Target identification method and device of millimeter wave radar and millimeter wave radar

Abstract

The invention provides a target identification method and device for a millimeter-wave radar and the millimeter-wave radar, and the method comprises the steps: carrying out the distance calculation of detection points of the millimeter-wave radar, obtaining a detection point graph, and determining a point cloud graph structure according to the detection point graph; performing multi-layer perceptron mapping on the point cloud image structure to obtain non-context coding features, and converting the non-context coding features into high-dimensional vector representation to obtain feature space data; performing graph convolution processing on the feature space data to obtain a context point code, and generating a target candidate frame according to the context point code; performing confidence evaluation on the target candidate box to obtain target confidence data, and calculating motion features and type probabilities of the candidate targets according to the target confidence data; and performing clustering and non-maximum suppression processing on the candidate targets according to the motion features and the type probability to obtain a target recognition result.

Description

Technical Field

[0001] This application relates to the field of radar technology, and particularly to a method and device for target recognition of millimeter-wave radar and a millimeter-wave radar. Background Art

[0002] Currently, millimeter-wave radar technology has broad application prospects in the field of target recognition, especially in industries such as autonomous driving, security monitoring, and industrial automation. However, when using neural networks to process millimeter-wave radar data, challenges such as slow speed and insufficient accuracy still exist. Neural network models usually require a large amount of computing resources to process complex radar signals, which leads to high latency. For application scenarios with high real-time requirements (such as autonomous driving), this latency may directly affect the response speed and safety of the system. In addition, training an efficient neural network model often depends on a large-scale well-annotated dataset, and the acquisition cost of high-quality data is high and time-consuming, which limits the further improvement of model performance. Secondly, in terms of accuracy, there are certain noise interferences and multipath effects in millimeter-wave radar data itself, and these factors will affect the target recognition results. Although neural networks can extract useful features from raw data through learning, it is still difficult to achieve ideal recognition accuracy when facing subtle differences or low signal-to-noise ratio situations in complex environments. Especially when multiple targets are close to each other or there is occlusion, traditional methods are prone to misjudgment or missed detection. Summary of the Invention

[0003] This application provides a method and device for target recognition of millimeter-wave radar and a millimeter-wave radar, which are used to improve the accuracy of target recognition.

[0004] In a first aspect, an embodiment of this application provides a method for target recognition of millimeter-wave radar, and the method includes: Calculate the distance of the detection points of the millimeter-wave radar to obtain a detection point map, and determine the point cloud map structure according to the detection point map; Perform multi-layer perceptron mapping on the point cloud map structure to obtain non-context encoded features, and convert the non-context encoded features into high-dimensional vector representations to obtain feature space data; Perform graph convolution processing on the feature space data to obtain context point encodings, and generate target candidate boxes according to the context point encodings; Perform confidence evaluation on the target candidate boxes to obtain target confidence data, and calculate the motion features and type probabilities of candidate targets according to the target confidence data; According to the motion features and the type probabilities, perform clustering and non-maximum suppression processing on the candidate targets to obtain the target recognition result.

[0005] Second aspect, an embodiment of the present application provides a target recognition device for a millimeter-wave radar, the device comprising: A cloud map calculation module, configured to perform distance calculation on detection points of the millimeter-wave radar to obtain a detection point map, and determine a point cloud map structure according to the detection point map; A feature mapping module, configured to perform multi-layer perceptron mapping on the point cloud map structure to obtain non-context encoded features, convert the non-context encoded features into high-dimensional vector representations, and obtain feature space data; A target locking module, configured to perform graph convolution processing on the feature space data to obtain context point encodings, and generate target candidate boxes according to the context point encodings; A data analysis module, configured to perform confidence evaluation on the target candidate boxes to obtain target confidence data, and calculate the motion features and type probabilities of candidate targets according to the target confidence data; A result output module, configured to perform clustering and non-maximum suppression processing on the candidate targets according to the motion features and the type probabilities to obtain a target recognition result.

[0006] Third aspect, an embodiment of the present application provides a millimeter-wave radar, the millimeter-wave radar comprising a memory and a processor; The memory is used to store a computer program; The processor is configured to execute the computer program and, when executing the computer program, implement the target recognition method for the millimeter-wave radar as described in any one of the embodiments of the present application.

[0007] Fourth aspect, an embodiment of the present application provides a computer-readable storage medium, the computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the processor is caused to implement the target recognition method for the millimeter-wave radar as described in any one of the embodiments of the present application.

[0008] The embodiments of the present application provide a method for target recognition of a millimeter-wave radar. The method includes: calculating the distance of detection points of the millimeter-wave radar to obtain a detection point map, and determining the point cloud map structure according to the detection point map; performing a multi-layer perceptron mapping on the point cloud map structure to obtain a non-context encoded feature, converting the non-context encoded feature into a high-dimensional vector representation to obtain feature space data; performing graph convolution processing on the feature space data to obtain a context point encoding, and generating a target candidate box according to the context point encoding; performing a confidence evaluation on the target candidate box to obtain target confidence data, and calculating the motion feature and type probability of the candidate target according to the target confidence data; performing clustering and non-maximum suppression processing on the candidate target according to the motion feature and type probability to obtain a target recognition result. Through the above method, by calculating the distance of detection points and generating a point cloud map, combining multi-layer perceptron mapping and graph convolution processing, efficient feature extraction and context information fusion are realized. By performing confidence evaluation and motion feature analysis on the target candidate box, further combining clustering analysis technology and non-maximum suppression technology, the accuracy and robustness of target recognition are effectively improved. This method not only improves the target recognition ability in complex environments, but also enhances the recognition accuracy of targets of different scales and types, significantly improving the overall performance of the millimeter-wave radar system, and is applicable to various application scenarios such as autonomous driving and security monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0010] Figure 1 It is a schematic flowchart of a method for target recognition of a millimeter-wave radar provided by an embodiment of the present application; Figure 2 It is a schematic block diagram of a device for target recognition of a millimeter-wave radar provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0011] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the protection scope of the present invention.

