A high-fidelity radar RCS prediction method, system, terminal and medium for autonomous driving real-time simulation

Abstract

The application relates to the technical field of automatic driving, and discloses a high-fidelity radar RCS prediction method, system, terminal and medium for automatic driving real-time simulation, the method comprising the following steps: acquiring radar data of multiple radars in a real road driving scene, and processing the radar data to obtain a high-dimensional feature vector; training an offline model according to the high-dimensional feature vector, and outputting a pre-training RCS generation model; acquiring true value state information of all targets in a simulation scene, and obtaining a target feature vector according to the true value state information; and obtaining radar RCS data according to the target feature vector based on the pre-training RCS generation model. The application can efficiently and in real time predict a dynamic RCS value conforming to a real world physical law, thereby guaranteeing the effectiveness of simulation test results.

Description

Technical Field

[0001] This application relates to the field of autonomous driving technology, and in particular to a high-fidelity radar RCS prediction method, system, terminal and medium for real-time simulation of autonomous driving. Background Technology

[0002] The development of autonomous driving systems, especially the training, testing, and verification of perception and decision-making algorithms, relies on large-scale testing. Field testing and road testing face problems such as poor safety, high costs, and low testing efficiency. Virtual simulation platforms can effectively alleviate these problems, generating massive amounts of diverse, even realistic, edge-and-dangerous driving scenarios that are difficult to reproduce in the real world. This is considered an inevitable choice to accelerate the development, technological iteration, and deployment of autonomous driving systems. Millimeter-wave radar, as a core sensor for achieving all-weather, all-time environmental perception, directly determines the effectiveness and reliability of simulation testing based on the modeling fidelity in the simulation environment. A high-fidelity radar model must accurately simulate the physical characteristics of radar sensors in the real world. Among these, RCS, as a key physical quantity describing the electromagnetic scattering characteristics of a target, comprehensively reflects the target's size, geometry, surface material, and orientation, directly affecting the signal-to-noise ratio and detection probability of the simulated radar point cloud. It is a core technical aspect of achieving high-fidelity radar simulation and one of the key technologies in autonomous driving simulation system modeling.

[0003] Currently, existing technologies for predicting RCS data in millimeter-wave radar models on autonomous driving simulation platforms mainly fall into three categories. The first category is electromagnetic simulation methods based on physical optics / geometric optics. These methods obtain RCS values ​​in a virtual 3D scene by calculating the interaction between electromagnetic waves and the simulated target in real time through ray tracing. However, accurate electromagnetic simulation requires intensive ray tracing and complex numerical solutions to physical equations, resulting in huge computational overhead and failing to meet the requirements of real-time simulation. For example, while ray tracing-based millimeter-wave radar simulation methods pursue high accuracy, their computational complexity makes them difficult to apply to simulations that require simultaneous calculation of the RCS of dozens or hundreds of traffic targets, and they also cannot support hardware-in-the-loop (HIL) testing that requires real-time calculation. In this type of method, the scattering center RCS estimation method decomposes complex targets into several discrete local scattering centers, extracts the characteristic parameters of each scattering center from the radar echo using a geometric diffraction model, and finally reconstructs the total RCS characteristics of the target through coherent superposition. This simplification achieves a moderate balance between computational efficiency and accuracy, but because its core modeling implementation is still based on simplified scenes and targets, it cannot include the more complex environmental factors in natural driving scenarios, resulting in limited simulation accuracy. The second category is based on static or simplified RCS lookup tables. The basic idea is that the simulation platform presets a fixed RCS value for each type of simulated target, or a two-dimensional / three-dimensional lookup table that only changes with the horizontal and pitch angles. This is the basic implementation method of radar models in most current autonomous driving simulation software. While similar methods solve the real-time simulation problem, the models are overly simplified and lack dynamism and realism. In the real physical world, the target's RCS is closely related to complex factors such as the observation angle, relative position, occlusion, and even its own minute dynamics, trajectory smoothness, and multipath effects. RCS values ​​fluctuate, and static or simplified lookup tables completely ignore these micro-dynamic characteristics, resulting in overly idealized radar data that cannot effectively test the robustness of the perception algorithm under real signal fluctuations. The third category is based on random value injection methods using simple statistical models. This involves superimposing random noise conforming to a specific statistical distribution onto the baseline RCS value given by the lookup table to simulate RCS fluctuations. Its main drawback is that it lacks physical basis for simulating RCS fluctuations, resulting in strong pseudo-randomness. Therefore, existing RCS modeling for millimeter-wave radar in the field of autonomous driving struggles to balance fidelity, computational efficiency, and dynamic realism. This "Sim-to-Real Gap" is one of the key bottlenecks restricting the development of autonomous driving technology and the application of autonomous driving simulation testing technology.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main purpose of this application is to provide a high-fidelity radar RCS prediction method, system, terminal and medium for real-time simulation of autonomous driving, aiming to solve the problem that in the prior art, millimeter-wave radar RCS modeling in autonomous driving simulation is difficult to balance fidelity, efficiency and dynamic realism, resulting in poor effectiveness of simulation testing.

