total of [number missing] weights in the local point cloud fragment. There are 1 query point, each denoted as _ . ,in For the query point index, the value is... For each query point From the implicit structure associated weight matrix Extract the first Row vector, which contains There are n elements, denoted as n. ,in Indicates query point In the global point cloud Points The implicit structural association weights between them. The formula for calculating the weighted sum of the coordinate features of the global point cloud is as follows:

[0040]

[0041] in, For querying points in the local point cloud fragments The migration feature field value at the location; The total number of points in the global point cloud; For the query point With the first in the global point cloud Points Implicit structural association weights between them; For the global point cloud, the first The coordinate feature encoding vector of each point is encoded by a learnable position encoding function. Three-dimensional spatial coordinates This is mapped to a 128-dimensional feature vector. For the global point cloud, the first The three-dimensional spatial coordinates of each point.

[0042] In some embodiments, the coordinates of the local point cloud fragments are input into a first multilayer perceptron to obtain local implicit features. The first multilayer perceptron includes an input layer, a hidden layer, and an output layer. Specifically, the coordinates of the local point cloud fragments are first position-encoded to obtain a high-frequency position-encoded vector. The position encoding uses sine and cosine functions to encode the three-dimensional coordinates of each point. The high-frequency position encoding vector is generated by mapping the vector to 10 trigonometric functions of different frequencies, resulting in a 60-dimensional high-frequency position encoding vector. This high-frequency position encoding vector is input to the first linear layer of the first multilayer perceptron, which has an input dimension of 60 and an output dimension of 256. A first intermediate feature is obtained through linear transformation. This first intermediate feature is then input to the first nonlinear activation layer of the first multilayer perceptron, which uses the ReLU activation function to obtain a first activation feature. Finally, this first activation feature is input to the second linear layer of the first multilayer perceptron, which has an input dimension of 256 and an output dimension of 128. A local implicit feature is obtained through linear transformation, with an output dimension of 128. This local implicit feature is denoted as... .

[0043] In some embodiments, the migration feature field is added element-wise to the local implicit features to obtain fused implicit features. The migration feature field With the aforementioned local implicit features All are 128-dimensional vectors, and the fused implicit features are obtained through element-wise addition. ,Right now .

[0044] In some embodiments, the fused implicit features are input to a second multilayer perceptron, which outputs a complete implicit neural field representation of the occluded target. The second multilayer perceptron comprises an input layer, two hidden layers, and an output layer. The input layer has a dimension of 128, the first hidden layer has a dimension of 256, the second hidden layer has a dimension of 128, and the output layer has a dimension of 1. All hidden layers use the ReLU activation function, and the output layer uses the Sigmoid activation function. The input to the second multilayer perceptron is the fused implicit features. The output is a scalar value representing the query point. The occupancy probability value at each query point. After performing the above operation on all query points, the occupancy probability field of the entire local point cloud fragment coverage area is obtained. This occupancy probability field is the complete implicit neural field representation of the occluded target.

[0045] Example 4: In a specific implementation, the three-dimensional spatial position of the occluded target is ray-projected and decoded using the complete implicit neural field representation to obtain the anti-occlusion 3D radar positioning coordinates, including the following steps.

[0046] See Figure 3 Multiple virtual rays are emitted in space in multiple preset directions, with the geometric center of the local point cloud fragment of the occluded target as the origin. The geometric center is obtained by averaging the three-dimensional spatial coordinates of all points in the local point cloud fragment. The multiple preset directions are obtained by uniformly sampling on a unit sphere using the Fibonacci grid algorithm. Specifically, the number of sampling points is set to 128, and 128 uniformly distributed three-dimensional direction vectors are generated on the unit sphere using the Fibonacci grid algorithm. Each direction vector is the emission direction of a virtual ray.

[0047] In some embodiments, multiple sampling points are collected at equal intervals along each virtual ray, and the spatial coordinates of each sampling point are input into the complete implicit neural field representation to obtain the occupancy probability value of each sampling point. Specifically, this includes determining the coordinates of the starting and ending sampling points for each virtual ray. The starting sampling point coordinates are set to the coordinates of the geometric center offset outwards by 0.1 meters along the ray direction; the ending sampling point coordinates are set to the coordinates of the geometric center offset outwards by 50 meters along the ray direction. The total number of sampling points is calculated based on the starting and ending sampling point coordinates and a preset sampling step size. The preset sampling step size is set to 0.2 meters, and the total number of sampling points is calculated using the formula... Calculation, where The distance from the terminating sampling point along the ray direction to the origin (i.e., 50 meters) is the distance. The distance from the starting sampling point along the ray direction to the origin (i.e., 0.1 meters). The preset sampling step size is 0.2 meters. Based on the total number of sampling points and the preset sampling step size, the spatial coordinates of each sampling point are generated. Starting from the initial sampling point, a sampling point is generated every 0.2 meters along the ray direction, for a total of [number missing] points. There are 3 sampling points, and the spatial coordinates of each sampling point are obtained through linear interpolation. The spatial coordinates of each sampling point are used sequentially as the query coordinates of the complete implicit neural field representation to obtain the occupancy probability value output by the complete implicit neural field representation. The complete implicit neural field representation accepts a three-dimensional spatial coordinate as input and outputs an occupancy probability value between 0 and 1, representing the probability that the coordinate position is occupied by the target.