[0012] The flowcharts shown in the accompanying drawings are merely illustrative examples and do not necessarily include all the content and operations / steps, nor do they necessarily need to be executed in the described order. For example, some operations / steps can be decomposed, combined, or partially merged, so the actual execution order may change according to the actual situation.

[0013] It should also be understood that the terms used in the specification of this application are merely for the purpose of describing specific embodiments and are not intended to limit this application. As used in the specification of this application and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an", and "the" are intended to include the plural forms.

[0014] It should be further understood that the term "and / or" used in the specification of this application and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0015] A millimeter-wave radar is a radar system that operates in the millimeter-wave frequency band (usually between 30 GHz and 300 GHz). The electromagnetic waves in this frequency band have a relatively short wavelength, generally between 1 mm and 10 mm, hence the name "millimeter-wave". Compared with traditional microwave radars, millimeter-wave radars can integrate antenna arrays in a smaller space, providing higher resolution and accuracy. At the same time, due to their shorter wavelength, they can better penetrate adverse weather conditions such as rain and fog to achieve effective detection of target objects.

[0016] Please refer to Figure 1 , Figure 1 which is a schematic flowchart of a method for target recognition of a millimeter-wave radar provided by an embodiment of this application. As Figure 1 shown, the specific steps of the method for target recognition of this millimeter-wave radar include: S101 - S105.

[0017] S101. Calculate the distances of the detection points of the millimeter-wave radar to obtain a detection point map, and determine the point cloud map structure based on the detection point map.

[0018] Exemplarily, the detection points of the millimeter-wave radar refer to the points with certain position and attribute information detected by the radar when detecting the environment. These points are generated by the millimeter-wave signals emitted by the radar and the echo signals received after being reflected by the object surface. Each detection point usually contains the following information: Spatial coordinates: The position of the detection point in three-dimensional space, usually represented by (x, y, z) coordinates. These coordinate information helps to describe the specific position of the object in the environment.

[0019] Reflection intensity: Also known as the radar cross section (RCS), which represents the intensity of the radar signal reflected from the object surface.

[0020] Speed information: Some millimeter-wave radars can also measure the relative speed of an object through the Doppler effect.

[0021] Timestamp: Used to mark the time when a detection point is collected.

[0022] By obtaining data of detection points in the environment through a millimeter-wave radar, this data includes the spatial coordinates of each detection point and possible reflected intensity information. Based on the above information, distance calculations are performed on the detection points of the millimeter-wave radar to obtain a detection point map. The point cloud map structure is determined according to the detection point map. Specifically, the distance between each detection point is calculated through the Euclidean distance formula to obtain distance information. These distance information are used to construct a point cloud map structure, where the detection points serve as the vertices of the point cloud map structure and the distance serves as the weight of the edge. This point cloud map structure represents the spatial relationship of the detection points. By analyzing the connectivity of the detection point map, the overall structural characteristics of the point cloud map can be further extracted, laying a foundation for feature extraction and analysis in subsequent steps.

[0023] S102. Perform a multi-layer perceptron mapping on the point cloud map structure to obtain non-contextual encoded features, and convert the non-contextual encoded features into a high-dimensional vector representation to obtain feature space data.

[0024] Exemplarily, a multi-layer perceptron (MLP) is used to process the point cloud map structure. MLP is a feedforward neural network that can map the input low-dimensional features to a high-dimensional space through a series of linear transformations and activation functions. Through this mapping, non-contextual encoded features of each detection point can be obtained. These non-contextual encoded features do not depend on the relationship between points but are based on the attributes of the points themselves, including: the spatial coordinates of the point, the reflected intensity of the point, the initial attribute label of the point, and the basic geometric features of the point. Next, these non-contextual feature encodings are converted into a high-dimensional vector representation, which helps to enhance the information expression ability and provides richer feature space data for subsequent graph convolution operations.

[0025] Exemplarily, for each point P in the point cloud map i , its feature vector is defined as: F i = [x i , y i , z i , r i , where x i , y i and z i are spatial coordinates, and r i is the reflected intensity of the point.

[0026] In the multi-layer perceptron mapping transformation, a multi-layer neural network structure is adopted, and the mapping function is defined as: ; where: W1, W2, and W3 are weight matrices, b1, b2, and b3 are bias terms, ReLU is the activation function, Sigmoid is used for non-linear mapping, and the feature vector is normalized to F ” i =(F i '-μ) / σ, where μ is the mean and σ is the standard deviation. Through dimensionality reduction and information compression, a compact feature representation is obtained: V i =PCA(F i '').

[0027] S103. Perform graph convolution processing on the feature space data to obtain context point encodings, and generate target candidate boxes based on the context point encodings.

[0028] Exemplarily, a graph convolutional network (GCN) is used to process the feature space data. The graph convolutional network can effectively capture the local structural information between points through convolutional operations on the graph structure. By performing graph convolution processing on the feature space data, context point encodings can be extracted. These encodings not only contain the features of individual points but also incorporate the information of their neighboring points, thus more comprehensively representing the local features of the target. Subsequently, target candidate boxes are generated based on the context point encodings. These boxes represent the regions where the target may exist, providing a basis for subsequent target recognition and classification.

[0029] S104. Perform confidence evaluation on the target candidate boxes to obtain target confidence data, and calculate the motion features and type probabilities of the candidate targets based on the target confidence data.