[0006] The first aspect of this application provides a high-fidelity radar RCS prediction method for real-time simulation of autonomous driving, the high-fidelity radar RCS prediction method for real-time simulation of autonomous driving includes the following steps: The radar data of multiple radars in real road driving scenarios are acquired, and the radar data is processed to obtain a high-dimensional feature vector. The offline model is trained based on the high-dimensional feature vectors, and the pre-trained RCS generation model is output. Obtain the truth state information of all targets in the simulation scenario, and obtain the target feature vector based on the truth state information, wherein the target feature vector has the same feature format as the high-dimensional feature vector; Based on the pre-trained RCS generation model, radar RCS data is obtained according to the target feature vector.

[0007] Optionally, in one embodiment of this application, the processing of the radar data to obtain a high-dimensional feature vector specifically includes: The radar data is grouped to obtain multiple data point groups. The multiple data point groups are then filtered for trajectory validity to obtain filtered trajectory groups. Finally, the filtered trajectory groups are sorted in time sequence to obtain an ordered trajectory sequence. The dynamic kinematic features and trajectory history features are calculated based on the ordered trajectory sequence, and the ordered trajectory sequence, the dynamic kinematic features and the trajectory history features are combined into feature vectors to obtain a high-dimensional feature vector.

[0008] Optionally, in one embodiment of this application, the ordered trajectory sequence includes timestamps, locations, azimuth angles, and trajectory indexes of multiple data points; The calculation of dynamic kinematic features and trajectory history features based on the ordered trajectory sequence specifically includes: The heading angle is calculated based on the timestamps, locations, and azimuth angles of multiple data points, and the observation angle is calculated based on the heading angle. Trigonometric transformations are performed on the heading angle and the observation angle to obtain the dynamic kinematic characteristics; The trajectory age is calculated based on the trajectory index, and the trajectory age is transformed to obtain the trajectory historical features.

[0009] Optionally, in one embodiment of this application, the step of training an offline model based on the high-dimensional feature vector and outputting a pre-trained RCS generation model specifically includes: The high-dimensional feature vector is processed to obtain the input tensor; The input tensor is input into the RCS prediction backbone network for feature extraction to obtain feature representation, and Gaussian distribution parameters are obtained based on the feature representation; Calculate the hybrid loss based on the Gaussian distribution parameters and the ordered trajectory sequence; The RCS prediction backbone network is updated based on the hybrid loss to output a pre-trained RCS generation model.

[0010] Optionally, in one embodiment of this application, the ordered trajectory sequence further includes the target RCS ground truth value, and the Gaussian distribution parameters include the mean and the logarithmic standard deviation; The calculation of the hybrid loss based on the Gaussian distribution parameters and the ordered trajectory sequence specifically includes: Based on the mean, the logarithmic standard deviation, and the target RCS true value, calculate the Gaussian negative log-likelihood loss; based on the mean, the logarithmic standard deviation, and the target RCS true value, calculate the Huber loss; and based on the logarithmic standard deviation, calculate the uncertainty regularization loss. The mixed loss is calculated based on the Gaussian negative log-likelihood loss, the Huber loss, and the uncertainty regularization loss.

[0011] Optionally, in one embodiment of this application, obtaining the target feature vector based on the truth state information specifically includes: Calculate the simulation dynamic features and simulation history features based on the truth state information; The target feature vector is obtained by combining the truth state information, the simulation dynamic features, and the simulation history features into feature vectors.

[0012] Optionally, in one embodiment of this application, obtaining radar RCS data based on the target feature vector using the pre-trained RCS generation model specifically includes: The target feature vector is input into the pre-trained RCS generation model, and the RCS prediction distribution parameters under simulation conditions are output. Radar RCS data is obtained by sampling based on the RCS prediction distribution parameters.

[0013] A second aspect of this application also provides a high-fidelity radar RCS prediction system for real-time simulation of autonomous driving, wherein the high-fidelity radar RCS prediction system for real-time simulation of autonomous driving is used to implement the high-fidelity radar RCS prediction method for real-time simulation of autonomous driving as described in any of the above-mentioned solutions; the high-fidelity radar RCS prediction system for real-time simulation of autonomous driving includes: The data acquisition and feature synthesis module is used to acquire radar data from multiple radars in real road driving scenarios and process the radar data to obtain high-dimensional feature vectors. The offline model training module is used to train an offline model based on the high-dimensional feature vector and output a pre-trained RCS generation model. The simulation state extraction and feature synthesis module is used to obtain the true state information of all targets in the simulation scene, and to obtain the target feature vector based on the true state information, wherein the target feature vector has the same feature format as the high-dimensional feature vector; The radar RCS data prediction output module is used to obtain radar RCS data based on the pre-trained RCS generation model and the target feature vector.