[0048] In some embodiments, the occupancy probability value is integrated along each virtual ray to obtain the marginal probability distribution of each virtual ray. The integration is achieved using discrete summation: for each virtual ray, the occupancy probability values ​​of all sampling points are accumulated sequentially from the start point to the end point, and the accumulated result is used as the cumulative response value of that ray. Then, the cumulative response values ​​of all rays are normalized so that the sum of the cumulative response values ​​of all rays is 1, thus obtaining the marginal probability distribution of each virtual ray. The marginal probability distribution reflects the probability of the target's presence in different directions.

[0049] In some embodiments, the spatial coordinates corresponding to the peak position of the edge probability distribution are used as the anti-occlusion 3D radar positioning coordinates. Specifically, in the edge probability distribution, a virtual ray with the largest cumulative response value is found; the three-dimensional spatial coordinates of the sampling point on the virtual ray corresponding to the largest cumulative response value (i.e., the position where the cumulative response value reaches its maximum value on the ray) are extracted and used as the anti-occlusion 3D radar positioning coordinates of the occluded target. If multiple rays share the same maximum cumulative response value, the average value of the coordinates of the sampling points corresponding to these rays is taken as the final positioning coordinates.

[0050] Example 5: In a specific implementation, after obtaining the anti-interference 3D radar positioning coordinates, the machine learning-based 3D radar positioning anti-interference method further includes the following steps.

[0051] The system acquires a first timestamp of the anti-occlusion 3D radar positioning coordinates and a second timestamp of the local point cloud fragments of the occluded target. The first timestamp is the acquisition time of the radar frame corresponding to the output of the anti-occlusion 3D radar positioning coordinates, and the second timestamp is the acquisition time of the radar frame corresponding to the extraction of the local point cloud fragments. Both the first and second timestamps are in milliseconds and are recorded by the radar system's time synchronization module.

[0052] In some embodiments, a temporal consistency loss function is constructed based on the time difference between the first timestamp and the second timestamp. Specifically, this includes: extracting the first displacement vector of the anti-occlusion 3D radar positioning coordinates between adjacent first timestamps; and extracting the second displacement vector of the local point cloud fragments of the occluded target between adjacent second timestamps. The adjacent first timestamps refer to the timestamps in two consecutive radar signal frames corresponding to the anti-occlusion 3D radar positioning coordinates, and the adjacent second timestamps refer to the timestamps in two consecutive radar signal frames corresponding to the local point cloud fragments. Let the first... The anti-occlusion 3D radar positioning coordinates corresponding to each first timestamp are , No. The anti-occlusion 3D radar positioning coordinates corresponding to each first timestamp are Then the first displacement vector The calculation formula is: Let the first... The geometric center coordinates of the local point cloud fragment corresponding to the second timestamp are: , No. The geometric center coordinates of the local point cloud fragment corresponding to the second timestamp are: Then the second displacement vector The calculation formula is: The geometric center coordinates are obtained by calculating the average of the three-dimensional spatial coordinates of all points in the local point cloud fragment.

[0053] In some embodiments, the Euclidean distance difference between the first displacement vector and the second displacement vector is calculated. The formula for calculating the Euclidean distance difference is:

[0054]

[0055] in, For the first The Euclidean distance difference between frames, Let L2 norm, or Euclidean distance, be the vector. The square of the Euclidean distance difference is used as the value of the temporal consistency loss function. The temporal consistency loss function... Defined as:

[0056]

[0057] in, This represents the total number of radar frames used in the training. From 2 to The value is obtained by calculating the square of the Euclidean distance difference for each adjacent frame pair and then taking the average value.

[0058] In some embodiments, the network parameters of the spatiotemporal attention graph neural network are updated via backpropagation according to the temporal consistency loss function. The spatiotemporal attention graph neural network is a pre-constructed neural network structure including a cross-target attention module, a temporal sliding window aggregation module, and a graph convolutional propagation module. During training, the temporal consistency loss function is weighted and combined with the task loss function (such as occupancy probability cross-entropy loss) previously used to generate the complete implicit neural field representation; the resulting total loss function is... ,in The hyperparameter is set to 0.1 to balance the importance of the two losses. The gradient of the total loss function with respect to all trainable parameters in the spatiotemporal attention graph neural network is calculated using the backpropagation algorithm, and the parameters are updated using the Adam optimizer. The learner rate of the optimizer is set to 0.001, the batch size to 16, and the training iterations to 100 epochs.

[0059] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A machine learning-based 3D radar positioning method to resist obstruction and interference, characterized in that, The method includes: Acquire multiple frames of radar echo signals, and extract local point cloud fragments of the obscured target and global point cloud of the unobscured reference target from the multiple frames of radar echo signals; The local point cloud fragments and the global point cloud are input into a spatiotemporal attention graph neural network, and the implicit structural association weights between the local point cloud fragments and the global point cloud are calculated through the cross-target attention module in the spatiotemporal attention graph neural network. Based on the implicit structural association weights, geometric features are transferred from the global point cloud to the local point cloud fragments to generate a complete implicit neural field representation of the occluded target; The three-dimensional spatial position of the occluded target is decoded by ray projection using the complete implicit neural field representation to obtain the anti-occlusion 3D radar positioning coordinates.