[0030] Exemplarily, in this stage, a confidence evaluation needs to be performed on each generated target candidate box to determine the likelihood that it contains a candidate target. In the target recognition process of a millimeter-wave radar, candidate targets usually refer to objects in the environment detected by the radar that are considered likely to be actual targets after certain processing, including: pedestrians, vehicles, cyclists, animals, and fixed obstacles.

[0031] Confidence evaluation is usually achieved through a classifier, which outputs the confidence scores of each target candidate box. Based on these scores, the motion features of the candidate targets, such as speed and direction, and the probability distribution of the target types can be further calculated.

[0032] Exemplarily, the specific calculation formula for the confidence score is: ; where C(b) represents the confidence score of target candidate box b, is the Sigmoid activation function, Φ(b) is the feature vector of target candidate box b, W is the learnable weight matrix, and β is the bias term.

[0033] The formula for calculating the weighted velocity is as follows: ; wherein, is the forgetting factor, and its value range is 0.7 - 0.9, V(t - 1) is the velocity at the previous moment, ΔP is the position change, and Δt is the time interval Direction angle weighted calculation: θ = C(b) * [arctan2(Δy, Δx) + ε * N(0, 1)]; wherein, arctan2(Δy, Δx) is the basic direction angle, ε is the noise modulation factor, and N(0, 1) is the Gaussian random perturbation.

[0034] The specific calculation formula for acceleration correction is: A(t) = C(b) * [V(t) - V(t - 1)] / Δt; The specific calculation formula for the motion uncertainty index is: U(motion) = (1 - C(b)) * ||V(t)||; The specific calculation formula for the probability distribution of the target type is: P(C|b) = softmax(G(Φ(b)) * Λ); wherein, Λ is the type discrimination weight matrix, and G is the non - linear mapping function.

[0035] This information will be used to distinguish real targets from noise or false targets, ensuring that only candidate boxes with high confidence enter the next step of processing.

[0036] S105. According to the motion characteristics and type probabilities, perform clustering and non - maximum suppression processing on the candidate targets to obtain the target recognition results.

[0037] Exemplarily, perform clustering processing on the candidate targets to identify targets with similar motion characteristics and type probabilities.

[0038] By analyzing the similarity between the motion characteristics and type probabilities, group the similar candidate targets into one category. Then, apply the non - maximum suppression (NMS) technique to eliminate the overlapping candidate boxes, ensuring that each target is represented by only one best candidate box. Non - maximum suppression obtains accurate target recognition results by retaining the candidate box with the highest confidence and deleting other candidate boxes with large overlaps with it. These results can be used for further applications, such as target tracking and obstacle avoidance in autonomous driving.

[0039] To introduce the technical solution of this application more clearly, the technical solution of this application will also be introduced through specific embodiments below. It should be noted that the specific embodiments are used to expand the description of the technical solution of this application, rather than limiting this application.

[0040] In some embodiments, distance calculations are performed on the detection points of the millimeter-wave radar to obtain a detection point map, and the point cloud map structure is determined according to the detection point map. The specific steps include: S1011 - S101 S1011. Perform translation and rotation processing on the detection points of the millimeter-wave radar for a preset number of frames to obtain time-sequential point cloud data, and superimpose the point clouds according to the spatial coordinates of the time-sequential point cloud data to obtain a superimposed point cloud.

[0041] Exemplarily, in the translation and rotation processing, the detection point data captured by the millimeter-wave radar within a preset number of frames is subjected to coordinate transformation. The preset number of frames is 5 frames to obtain point cloud data, so as to eliminate the influence caused by the movement of the radar itself or environmental changes. Then, the point cloud data of multiple time frames are arranged in chronological order, recording the position changes of each detection point in different time frames. Next, the point cloud data of multiple time frames are superimposed together to form a superimposed point cloud containing all detection points. This superimposing method can better reflect the overall structure and movement trajectory of the target object.

[0042] S1012. Perform spatial distance calculations on the points in the superimposed point cloud to obtain a point cloud spatial distance matrix, and screen the point cloud spatial distance matrix according to a preset radius threshold to obtain a neighborhood point set.

[0043] Exemplarily, in this step, the superimposed point cloud is obtained by superimposing the point cloud data of multiple frames in the same coordinate system. It is necessary to calculate the spatial distance between each point in these superimposed point clouds and other points to obtain a point cloud spatial distance matrix. Each element in this matrix represents the Euclidean distance between two points. Next, screen the point cloud spatial distance matrix according to a preset radius threshold. The value range of the preset radius threshold is: 0.5m - 5m, and different values are selected according to different scenario requirements. Retain the point pairs whose distances are within the range of the preset radius threshold to form a neighborhood point set.

[0044] S1013. Perform edge connection processing on the neighborhood point set to obtain graph structure vertex data, and construct self-loop edges according to the graph structure vertex data to obtain an initial point map.

[0045] Exemplarily, based on the neighborhood point set, an edge connection operation is performed between points to generate vertex data of the graph structure. This process is similar to constructing a graph in point cloud data, where each point is a vertex, and edges are connected between eligible point pairs. The existence of an edge indicates that two points are spatially adjacent. Based on these edges and vertex data, a graph structure can be constructed, and on this basis, self-loop edges (self-loop edges are edges of the vertex itself) are added to obtain the initial point graph.

[0046] S1014. Perform an Euclidean coordinate system transformation on the points in the initial point graph to obtain target velocity data, and calculate the number of connected edges of the points based on the target velocity data to obtain point density characteristics.