[0014] A third aspect of this application also provides a terminal, wherein the terminal includes: a memory, a processor, and a high-fidelity radar RCS prediction program for real-time simulation of autonomous driving stored in the memory and executable on the processor, wherein when the high-fidelity radar RCS prediction program for real-time simulation of autonomous driving is executed by the processor, it implements the steps of the high-fidelity radar RCS prediction method for real-time simulation of autonomous driving as described above.

[0015] A fourth aspect of this application also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a high-fidelity radar RCS prediction program for real-time simulation of autonomous driving, and the high-fidelity radar RCS prediction program for real-time simulation of autonomous driving implements the steps of the high-fidelity radar RCS prediction method for real-time simulation of autonomous driving as described above when executed by a processor.

[0016] Beneficial effects: This application provides a high-fidelity radar RCS prediction method, system, terminal, and medium for real-time simulation of autonomous driving. The invention utilizes massive amounts of radar data collected from the real world and trains a probabilistic RCS prediction model through deep learning. This model can learn and reproduce the complex nonlinear mapping relationship between the target RCS and various scene contexts in the real world. By directly extracting the input parameters required by the generator from the simulation environment and calling the already offline trained RCS prediction model, the dynamic RCS value that conforms to the physical laws of the real world can be predicted in real time and efficiently, ensuring the validity of the simulation test results. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a preferred embodiment of the high-fidelity radar RCS prediction method for real-time simulation of autonomous driving according to this application; Figure 2 This is a diagram showing the installation location and numbering of the vehicle-mounted millimeter-wave radar in a preferred embodiment of the high-fidelity radar RCS prediction method for real-time simulation of autonomous driving in this application. Figure 3 This is a flowchart of the offline training process of the RCS prediction model in a preferred embodiment of the high-fidelity radar RCS prediction method for real-time simulation of autonomous driving in this application. Figure 4 This is a block diagram of a deep probabilistic neural network model system in a preferred embodiment of the high-fidelity radar RCS prediction method for real-time simulation of autonomous driving in this application. Figure 5 This is a schematic diagram of the application architecture of the high-fidelity radar RCS prediction method for real-time simulation of autonomous driving in the autonomous driving simulation platform, as described in the preferred embodiment of this application. Figure 6 This is a structural diagram of a preferred embodiment of the high-fidelity radar RCS prediction system for real-time simulation of autonomous driving according to this application; Figure 7 This is a structural diagram of a preferred embodiment of the terminal of this application.

[0019] Explanation of reference numerals in the attached figures: 100. Data acquisition and feature synthesis module; 200. Offline model training module; 300. Simulation state extraction and feature synthesis module; 400. Radar RCS data prediction output module. Detailed Implementation

[0020] To make the objectives, technical solutions, and effects of this application clearer and more explicit, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are only possible technical implementations of this application and not all possible implementations. Based on the embodiments in this application, those skilled in the art can obtain other embodiments without creative effort, and these embodiments are also within the protection scope of this application.

[0021] To address the technical problems of high computational cost, low fidelity, and lack of dynamic realism in existing autonomous driving simulation technologies for RCS modeling, this invention employs a data-driven, millimeter-wave radar RCS prediction method for autonomous driving simulation. It abandons traditional high-cost physical simulations and low-fidelity simplified models, instead utilizing massive amounts of radar data collected from the real world to train a probabilistic RCS prediction model through deep learning. This model can learn and reproduce the complex nonlinear mapping relationship between the target RCS and various scene contexts in the real world. In the simulation platform, this invention no longer performs physical optical calculations but directly extracts the input parameters required by the generator from the simulation environment and calls the pre-trained offline RCS prediction model to predict dynamically changing RCS values ​​and their physical uncertainties in real time and efficiently, conforming to the physical laws of the real world.

[0022] This invention belongs to the field of autonomous driving system testing and verification technology. In a virtual simulation platform, it can generate high-fidelity, real-time radar cross-section (RCS) data for millimeter-wave radar sensor models.

[0023] The present invention has the following technical effects: 1. The RCS prediction of scene targets is transformed from complex physical ray tracing calculations into a lightweight neural network forward inference, which improves the computational efficiency by several orders of magnitude, making large-scale, real-time radar simulation possible and fully supporting the real-time requirements of HIL closed-loop testing.

[0024] 2. Since the model learns from real-world data, the generated RCS value naturally contains the dynamic changes, random fluctuations and physical laws of the real world, which significantly improves the authenticity and credibility of the simulation data and effectively reduces the gap between simulation and reality.

[0025] 3. This invention can output the probability distribution of RCS prediction, rather than a single point estimate. This allows for the simulation of the real uncertainty and physical randomness of radar signals, providing crucial support for testing the robustness of sensing algorithms under low-confidence signals.

[0026] 4. This invention eliminates the need for complex physical modeling and measurement for new vehicle models. The model can learn its RCS characteristics simply by including the vehicle model data in the real data, which greatly improves the expansion efficiency of the simulation scenario.