2. The 3D radar positioning anti-interference method based on machine learning according to claim 1, characterized in that, Extracting local point cloud fragments of the obscured target and global point cloud of the unobscured reference target from the multi-frame radar echo signals includes: Range Doppler processing is performed on the multi-frame radar echo signals to generate a range Doppler spectrum. A first signal region with continuous Doppler frequency shift characteristics and a second signal region with discrete Doppler frequency shift characteristics are segmented from the range Doppler spectrum. The first signal region is marked as the global point cloud of the unobstructed reference target; The subset of point clouds in the second signal region that meets the preset spatial continuity condition is marked as local point cloud fragments of the occluded target.

3. The 3D radar positioning anti-interference method based on machine learning according to claim 2, characterized in that, When segmenting a first signal region with continuous Doppler frequency shift characteristics and a second signal region with discrete Doppler frequency shift characteristics from the range-Doppler spectrum, a density-based spatial clustering algorithm is used to cluster and segment the connected components in the range-Doppler spectrum.

4. The 3D radar positioning anti-interference method based on machine learning according to claim 1, characterized in that, The implicit structural association weights between the local point cloud fragments and the global point cloud are calculated using the cross-target attention module in the spatiotemporal attention graph neural network, including: Using each point in the local point cloud fragment as a query node and each point in the global point cloud as a key node, calculate the feature similarity matrix between the query node and the key node; The feature similarity matrix is ​​aggregated using a sliding window over a time dimension to obtain a time-aggregated similarity matrix. The spatial dimension of the temporal aggregation similarity matrix is ​​propagated by graph convolution to obtain implicit structural association weights.

5. The 3D radar positioning anti-interference method based on machine learning according to claim 4, characterized in that, Using each point in the local point cloud fragment as a query node and each point in the global point cloud as a key node, the feature similarity matrix between the query node and the key node is calculated as follows: Extract the first local density feature and the first normal vector feature of each point in the local point cloud fragment; Extract the second local density feature and the second normal vector feature of each point in the global point cloud; The first local density feature and the first normal vector feature are concatenated to form the query feature vector; The second local density feature and the second normal vector feature are concatenated to form a key feature vector; Calculate the dot product between the query feature vector and the key feature vector, and use the result of the dot product as an element in the feature similarity matrix.

6. The 3D radar positioning anti-interference method based on machine learning according to claim 5, characterized in that, When extracting the first local density feature and the first normal vector feature, principal component analysis is used to decompose the covariance matrix of the neighborhood point set of each point in the local point cloud fragment.

7. The 3D radar positioning anti-interference method based on machine learning according to claim 1, characterized in that, Based on the implicit structural association weights, geometric features are transferred from the global point cloud to the local point cloud fragments to generate a complete implicit neural field representation of the occluded target, including: Based on the implicit structural association weights, the coordinate features of the global point cloud are weighted and summed to obtain the migration feature field; The coordinates of the local point cloud fragments are input into the first multilayer perceptron to obtain local implicit features; The transferred feature field is added element-wise to the local implicit features to obtain the fused implicit features; The fused implicit features are input into a second multilayer perceptron, which outputs a complete implicit neural field representation of the occluded target.

8. The 3D radar positioning anti-interference method based on machine learning according to claim 7, characterized in that, The coordinates of the local point cloud fragments are input into the first multilayer perceptron to obtain local implicit features, including: The coordinates of the local point cloud fragments are position-encoded to obtain a high-frequency position-encoded vector; The high-frequency position encoding vector is input into the first linear layer of the first multilayer perceptron to obtain the first intermediate feature; The first intermediate feature is input into the first nonlinear activation layer of the first multilayer perceptron to obtain the first activation feature; The first activation feature is input into the second linear layer of the first multilayer perceptron to obtain local implicit features.

9. The 3D radar positioning anti-interference method based on machine learning according to claim 1, characterized in that, Using the complete implicit neural field representation to perform ray casting decoding on the three-dimensional spatial position of the occluded target, the resulting anti-occlusion 3D radar positioning coordinates include: Using the geometric center of the local point cloud fragment of the occluded target as the origin, multiple virtual rays are emitted in multiple preset directions in space; Multiple sampling points are collected at equal intervals along each virtual ray, and the spatial coordinates of each sampling point are input into the complete implicit neural field representation to obtain the occupancy probability value of each sampling point; Integrating the occupancy probability value along each virtual ray yields the marginal probability distribution of each virtual ray; The spatial coordinates corresponding to the peak position of the edge probability distribution are used as the 3D radar positioning coordinates after anti-occlusion.

10. The 3D radar positioning anti-interference method based on machine learning according to claim 9, characterized in that, The multiple preset directions are obtained by uniformly sampling on a unit sphere using the Fibonacci grid algorithm.

Images