Simulation vehicle-mounted sensing equipment data correction method and device

Abstract

The invention discloses a simulation vehicle-mounted sensing equipment data correction method and device. The automatic scene generation method based on the railway topological relation comprises the following steps: correcting camera sensor data in a simulation environment; and correcting the radar sensor data in the simulation environment. In the process of constructing the virtual simulation platform oriented to the rail transit industry based on the physical engine, the dynamic characteristics of high-speed movement of the train are considered, a data correction method of the train-mounted camera and the laser radar is introduced, and the authenticity of sensor data acquisition is ensured.

Description

Technical Field

[0001] The present application relates to the technical field of perception device data correction, and in particular to a simulated vehicle-mounted perception device data correction method and a simulated vehicle-mounted perception device data correction device. Background Art

[0002] The Train Autonomous Perception System (TAPS), a key technology for automated train operation, plays a vital role in improving railway transportation efficiency and ensuring operational safety. Train autonomous perception can generally be categorized as active or passive. The former relies on onboard sensors and processing terminals to acquire road condition information and output corresponding control decisions, while the latter relies on trackside sensors, communication modules, and cloud-based processors to transmit decision instructions to the train. The development of active perception technology is crucial for achieving fully autonomous train operation (FAO). This development relies heavily on large amounts of sensor data for training and testing. However, data collection in real-world scenarios in the rail sector faces challenges such as high cost, low coverage, and diverse weather conditions, leading to long R&D cycles and insufficient completeness. Simulation-based data acquisition, with its efficiency, cost-effectiveness, and wide-area coverage, can alleviate these challenges.

[0003] Currently, there is no mature simulation platform in China that can achieve the above functions. This solution builds a virtual simulation platform for the rail transit industry based on a physical engine to achieve sensor data acquisition under multiple working conditions, and introduces a sensor data correction method considering the dynamic factors of high-speed train operation.

[0004] Therefore, it is hoped that there will be a technical solution to solve or at least alleviate the above-mentioned deficiencies in the prior art. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for correcting data of a simulated vehicle-mounted sensing device to at least solve one of the above-mentioned technical problems.

[0006] The present invention provides the following solutions:

[0007] According to one aspect of the present invention, a method for correcting data of a simulated vehicle-mounted perception device is provided, which is used in a virtual simulation platform, wherein the method corrects the camera sensor data in a simulation environment;

[0008] Correct radar sensor data in a simulation environment.

[0009] Optionally, the correcting the camera sensor data in the simulation environment includes:

[0010] Perform noise correction on the data transmitted by the simulated camera;

[0011] Perform motion blur simulation on the data transmitted by the simulated camera;

[0012] Dynamically expose the data transmitted by the simulated camera.

[0013] Optionally, performing noise correction on the data transmitted by the simulation camera includes:

[0014] Acquire an original data set, wherein the original data set includes paired data of a noise-free image and a real noise image;

[0015] Performing illumination intensity estimation on each noise-free image in the training set to obtain a calibrated noise-free image, wherein the calibrated noise-free image and the real noise image constitute the training set;

[0016] Obtain a noise generation model;

[0017] Training the noise generation model using the training set to obtain a trained noise generation model;

[0018] The generator in the trained noise generation model is embedded in the virtual simulation engine, which receives the noise-free image of the simulated light intensity signal and outputs the noise image in real time.

[0019] Optionally, performing motion blur simulation on the data transmitted by the simulation camera includes:

[0020] Establish the position model and posture model of the vehicle at the end of exposure;

[0021] Get the total displacement;

[0022] Perform displacement decomposition and vibration simulation on the total displacement to correct the sensor position;

[0023] An intermediate image is rendered at each corrected sensor position, containing the scene characteristics within the sensor, and an exponential decay model is used to superimpose the intermediate images to generate the final blurred image.

[0024] Optionally, dynamically exposing the data transmitted by the simulation camera includes:

[0025] Get the initial image;

[0026] Obtaining a current exposure value according to a histogram of the initial image;

[0027] Get exposure compensation value according to exposure value;

[0028] Apply the exposure compensation value to the original image to produce a corrected image.

[0029] Optionally, the correcting the radar sensor data in the simulation environment includes:

[0030] Dynamic generation of radar point cloud motion distortion through IMU data;

[0031] Achieve high-fidelity radar point cloud loss simulation through mathematical modeling.