[0027] The technical solutions of this application will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0028] The high-fidelity radar RCS prediction method for real-time simulation of autonomous driving described in the preferred embodiment of this application, such as Figure 1 As shown, the high-fidelity radar RCS prediction method for real-time simulation of autonomous driving includes the following steps: In step S1, radar data from multiple radars in a real road driving scenario is acquired, and the radar data is processed to obtain a high-dimensional feature vector.

[0029] This invention utilizes massive amounts of millimeter-wave radar data collected from real-world driving scenarios. The data collection vehicle is equipped with four millimeter-wave radars, such as... Figure 2 As shown, the numbers on the sensors represent sensor numbers. The actual data acquisition vehicle is equipped with four millimeter-wave radars, numbered 1 to 4. The horizontal field of view of each radar is [-60°, +60°]. The two lateral sensors, No. 1 and No. 4, are offset from the main axis of the vehicle by about 70°, while the two forward sensors, No. 2 and No. 3, are offset from the main axis of the vehicle by about 20°.

[0030] It is worth noting that steps S11 and S12 constitute the first stage: collecting the dataset and training the offline model, such as... Figure 3 As shown. The goal of this stage is to train a high-fidelity probabilistic millimeter-wave radar RCS prediction model using real-world data.

[0031] Step S11 specifically includes: grouping the radar data to obtain multiple data point groups; filtering the multiple data point groups for trajectory validity to obtain filtered trajectory groups; sorting the filtered trajectory groups in time sequence to obtain an ordered trajectory sequence; calculating dynamic kinematic features and trajectory history features based on the ordered trajectory sequence; and combining the ordered trajectory sequence, the dynamic kinematic features, and the trajectory history features into feature vectors to obtain a high-dimensional feature vector.

[0032] In one possible implementation, the ordered trajectory sequence includes timestamps, positions, azimuth angles, and trajectory indices for multiple data points.

[0033] The specific implementation of the steps for calculating dynamic kinematic features and trajectory history features is as follows: calculate the heading angle based on the timestamps, positions, and azimuth angles of multiple data points, and calculate the observation angle based on the heading angle; perform trigonometric transformation on the heading angle and the observation angle to obtain dynamic kinematic features; calculate the trajectory age based on the trajectory index, and transform the trajectory age to obtain trajectory history features.

[0034] like Figure 3As shown, step S11 specifically includes: step S201, aligning time-series data to construct a trajectory dataset; step S202, creating dynamic kinematics and trajectory history features for data points; step S203, multimodal feature input and embedding, including feature concatenation; step S204, forward propagation of a deep probabilistic neural network, the core of which is... Figure 4 Model 100; Step S205, Hybrid loss function and optimization, including loss function formula and optimization strategy; Step S206, Output pre-trained RCS prediction model, which represents the final product of the offline training stage, i.e. Figure 5 Module S100 in it.

[0035] Furthermore, S201 constructs a trajectory dataset to align the time-series data. Each data point collected by the data acquisition vehicle includes a high-precision timestamp t and the ID of the onboard millimeter-wave sensor that detected the target. Target distance r Azimuth radial velocity Target RCS truth value The target's position in the global coordinate system Target trajectory ID Target Category ID The system first based on Time-series point clouds belonging to the same target are grouped. To ensure robustness, the system adds [a feature related to...]. Availability determination. If it is in the sequence, The system will only accept the application if the cumulative number of rejections exceeds 3. The record. Unreliable data points will be discarded. Then, within each trajectory, data will be processed according to timestamps. t The data points are strictly sorted in ascending order to form an ordered trajectory sequence. ,in For the first i There are 10 data points. Where t is the timestamp. For sensor ID, r Radial distance; It is the azimuth angle; Radial velocity, This is the true value of the radar cross section (i.e., the true value of the target's RCS). It is a global coordinate sequence. For trajectory ID, This is the category ID.

[0036] The process can be understood as follows: First, all input raw data points are scanned. For each data point, its trackid (target trajectory ID) field is read, and all data points with the same trackid value are grouped into the same data set. That is, the input is an unordered list of raw data points, and the output is a group of data points. Each group theoretically corresponds to a physical target (such as a vehicle or pedestrian) continuously tracked by radar in the real world. Second, trajectory validity is filtered. For each group, the total number of data points within that group is calculated, and a filtering rule is applied to retain long trajectories, ensuring that the learning samples come from stable and real targets. Third, the trajectory is sorted temporally (based on timestamp t). For each filtered trajectory group, the system reads the high-precision timestamp t of each data point and sorts them in strict ascending order (from smallest to largest) according to the value of t, forming an ordered trajectory sequence.

[0037] Step S202 involves creating dynamic kinematics and trajectory history features for the data points.

[0038] This invention applies to each data point in the trajectory sequence T. ( i >1), perform the following feature calculations: (1) Dynamic kinematic characteristics.

[0039] ① Time difference : Calculate the time difference between two adjacent data points in the trajectory sequence.