[0047] Exemplarily, at this stage, it is necessary to perform an Euclidean coordinate system transformation on the points in the initial point graph and calculate the target velocity data. This process involves converting the points from the current coordinate system to a new coordinate system to better describe the motion state of the points. Based on the target velocity data, the number of connected edges of each point, that is, the connection quantity between this point and other points, can be calculated, thereby obtaining point density characteristics. Point density characteristics are important indicators for describing the spatial distribution density of points.

[0048] S1015. Perform time frame marking on the points in the initial point graph according to the point density characteristics to obtain marked point cloud data, and perform state vector configuration on the marked point cloud data to obtain point attribute data.

[0049] Exemplarily, according to the point density characteristics, time frame marking can be performed on the points in the initial point graph. Time frame marking means assigning a time frame label to each point to represent its state in different time periods. Next, state vector configuration is performed on these marked point cloud data. The state vector can include information such as the velocity, acceleration, and position of the points to obtain point attribute data, which are the basis for subsequent analysis and processing.

[0050] S1016. Extract features from the millimeter-wave radar cross section in the point attribute data to obtain an RCS feature vector, and construct a classification covariate based on the RCS feature vector and the point density characteristics to obtain a point cloud graph structure.

[0051] Exemplarily, extract features from the millimeter-wave radar cross section (RCS, Radar Cross Section) in the point attribute data to obtain an RCS feature vector. The RCS feature vector is an important indicator for describing the radar reflection characteristics of target objects. Combining the RCS feature vector and the point density characteristics, a classification covariate can be constructed. Classification covariates are important inputs for classification analysis, and through these covariates, the point cloud graph structure can be determined.

[0052] In some embodiments, a multi-layer perceptron mapping is performed on the point cloud map structure to obtain non-contextual encoded features, and the non-contextual encoded features are converted into a high-dimensional vector representation to obtain feature space data. The specific steps include: S1021 - S1026.

[0053] S1021. Perform initial feature extraction on the point covariates in the point cloud map structure to obtain a set of point covariance features.

[0054] Exemplarily, the point covariates include covariates such as the spatial coordinates, reflection intensity, velocity information, and timestamp of each point. According to the physical meaning and statistical characteristics of the point covariates, a covariance feature vector is constructed. Local geometric feature calculations are performed on the covariance feature vector to extract geometric features such as local curvature, normal vector, and point density, obtaining a geometric feature vector; the covariance feature vector and the geometric feature vector are concatenated to form a set of point covariance features.

[0055] S1022. Perform a non-linear mapping transformation on the set of point covariance features through a preset shared multi-layer perceptron to obtain an initial feature encoding at the point level. Input the set of point covariance features into the shared multi-layer perceptron. The shared multi-layer perceptron includes multiple fully connected layers and an activation function layer. Perform weighted summation and bias addition operations on the input set of point covariance features in each fully connected layer to obtain an intermediate activation value, and input the intermediate activation value into a non-linear activation function for non-linear transformation to obtain a non-linear activation output. Connect the non-linear activation output through the activation function layer of the multi-layer perceptron and process it layer by layer to obtain an initial feature encoding at the point level.

[0056] S1023. According to the initial feature encoding at the point level, construct a feature mapping matrix, and perform dimensionality transformation and normalization processing on the feature mapping matrix to obtain feature-normalized data.

[0057] Exemplarily, arrange the initial feature encoding at the point level into a feature mapping matrix according to the spatial position and topological relationship of the points; perform dimensionality transformation methods such as principal component analysis (PCA) or singular value decomposition (SVD) on the feature mapping matrix to reduce feature redundancy and obtain a reduced-dimensional feature matrix; perform normalization processing on the reduced-dimensional feature matrix, including mean normalization and variance normalization, to obtain feature-normalized data.

[0058] S1024. Perform attention mechanism feature weighting on the feature-normalized data to obtain a weighted feature vector.

[0059] Exemplarily, perform positional encoding on the feature-normalized data to obtain feature data with positional information. Input the feature data with positional information into the attention mechanism network and calculate the attention weights for each feature. The attention mechanism includes self-attention and multi-head attention mechanism. According to the calculated attention weights, perform weighted processing on the feature-normalized data to obtain a weighted feature vector.

[0060] In one embodiment, perform attention mechanism feature weighting on the feature-normalized data. The specific steps include: 1) Perform positional encoding on the feature-normalized data to obtain feature data with positional information, including: Concatenate the feature vector of each point in the feature-normalized data with its spatial position coordinates to form feature data with positional information. Perform a linear transformation on the feature data with positional information to embed the positional information into the feature vector.

[0061] 2) Input the feature data with positional information into the self-attention mechanism to calculate the self-attention weights, including: For each feature vector, copy it into query, key, and value vectors. Calculate the similarity between the query vector and the key vector through dot product operation to obtain an attention score matrix. Normalize each element in the attention score matrix and use the softmax function to convert it into an attention weight. Multiply the attention weight by the value vector and sum all the results to obtain a weighted feature vector.

[0062] 3) Input the weighted feature vector into the multi-head attention mechanism to calculate the multi-head attention weights, including: Pass the weighted feature vector through multiple different linear transformation layers to generate multiple different query, key, and value vectors, i.e., multiple heads in the multi-head attention mechanism. For each head, repeat the steps in the self-attention mechanism to calculate the attention weights and weighted features for each head. Concatenate the weighted feature vectors of all heads to form a joint feature representation of multiple heads. Perform a linear transformation on the concatenated joint feature representation to obtain a multi-head weighted feature vector.

[0063] 4) According to the calculated multi-head weighted feature vector, perform weighted processing on the feature-normalized data to obtain a weighted feature vector, including: Perform weighted summation on the feature-normalized data and the multi-head weighted feature vector to obtain a fused feature vector; Perform normalization processing on the fused feature vector, including mean normalization and variance normalization, to obtain a weighted feature vector.