[0040] ;in, For the time difference, For the first Frame timestamps For the first Frame timestamp; ② Displacement difference : Calculate the position difference between two adjacent data points in the trajectory sequence.

[0041] ;in, for x The amount of displacement in the direction, For the first i Global points x coordinate, For the first Global points x coordinate; ;in, for y The amount of displacement in the direction, For the first i Global points y coordinate, For the first Global points y coordinate; ③ Instantaneous velocity Based on the displacement difference and time difference, assuming that the target moves at a constant speed between two adjacent data points, calculate the instantaneous velocity of the two adjacent data points.

[0042] ;in, for x Instantaneous velocity in the direction; ;in, for y Instantaneous velocity in the direction; ④ Absolute heading angle φ ): This angle represents the direction of the target's movement in the global coordinate system.

[0043] ;in, For heading angle, It is the arctangent function in the four quadrants; ⑤ Radar observation perspective α This angle represents the angle between the target's direction of motion and the centerline of the radar's field of view (FOV), and is a core physical quantity affecting the radar cross-section (RCS).

[0044] ;in, From the perspective of observation, It is the azimuth angle; The direction of movement of the target (the direction the front of the car is facing). The direction the radar is looking at the target. Let be the angle between these two directions.

[0045] ⑥ Angle characteristic stabilization: To solve the problem of 2π periodicity of angles, the absolute heading angle is stabilized. φ and radar observation perspective α Perform trigonometric transformations.

[0046] ; The sine of the heading angle; ; The cosine of the heading angle; ; The sine value of the observation angle; ; The cosine value of the observation angle.

[0047] (2) Historical characteristics of the trajectory.

[0048] ① Trajectory Age ( ): This indicates the lifespan of the trajectory; a longer trajectory indicates more stable tracking. Counting starts from 0, therefore .

[0049] ② First point identifier: Indicates whether this data point is the first point of the trajectory. If it is the first point, is_first_ i =1, otherwise is_first_i =0.

[0050] ③ Logarithmic Trajectory Age: To make it easier for the model to learn the nonlinear effects of trajectory stability, the trajectory age value is logarithmically transformed, which can smooth the feature distribution.

[0051] = ; Logarithmic age; Ultimately, through the aforementioned preliminary feature engineering, each data point It is transformed into a high-dimensional numerical feature vector , It includes all the derived features mentioned above as well as the original features.

[0052] Understandably, this invention transforms raw data into a physical language that describes how RCS changes. The observation perspective is one of the most crucial variables affecting RCS (e.g., the RCS of a vehicle's side and front differ significantly). By calculating speed and heading, the observation perspective is ultimately obtained, providing the model with the key physical basis for predicting RCS changes. Trajectory history features provide context for target tracking stability.

[0053] In step S12, an offline model is trained based on the high-dimensional feature vector, and a pre-trained RCS generation model is output.

[0054] Step S12 specifically includes: processing the high-dimensional feature vector to obtain an input tensor; inputting the input tensor into the RCS prediction backbone network for feature extraction to obtain a feature representation, and obtaining Gaussian distribution parameters based on the feature representation; calculating the hybrid loss based on the Gaussian distribution parameters and the ordered trajectory sequence; updating the RCS prediction backbone network based on the hybrid loss, and outputting a pre-trained RCS generation model.

[0055] In one possible implementation, the ordered trajectory sequence further includes a target RCS ground truth value, and the Gaussian distribution parameters include the mean and log-standard deviation. Specifically, the step of calculating the mixed loss is implemented as follows: a Gaussian negative log-likelihood loss is calculated based on the mean, the log-standard deviation, and the target RCS ground truth value; a Huber loss is calculated based on the mean, the log-standard deviation, and the target RCS ground truth value; and an uncertainty regularization loss is calculated based on the log-standard deviation; finally, the mixed loss is calculated based on the Gaussian negative log-likelihood loss, the Huber loss, and the uncertainty regularization loss.

[0056] Furthermore, step S203 involves multimodal feature input and embedding (refer to...). Figure 5 Model). (1) The numerical feature vector generated in the previous step. Z-score standardization is performed to transform all features into a standard normal distribution with a mean of 0 and a standard deviation of 1. For the first i The original feature vectors of each data point. (2) The discrete type and The data is mapped to a low-dimensional, dense vector representation through both a category embedding layer and a sensor embedding layer. and .in, Representing the target category, For sensor ID, For category embedding vectors, (3) Concatenate the three to form the final input tensor. : .

[0057] Understandably, multimodal feature inputs and embeddings prepare standardized and semantic inputs for neural networks. The embedding layer allows the model to autonomously learn the similarity and difference relationships of categories such as "car" and "truck" in terms of RCS characteristics, and the concatenation operation integrates information from three different modalities: continuous features (motion state), category semantics, and sensor characteristics.