[0064] S1025. According to the weighted feature vector, construct a multi-scale feature mapping structure, perform dimensionality reduction and information aggregation on the multi-scale feature mapping structure to obtain a compressed high-dimensional feature representation.

[0065] Input the weighted feature vector into the multi-scale feature mapping module, which includes convolutional layers or pooling layers with different receptive fields; extract features from the weighted feature vector at different scales to obtain multi-scale feature representations; perform dimensionality reduction on the multi-scale feature representations, using fully connected layers or dimensionality reduction algorithms to reduce the feature dimensions; aggregate the features at different scales through information aggregation operations, such as weighted summation, concatenation, etc., to obtain a compressed high-dimensional feature representation.

[0066] S1026. Project the compressed high-dimensional feature representation into the semantic space related to target recognition through a non-linear transformation to obtain feature space data.

[0067] Input the compressed high-dimensional feature representation into the non-linear transformation module, which includes an activation function and a Dropout layer. Perform a non-linear transformation on the input high-dimensional feature representation to obtain a semantic feature vector. Project the semantic feature vector through the Dropout layer to map it into the semantic space related to target recognition to obtain feature space data.

[0068] In some embodiments, perform graph convolution processing on the feature space data to obtain context point encodings, and generate target candidate boxes according to the context point encodings, including: S1031 - S1035.

[0069] S1031. Calculate the corner points and relative positions of the point features in the feature space data to obtain edge feature vectors, and construct a spatial coordinate transformation invariance mapping according to the edge feature vectors.

[0070] Exemplarily, extract the information of each point feature from the input feature space data. The information of the point feature includes: position coordinates, color, normal direction, etc. These points are usually discrete data in three-dimensional space. To identify the structural information in the point cloud, the corner points and relative position relationships between the points in the point cloud can be calculated. Corner points are points where the local geometric shape changes drastically, and corner features can be extracted through the distance changes between adjacent points. Relative position calculation describes the spatial relationship through the relative positions between a point and its neighboring points, which provides the necessary spatial information for constructing a graph convolution network. By calculating the geometric relationships of each point relative to other points, edge feature vectors can be generated. This vector describes the morphological features of the edges in the point cloud, such as the thickness and direction of the edges. The edge feature vectors help to effectively distinguish different geometric structures in subsequent processing. Through the spatial coordinate transformation invariance mapping, regardless of how the input point cloud data rotates or transforms, the obtained feature mapping should be insensitive to these changes, ensuring the robustness of the model.

[0071] S1032. Perform max-pooling feature extraction on the spatial coordinate transformation invariance mapping, and perform a non-linear encoding transformation on the pooled features to obtain local feature encodings.

[0072] Exemplarily, the maximum value in the local neighborhood is selected as the representative feature of the area through max pooling, which helps to retain the most important geometric information in the graph, including local geometric information such as edges, surfaces, etc. The pooled features will undergo a non-linear transformation (such as the ReLU activation function) to convert them into a feature space more suitable for subsequent processing. The purpose of this step is to improve the expressive ability of the features so that complex spatial relationships can be captured.

[0073] S1033. Based on the local feature encoding, construct a multi-layer perceptron network, perform iterative graph convolution processing on the point features in the multi-layer perceptron network, and layer by layer extract the local and global semantic features of the point cloud space to obtain the feature vector after iterative graph convolution.

[0074] Exemplarily, the local feature encoding is further processed through a multi-layer perceptron network (MLP). The role of the MLP network in this step is to map the local features to a higher-dimensional feature space to provide richer information for graph convolution. Through multiple iterations of graph convolution, the feature vector of each point will gradually fuse with the information of its neighboring points, thereby extracting the semantic features of this point in the global space. Each layer of graph convolution focuses on the local structure of neighboring points and extracts local spatial information. As the iteration progresses, the propagation of information gradually involves more distant points, thereby gradually capturing the global spatial features in the point cloud. After iterative graph convolution processing, each point will have a new feature vector, which contains the information of this point in terms of local and global semantics and can be used for subsequent generation of target candidate boxes.

[0075] S1034. Perform dimensional mapping and decoding on the feature vector after iterative graph convolution, and generate two-dimensional target candidate box parameters for each encoded point through a preset three-layer multi-layer perceptron decoder.

[0076] Exemplarily, the feature vector after graph convolution usually needs to be dimensionally mapped in some way to make it suitable for subsequent decoding tasks. Dimensional mapping usually includes a linear transformation to map the features to a two-dimensional space for generating two-dimensional candidate boxes. In order to recover the parameters of the two-dimensional candidate box from the high-dimensional features, a three-layer MLP decoder is used. The role of this three-layer MLP decoder is to obtain the specific parameters of the two-dimensional target candidate box of this point through non-linear transformation according to the feature vector of each point. These parameters include the coordinates of the candidate box (such as the position of the center point) and scale information.

[0077] S1035. Perform direction prediction and speed estimation on the two-dimensional target candidate box parameters, predict the direction of the bounding box as the probability distribution in 8 equivalent directions, and calculate the squared scalar value of the target motion speed according to the probability distribution to obtain the complete target candidate box information.

[0078] Exemplarily, the direction of the candidate box is predicted through the network. The predicted direction is usually to discretize the orientation of the candidate box into multiple directions (such as 8 equivalent directions). The direction prediction is given in the form of a probability distribution, reflecting the possibility of each direction. Using the historical information and current position of the target candidate box, the target's movement speed is estimated. This can be done by calculating the square value of the speed of each candidate box, and further inferring the target's movement state in the future. Finally, through direction prediction and speed estimation, complete target candidate box information can be obtained. This information includes not only the position and size, but also the target's movement trajectory and speed, providing a comprehensive reference for subsequent target tracking or detection tasks.

[0079] In some embodiments, a confidence assessment is performed on the target candidate frame to obtain target confidence data, and motion features and type probabilities of the candidate targets are calculated based on the target confidence data, including: S1041-S1045.

[0080] S1041, assigning class labels to the detection points in the target candidate frame to obtain class label data, and calculating the frame overlap according to the class label data; Exemplarily, the process of assigning class labels to detection points within the target candidate frame can be implemented in a variety of ways, for example, using a pre-trained model for classification, or using a manually annotated data set for supervised learning. Each detection point is assigned a specific class label to identify which category the detection point belongs to. Based on these class label data, the overlap between the candidate frame and the true target frame can be calculated (usually using IoU, Intersection over Union). Frame overlap is an important indicator to measure the degree of match between the candidate frame and the actual target frame. The higher the overlap, the closer the candidate frame is to the true target frame.

[0081] S1042. Perform cluster prediction on the candidate points according to the class label data to obtain cross entropy loss data, and perform superposition calculation on the cross entropy loss data to obtain a classification loss value.

[0082] Cluster prediction can be achieved through a variety of algorithms, such as K-means, DBSCAN, or neural network-based clustering methods. The cross entropy loss function is used to evaluate the effect of cluster prediction. The cross entropy loss function is a commonly used loss function in classification problems, which is used to measure the difference between the predicted distribution and the true distribution. By superimposing the cross entropy loss of each candidate point, the overall classification loss value can be obtained. The smaller the classification loss value, the better the cluster prediction effect and the more accurate the class label assignment.

[0083] S1043. Perform Huber loss function calculation on the target candidate frame to obtain predicted loss data, and calculate the target pose loss based on the predicted loss data.

[0084] Exemplarily, the Huber loss function is a loss function that is insensitive to outliers. It combines the advantages of the mean squared error (MSE) and the mean absolute error (MAE). It behaves as the mean squared error when the error is small and as the mean absolute error when the error is large. Through the Huber loss function, the predicted loss data of the target candidate box can be calculated. Based on these predicted loss data, the target pose loss is further calculated. The target pose loss is used to measure the difference between the position and pose information of the target candidate box and the true target box. By minimizing the pose loss, the accuracy of the target candidate box can be improved to better match the true target box.

[0085] S1044. Perform L2 regularization on the pose loss to obtain the target function parameters, and perform Adam optimization calculation according to the target function parameters.

[0086] S1045. Calculate the probability distribution of the target candidate box according to the optimization calculation result to obtain the target type probability as the confidence evaluation result.

[0087] Exemplarily, according to the target function parameters obtained by the optimization calculation, the probability distribution of the target candidate box is calculated. The purpose of the probability distribution calculation is to determine the probability that the candidate box belongs to different classes. Usually, the Softmax function is used to convert the output value into the target type probability as the confidence evaluation result.

[0088] In some embodiments, according to the motion characteristics and type probability, clustering and non-maximum suppression processing are performed on the candidate targets to obtain the target recognition result, including: S1051 - S1056.

[0089] S1051. Perform K-means clustering calculation on the motion characteristics to obtain the target motion trajectory data, and perform temporal association on the candidate targets according to the target motion trajectory data to obtain the target trajectory association matrix.

[0090] Exemplarily, the goal of K-means clustering is to divide the data points into K clusters so that the points within each cluster are as similar as possible, while the clusters are as different as possible. In target tracking, K-means clustering can help identify the motion trajectories of different targets. Each cluster obtained by K-value clustering calculation represents the motion trajectory of a target. The motion trajectory of each target may include dynamic characteristics such as its position, speed, and acceleration. Using the target motion trajectory data to perform temporal association on the candidate targets at different time points, matching the motion characteristics of the same target at different time points to form the temporal motion pattern of the target. This matrix represents the association relationship of different targets in the time series. For example, the elements of the matrix can represent the association degree between the targets at two time points.

[0091] S1052. Perform feature normalization on the association data in the target trajectory association matrix to obtain a normalized feature vector. Calculate the distance metric for candidate targets based on the normalized feature vector to obtain a target distance matrix.

[0092] Exemplarily, the normalized feature vector contains the temporal information of the target movement. The normalized vector is more adaptable to candidate targets with different scales and different movement characteristics. Using the normalized feature vector to perform distance metric on candidate targets to obtain a target distance matrix, common metric methods include Euclidean distance, Manhattan distance, etc. These metric methods help to quantify the similarity between targets. This matrix reflects the similarity or difference between targets by calculating the distances between candidate targets. Each element of the matrix represents the distance between two targets.

[0093] S1053. Perform hierarchical clustering on the target distance matrix to obtain clustering result data. Group the candidate target boxes according to the clustering result data to obtain target box combination data.

[0094] Exemplarily, in hierarchical clustering, calculate the distances or similarities of candidate targets based on the target distance matrix, and gradually merge the candidate targets to form a hierarchical structure. Each cluster in the clustering result data may represent a specific candidate target or an aggregation of multiple candidate targets. In this way, the number of candidate target boxes can be effectively reduced, facilitating subsequent target recognition and screening.

[0095] S1054. Calculate the intersection over union (IoU) threshold for the target box combination data to obtain overlap degree data. Perform non-maximum suppression on the overlapping target boxes according to the overlap degree data to obtain the filtered target boxes.