[0058] Step S204 is the forward propagation of the deep probabilistic neural network. (1) Convert the input tensor Send in a by N The RCS prediction backbone network (103) is composed of stacked residual blocks, and its model system block diagram is as follows: Figure 4 As shown. (The following is a list of numbers / items / etc.) j Each residual block The calculation can be expressed as: ;in (2) High-level features extracted by the backbone network It is fed into two independent linear output heads (104, 105): mean prediction head (104): Uncertainty prediction head (105): (3) The final model output is a Gaussian distribution parameter with an RCS value. See also Figure 4 The network consists of several layers: a ClassEmbedding Layer, which receives discrete class IDs and outputs class feature vectors; a SensorEmbedding Layer, which receives discrete sensor IDs and outputs sensor feature vectors; an RCS prediction backbone, which receives concatenated multimodal features and extracts high-level feature representations through multiple residual blocks; a Mean Head, which predicts the expected value μ of the RCS from the high-level features; and an Uncertainty Head, which predicts the logarithmic standard deviation log(σ) of the RCS from the high-level features. For the features of layer 0, This is the output of the (j-1)th residual block. For the output of the j-th residual block, N The number of residual blocks, This is the output of the last residual block (high-level feature).

[0059] Understandably, step S203 constructs an intelligent core capable of learning complex mapping relationships and quantifying its own predictive uncertainty. The residual network allows for deep model design, resulting in powerful learning capabilities and capturing the complex nonlinear relationship between features and RCS. The probabilistic output recognizes that real-world RCS measurements inherently contain fluctuations and noise; the model not only predicts the value but also how certain that prediction is. This lays the foundation for subsequent simulations of real-world fluctuations.

[0060] Step S205 involves the design of the loss function and optimization strategy. (1) Loss function design. In order to perform end-to-end optimization of the model, this invention adopts a hybrid loss function. . Among them, Gaussian negative log-likelihood loss As shown below, this loss function drives the model to learn the entire RCS probability distribution. Huber's losses As a robust regularization term for mean prediction, it is insensitive to outliers. Its weight: ; Uncertainty Regularization Loss To prevent the model from predicting meaningless maximum uncertainty in order to avoid the penalty of Gaussian negative log-likelihood loss, Its weight: (2) Optimization strategy design. This invention uses the AdamW optimizer and combines the OneCycle learning rate scheduling strategy and the uncertainty warm-up strategy to train the model, and finally obtains a convergent pre-trained RCS generation model (S100) that can be used for online generation.

[0061] Understandably, this model S100 encapsulates the "mapping pattern from scene state to RCS probability distribution" learned from massive amounts of real data, becoming the engine for efficiently generating high-fidelity RCS data in the second stage of online simulation.

[0062] In step S13, the true state information of all targets in the simulation scenario is obtained, and the target feature vector is obtained based on the true state information, wherein the feature format of the target feature vector is consistent with that of the high-dimensional feature vector.

[0063] It is worth noting that, see Figure 5 Steps 13 and S14 constitute the second stage: online simulation to generate RCS data.

[0064] Step S13 specifically includes: calculating simulation dynamic features and simulation history features based on the truth state information; combining the truth state information, the simulation dynamic features, and the simulation history features into feature vectors to obtain a target feature vector.

[0065] like Figure 5 As shown, S100 is the pre-trained deep learning RCS prediction model, which is the core algorithm module of this invention and has been trained offline. S101 is the simulation platform state extraction module. It calls the simulation engine API to obtain and output the ground truth state of the simulation target. S102 is the online feature synthesis module. This module constructs the model input in real time based on the target state information in the scene obtained by S101. S103 is the RCS model forward inference module. This module calls the pre-trained RCS prediction model in S100 to make predictions. S104 is the probabilistic RCS generation and application module. This module samples and generates the final RCS value and sends it to the millimeter-wave radar object model as the simulation output of the RCS value. Simulation engine: the core of the entire simulation platform; millimeter-wave radar object model: the module in the simulation platform responsible for simulating millimeter-wave radar modeling and calculating simulation output.

[0066] Furthermore, steps S101 and S102 involve state extraction and online feature synthesis for the simulation platform. The simulation engine is responsible for managing all targets and scheduling all simulation models within the simulation scenario. At each time step of the simulation platform... For each dynamic target object `obj` in the simulation scenario, its ground truth state information is extracted through the simulation engine API, and a feature vector consistent with the offline training format is synthesized in real time. .

[0067] In step S14, radar RCS data is obtained based on the pre-trained RCS generation model and the target feature vector.

[0068] Step S14 specifically includes: inputting the target feature vector into the pre-trained RCS generation model and outputting the RCS prediction distribution parameters under simulation conditions; sampling according to the RCS prediction distribution parameters to obtain radar RCS data.

[0069] Further, step S103 is the forward inference of the RCS model. And the corresponding target category ID and sensor ID, generated according to the basic steps of S203. The input is fed into the preloaded RCS generative model (S100), and forward inference is performed once to obtain the output. It should be noted here that the millimeter-wave radar sensor IDs and deployments in the simulation platform must be consistent with the sensor deployments of the vehicles collecting data in real-world driving scenarios.