[0096] Exemplarily, IoU is an index used in object detection to measure the overlapping degree of two target boxes. The higher the IoU value, the greater the overlapping degree of the two target boxes. The calculation formula is: IoU = (area of intersection of target boxes) / (area of union of target boxes). Set an IoU threshold, usually 0.5 or 0.7, indicating that if the IoU of two target boxes is greater than this threshold, it is considered that they represent the same target and may need to be merged or removed. The removal process requires non-maximum suppression (NMS). NMS is a common post-processing method used to remove duplicate target boxes. For target boxes with a high overlapping degree, NMS will select the box with the highest confidence as the final box and suppress (delete) other target boxes with a high overlapping degree.

[0097] S1055. Calculate the class probability for the filtered target boxes to obtain class weight data. Perform weighted average processing on the target boxes according to the class weight data to obtain target box correction data.

[0098] Exemplarily, the Softmax function is usually used for multi-class classification, that is, to calculate the class probabilities of the filtered target boxes, so as to obtain the class weight data of each target box, indicating the credibility of each class. In the target box, if there are prediction results of multiple classes, the class prediction of the target box is corrected by means of weighted average. This weighted average is usually based on the probabilities or weights of the classes to ensure that the final result reflects the most likely class.

[0099] S1056. Perform boundary optimization processing on the target box correction data to obtain target box data, and output the position, speed, class, and confidence information of the target according to the target box data to obtain the target recognition result.

[0100] Exemplarily, the data of the corrected target box may need further boundary optimization, mainly for fine-tuning the position, size, etc. of the target box. This step usually uses a regression model for boundary correction, or uses some geometric optimization methods such as edge alignment and box shape adjustment. The output includes information such as the bounding box, class, speed, and confidence of the target. These information are very important for subsequent applications such as target tracking, behavior analysis, and autonomous driving. The position can be represented by coordinates, the speed is usually calculated based on the movement trajectory of the target, and the confidence is the degree of confidence in this prediction.

[0101] Please refer to Figure 2 , Figure 2 FIG. is a schematic block diagram of an object recognition device of a millimeter-wave radar provided by an embodiment of the present application. The object recognition device 200 of the millimeter-wave radar is used to execute the foregoing object recognition method of the millimeter-wave radar. Among them, the object recognition device 200 of the millimeter-wave radar can be configured in a server.

[0102] Among them, the server can be an independent server, a server cluster, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms.

[0103] As Figure 2 shown, the object recognition device 200 of the millimeter-wave radar includes: a cloud map calculation module 201, a feature mapping module 202, a target locking module 203, a data analysis module 204, and a result output module 205.

[0104] The cloud map calculation module 201 is used to calculate the distance of the detection points of the millimeter-wave radar to obtain a detection point map, and determine the point cloud map structure according to the detection point map; The feature mapping module 202 is used to perform multi-layer perceptron mapping on the point cloud map structure to obtain non-context-encoded features, convert the non-context-encoded features into high-dimensional vector representations, and obtain feature space data; The target locking module 203 is used to perform graph convolution processing on the feature space data to obtain context point encodings, and generate target candidate boxes according to the context point encodings; The data analysis module 204 is used to evaluate the confidence of the target candidate boxes to obtain target confidence data, and calculate the motion features and type probabilities of the candidate targets according to the target confidence data; The result output module 205 is used to perform clustering and non-maximum suppression processing on the candidate targets according to the motion features and type probabilities to obtain target recognition results.

[0105] An embodiment of the present application provides a millimeter-wave radar, which includes a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program and implement the target recognition method of the millimeter-wave radar according to any one of the embodiments of the present application when executing the computer program.

[0106] An embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the processor is caused to implement the target recognition method of the millimeter-wave radar according to any one of the embodiments of the present application.

[0107] The above is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of various equivalent modifications or substitutions, and these modifications or substitutions should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A target recognition method for a millimeter-wave radar, characterized in that, The method includes: Calculating the distances of the detection points of the millimeter-wave radar to obtain a detection point map, and determining the point cloud map structure according to the detection point map; Performing multi-layer perceptron mapping on the point cloud map structure to obtain non-context encoded features, converting the non-context encoded features into high-dimensional vector representations, and obtaining feature space data; Performing graph convolution processing on the feature space data to obtain context point encodings, and generating target candidate boxes according to the context point encodings; Performing confidence evaluation on the target candidate boxes to obtain target confidence data, and calculating the motion features and type probabilities of the candidate targets according to the target confidence data; Performing clustering and non-maximum suppression processing on the candidate targets according to the motion features and the type probabilities to obtain target recognition results.

2. The method for target recognition of a millimeter-wave radar according to claim 1, characterized in that, The calculating the distances of the detection points of the millimeter-wave radar to obtain a detection point map, and determining the point cloud map structure according to the detection point map includes: Performing translation and rotation processing on the detection points of the millimeter-wave radar for a preset number of frames to obtain time-series point cloud data, and superimposing the point clouds according to the spatial coordinates of the time-series point cloud data to obtain superimposed point clouds; Calculating the spatial distances between the points in the superimposed point clouds to obtain a point cloud spatial distance matrix, and screening the point cloud spatial distance matrix according to a preset radius threshold to obtain a neighborhood point set; Performing edge connection processing on the neighborhood point set to obtain graph structure vertex data, and constructing self-loop edges according to the graph structure vertex data to obtain an initial point map; Performing Euclidean coordinate system transformation on the points in the initial point map to obtain target velocity data, and calculating the number of connected edges of the points according to the target velocity data to obtain point density features; Performing time frame marking on the points of the initial point map according to the point density features to obtain marked point cloud data, and performing state vector configuration on the marked point cloud data to obtain point attribute data; Extracting features of the millimeter-wave radar cross section in the point attribute data to obtain an RCS feature vector, and constructing a classification covariate according to the RCS feature vector and the point density features to obtain the point cloud map structure.