[0070] Step S104 involves probabilistic RCS generation and application. This is based on a Gaussian distribution. A random sample is taken to obtain the final RCS value, as follows: ;in N (0,1) is a random sample from a standard normal distribution. This high-fidelity... The value is then fed into the radar sensor model of the simulation platform for subsequent signal-to-noise ratio calculation and point cloud generation, or it can be directly output as the attribute feature of obj.

[0071] In this embodiment, the present invention uses a data-driven, pre-trained probabilistic deep learning model to replace traditional physical simulation calculations or simplified lookup tables, generating high-fidelity RCS data with physical randomness in real time and efficiently during the simulation loop. In the simulation, the probability distribution output by the model is used for sampling to generate RCS values, injecting randomness consistent with real physical laws into the simulated radar signal within a data-driven framework. Simultaneously, instantaneous dynamic kinematic information (velocity, viewpoint) and temporal trajectory history information (trajectory age) are considered, enabling the generated RCS to reflect the target's motion behavior and state history.

[0072] Next, referring to the accompanying drawings, a high-fidelity radar RCS prediction system for real-time simulation of autonomous driving proposed according to the embodiments of this application is described, which is used to implement the high-fidelity radar RCS prediction method for real-time simulation of autonomous driving as described in any of the above schemes.

[0073] Figure 6 This is a structural diagram of a high-fidelity radar RCS prediction system for real-time simulation of autonomous driving, according to an embodiment of this application.

[0074] like Figure 6 As shown, the high-fidelity radar RCS prediction system for real-time simulation of autonomous driving includes: a data acquisition and feature synthesis module 100, an offline model training module 200, a simulation state extraction and feature synthesis module 300, and a radar RCS data prediction output module 400.

[0075] Specifically, the data acquisition and feature synthesis module 100 is used to acquire radar data from multiple radars in real road driving scenarios, and process the radar data to obtain high-dimensional feature vectors. The offline model training module 200 is used to train an offline model based on the high-dimensional feature vector and output a pre-trained RCS generation model. The simulation state extraction and feature synthesis module 300 is used to obtain the true state information of all targets in the simulation scene, and obtain the target feature vector based on the true state information, wherein the target feature vector has the same feature format as the high-dimensional feature vector; The radar RCS data prediction output module 400 is used to obtain radar RCS data based on the pre-trained RCS generation model and the target feature vector.

[0076] Figure 7 A structural diagram of a terminal provided in an embodiment of this application. The terminal may include: The memory 501, the processor 502, and the computer program stored on the memory 501 and capable of running on the processor 502.

[0077] When processor 502 executes the program, it implements the high-fidelity radar RCS prediction method for real-time simulation of autonomous driving provided in the above embodiments.

[0078] Furthermore, the terminal also includes: Communication interface 503 is used for communication between memory 501 and processor 502.

[0079] The memory 501 is used to store computer programs that can run on the processor 502.

[0080] Memory 501 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0081] If the memory 501, processor 502, and communication interface 503 are implemented independently, then the communication interface 503, memory 501, and processor 502 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EIS) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 7 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0082] Optionally, in a specific implementation, if the memory 501, processor 502, and communication interface 503 are integrated on a single chip, then the memory 501, processor 502, and communication interface 503 can communicate with each other through an internal interface.

[0083] Processor 502 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0084] This embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the high-fidelity radar RCS prediction method for real-time simulation of autonomous driving as described above.

[0085] One embodiment of this application provides a computer program product, including a computer program that, when executed by a processor, implements the features described in this application. Figure 1 The corresponding embodiments provide a high-fidelity radar RCS prediction method for real-time simulation of autonomous driving.

[0086] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0087] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0088] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0089] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable storage medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable storage medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable storage medium could be paper or other suitable media on which the program can be printed, since the program can be obtained electronically by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0090] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0091] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it includes one or a combination of the steps of the method embodiments.

[0092] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0093] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

[0094] It should be understood that the application of this application is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A high-fidelity radar RCS prediction method for real-time simulation of autonomous driving, characterized in that, The high-fidelity radar RCS prediction method for real-time simulation of autonomous driving includes: The radar data of multiple radars in real road driving scenarios are acquired, and the radar data is processed to obtain a high-dimensional feature vector. The offline model is trained based on the high-dimensional feature vectors, and the pre-trained RCS generation model is output. Obtain the truth state information of all targets in the simulation scenario, and obtain the target feature vector based on the truth state information, wherein the target feature vector has the same feature format as the high-dimensional feature vector; Based on the pre-trained RCS generation model, radar RCS data is obtained according to the target feature vector.