3. The target recognition method of the millimeter-wave radar according to claim 1, characterized in that, The performing multi-layer perceptron mapping on the point cloud map structure to obtain non-context encoded features, converting the non-context encoded features into high-dimensional vector representations, and obtaining feature space data includes: Performing initial feature extraction on the point covariates in the point cloud map structure to obtain a point covariance feature set; Performing non-linear mapping transformation on the point covariance feature set through a preset shared multi-layer perceptron to obtain initial feature encodings at the point level; Constructing a feature mapping matrix according to the initial feature encodings at the point level, and performing dimension transformation and normalization processing on the feature mapping matrix to obtain feature normalization data; Performing attention mechanism feature weighting on the feature normalization data to obtain weighted feature vectors; Constructing a multi-scale feature mapping structure according to the weighted feature vectors, and performing dimensionality reduction and information aggregation on the multi-scale feature mapping structure to obtain a compressed high-dimensional feature representation; Projecting the compressed high-dimensional feature representation into a semantic space related to target recognition through non-linear transformation to obtain the feature space data.

4. The method for target recognition of a millimeter-wave radar according to claim 1, characterized in that, Performing graph convolution processing on the feature space data to obtain context point encodings, and generating target candidate boxes based on the context point encodings, includes: Calculating corner points and relative positions of the point features in the feature space data to obtain edge feature vectors, and constructing a space coordinate transformation invariance mapping based on the edge feature vectors; Performing max-pooling feature extraction on the space coordinate transformation invariance mapping, and performing a non-linear encoding transformation on the pooled features to obtain local feature encodings; Constructing a multi-layer perceptron network based on the local feature encodings, performing iterative graph convolution processing on the point features in the multi-layer perceptron network, and layer-by-layer extracting local and global semantic features of the point cloud space to obtain an iteratively graph-convolved feature vector; Performing dimension mapping and decoding on the iteratively graph-convolved feature vector, and generating two-dimensional target candidate box parameters for each encoded point through a preset three-layer multi-layer perceptron decoder; Performing direction prediction and speed estimation on the two-dimensional target candidate box parameters, predicting the bounding box direction as a probability distribution in 8 equivalent directions, and calculating the squared scalar value of the target motion speed based on the probability distribution to obtain complete target candidate box information.

5. The target recognition method of the millimeter-wave radar according to claim 1, wherein, Performing confidence evaluation on the target candidate boxes to obtain target confidence data, and calculating the motion features and type probabilities of the candidate targets based on the target confidence data, includes: Performing class label assignment on the detection points within the target candidate boxes to obtain class label data, and calculating the box overlap degree based on the class label data; Performing clustering prediction on the candidate points based on the class label data to obtain cross-entropy loss data, and performing superposition calculation on the cross-entropy loss data to obtain a classification loss value; Performing Huber loss function calculation on the target candidate boxes to obtain prediction loss data, and calculating the target pose loss based on the prediction loss data; Performing L2 regularization processing on the pose loss to obtain target function parameters, and performing Adam optimization calculation based on the target function parameters; Performing probability distribution calculation on the target candidate boxes based on the optimization calculation results to obtain target type probabilities, as the confidence evaluation results.

6. The target recognition method of the millimeter-wave radar according to claim 1, characterized in that, Based on the motion features and the type probabilities, performing clustering and non-maximum suppression processing on the candidate targets to obtain target recognition results, includes: Performing K-means clustering calculation on the motion features to obtain target motion trajectory data, and performing temporal association on the candidate targets based on the target motion trajectory data to obtain a target trajectory association matrix; Performing feature normalization processing on the association data in the target trajectory association matrix to obtain a normalized feature vector, and performing distance metric calculation on the candidate targets based on the normalized feature vector to obtain a target distance matrix; Performing hierarchical clustering processing on the target distance matrix to obtain clustering result data, and grouping the candidate target boxes based on the clustering result data to obtain target box combination data; Performing intersection-over-union threshold calculation on the target box combination data to obtain overlap degree data, and performing non-maximum suppression processing on the overlapping target boxes based on the overlap degree data to obtain filtered target boxes; Calculate the class probability of the filtered target bounding boxes to obtain class weight data, and perform weighted average processing on the target bounding boxes according to the class weight data to obtain target bounding box correction data; Perform boundary optimization processing on the target bounding box correction data to obtain target bounding box data, and output the position, speed, class, and confidence information of the candidate target according to the target bounding box data to obtain the target recognition result.

7. A target recognition device for a millimeter-wave radar, characterized in that, The target recognition device of the millimeter wave radar includes: A cloud map calculation module, configured to calculate the distance of the detection points of the millimeter wave radar to obtain a detection point map, and determine the point cloud map structure according to the detection point map; A feature mapping module, configured to perform multi-layer perceptron mapping on the point cloud map structure to obtain non-context encoded features, and convert the non-context encoded features into high-dimensional vector representations to obtain feature space data; A target locking module, configured to perform graph convolution processing on the feature space data to obtain context point encoding, and generate target candidate bounding boxes according to the context point encoding; A data analysis module, configured to evaluate the confidence of the target candidate bounding boxes to obtain target confidence data, and calculate the motion characteristics and type probabilities of the candidate targets according to the target confidence data; A result output module, configured to perform clustering and non-maximum suppression processing on the candidate targets according to the motion characteristics and the type probabilities to obtain the target recognition result.

8. A millimeter-wave radar, characterized in that, The millimeter wave radar includes a memory and a processor; The memory is used to store a computer program; The processor is configured to execute the computer program and implement the target recognition method of the millimeter wave radar according to any one of claims 1 to 6 when executing the computer program.

Images