2. The high-fidelity radar RCS prediction method for real-time simulation of autonomous driving according to claim 1, characterized in that, The process of processing the radar data to obtain a high-dimensional feature vector specifically includes: The radar data is grouped to obtain multiple data point groups. The multiple data point groups are then filtered for trajectory validity to obtain filtered trajectory groups. Finally, the filtered trajectory groups are sorted in time sequence to obtain an ordered trajectory sequence. The dynamic kinematic features and trajectory history features are calculated based on the ordered trajectory sequence, and the ordered trajectory sequence, the dynamic kinematic features and the trajectory history features are combined into feature vectors to obtain a high-dimensional feature vector.

3. The high-fidelity radar RCS prediction method for real-time simulation of autonomous driving according to claim 2, characterized in that, The ordered trajectory sequence includes timestamps, locations, azimuth angles, and trajectory indexes for multiple data points; The calculation of dynamic kinematic features and trajectory history features based on the ordered trajectory sequence specifically includes: The heading angle is calculated based on the timestamps, locations, and azimuth angles of multiple data points, and the observation angle is calculated based on the heading angle. Trigonometric transformations are performed on the heading angle and the observation angle to obtain the dynamic kinematic characteristics; The trajectory age is calculated based on the trajectory index, and the trajectory age is transformed to obtain the trajectory historical features.

4. The high-fidelity radar RCS prediction method for real-time simulation of autonomous driving according to claim 2, characterized in that, The step of training an offline model based on the high-dimensional feature vector and outputting a pre-trained RCS generation model specifically includes: The high-dimensional feature vector is processed to obtain the input tensor; The input tensor is input into the RCS prediction backbone network for feature extraction to obtain feature representation, and Gaussian distribution parameters are obtained based on the feature representation; Calculate the hybrid loss based on the Gaussian distribution parameters and the ordered trajectory sequence; The RCS prediction backbone network is updated based on the hybrid loss to output a pre-trained RCS generation model.

5. The high-fidelity radar RCS prediction method for real-time simulation of autonomous driving according to claim 4, characterized in that, The ordered trajectory sequence also includes the target RCS true value, and the Gaussian distribution parameters include the mean and the log-standard deviation; The calculation of the hybrid loss based on the Gaussian distribution parameters and the ordered trajectory sequence specifically includes: Based on the mean, the logarithmic standard deviation, and the target RCS true value, calculate the Gaussian negative log-likelihood loss; based on the mean, the logarithmic standard deviation, and the target RCS true value, calculate the Huber loss; and based on the logarithmic standard deviation, calculate the uncertainty regularization loss. The mixed loss is calculated based on the Gaussian negative log-likelihood loss, the Huber loss, and the uncertainty regularization loss.

6. The high-fidelity radar RCS prediction method for real-time simulation of autonomous driving according to claim 4, characterized in that, The step of obtaining the target feature vector based on the truth state information specifically includes: Calculate the simulation dynamic features and simulation history features based on the truth state information; The target feature vector is obtained by combining the truth state information, the simulation dynamic features, and the simulation history features into feature vectors.

7. The high-fidelity radar RCS prediction method for real-time simulation of autonomous driving according to claim 6, characterized in that, The step of obtaining radar RCS data based on the pre-trained RCS generation model and the target feature vector specifically includes: The target feature vector is input into the pre-trained RCS generation model, and the RCS prediction distribution parameters under simulation conditions are output. Radar RCS data is obtained by sampling based on the RCS prediction distribution parameters.

8. A high-fidelity radar RCS prediction system for real-time simulation of autonomous driving, characterized in that, The high-fidelity radar RCS prediction system for real-time simulation of autonomous driving is used to implement the high-fidelity radar RCS prediction method for real-time simulation of autonomous driving as described in any one of claims 1-7, wherein the high-fidelity radar RCS prediction system for real-time simulation of autonomous driving includes: The data acquisition and feature synthesis module is used to acquire radar data from multiple radars in real road driving scenarios and process the radar data to obtain high-dimensional feature vectors. The offline model training module is used to train an offline model based on the high-dimensional feature vector and output a pre-trained RCS generation model. The simulation state extraction and feature synthesis module is used to obtain the true state information of all targets in the simulation scene, and to obtain the target feature vector based on the true state information, wherein the target feature vector has the same feature format as the high-dimensional feature vector; The radar RCS data prediction output module is used to obtain radar RCS data based on the pre-trained RCS generation model and the target feature vector.

9. A terminal, characterized in that, The terminal includes: a memory, a processor, and a high-fidelity radar RCS prediction program for real-time simulation of autonomous driving stored in the memory and executable on the processor. When the high-fidelity radar RCS prediction program for real-time simulation of autonomous driving is executed by the processor, it implements the steps of the high-fidelity radar RCS prediction method for real-time simulation of autonomous driving as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a high-fidelity radar RCS prediction program for real-time simulation of autonomous driving. When the high-fidelity radar RCS prediction program for real-time simulation of autonomous driving is executed by a processor, it implements the steps of the high-fidelity radar RCS prediction method for real-time simulation of autonomous driving as described in any one of claims 1-7.

Images