[0032] Optionally, the dynamically generating radar point cloud motion distortion using IMU data includes:

[0033] Get the original point cloud data;

[0034] Get the coordinate offset value at a fixed time interval;

[0035] Get the radar scanning time interval;

[0036] Point cloud correction data is obtained according to the original point cloud data, the coordinate offset value and the radar scanning time interval.

[0037] Optionally, the high-fidelity radar point cloud loss simulation achieved through mathematical modeling includes:

[0038] Get the original point cloud data;

[0039] Processing the original point cloud data by randomly discarding the original point cloud data, thereby obtaining point cloud data after random discarding;

[0040] The original point cloud data after random discarding is filtered by intensity filtering to obtain filtered point cloud data.

[0041] The present application also provides a simulated vehicle-mounted sensing device data correction device, the simulated vehicle-mounted sensing device data correction device comprising:

[0042] A camera sensor data correction module, wherein the camera sensor data correction module is used to correct the camera sensor data in the simulation environment;

[0043] A radar sensor data correction module is used to correct radar sensor data in a simulation environment.

[0044] In the process of building a virtual simulation platform for the rail transit industry based on a physics engine, this application takes into account the dynamic characteristics of high-speed movement of trains and introduces a data correction method for train-mounted cameras and lidars to ensure the authenticity of sensor data acquisition. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flowchart of a method for correcting data of a simulated vehicle-mounted sensing device in one embodiment of the present application.

[0046] Figure 2Schematic diagram of the noise image simulation process of the generative adversarial network in one embodiment of the present application.

[0047] Figure 3 Schematic diagram of noise image verification and comparison in one embodiment of the present application.

[0048] Figure 4 Schematic diagram of motion blur simulation based on dynamic displacement decomposition in one embodiment of the present application.

[0049] Figure 5 It is a schematic diagram of dynamic exposure in one embodiment of the present application.

[0050] Figure 6 Schematic diagram of the radar point cloud motion distortion correction principle in one embodiment of the present application. DETAILED DESCRIPTION

[0051] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0052] like Figure 1 The data correction method for the simulated vehicle-mounted perception device shown is used for a virtual simulation platform. The automatic scene generation method based on railway topology relationships includes:

[0053] Correct the camera sensor data in the simulation environment;

[0054] Correct radar sensor data in a simulation environment.

[0055] In this embodiment, the correction of camera sensor data in the simulation environment includes:

[0056] Perform noise correction on the data transmitted by the simulated camera;

[0057] Perform motion blur simulation on the data transmitted by the simulated camera;

[0058] Dynamically expose the data transmitted by the simulated camera.

[0059] See also Figure 2 In this embodiment, performing noise correction on the data transmitted by the simulation camera includes:

[0060] Acquire an original data set, wherein the original data set includes paired data of a noise-free image and a real noise image;

[0061] Performing illumination intensity estimation on each noise-free image in the training set to obtain a calibrated noise-free image, wherein the calibrated noise-free image and the real noise image constitute the training set;

[0062] Obtain a noise generation model;

[0063] Training the noise generation model using the training set to obtain a trained noise generation model;

[0064] The generator in the trained noise generation model is embedded in the virtual simulation engine, which receives the noise-free image of the simulated light intensity signal and outputs the noise image in real time.

[0065] When modeling using virtual simulation platforms, noise is generated during the camera imaging process by a combination of hardware randomness (such as shot noise, electronic noise, and gain noise) and signal processing (analog-to-digital conversion, gamma correction, and compression encoding). Existing methods struggle to accurately model the noise characteristics of different sensors at the imaging mechanism level, and are particularly unable to dynamically simulate the noise degradation effects under varying lighting conditions. Furthermore, existing noise simulation techniques are often based on fixed parameter models (such as Gaussian noise superposition) and lack the ability to model the nonlinearities of real-world noise distributions. Supervised learning-based methods rely on one-to-one noisy and noise-free image pairs, resulting in poor generalization and inability to adapt to the dynamic lighting scenarios required in virtual simulation environments.

[0066] This application uses the generative adversarial network (GAN) infrastructure and the image denoising open source dataset (SIDD) to train an end-to-end noise generation network under different ISOs to generate highly credible noise images in a virtual environment. It solves the problems of poor adaptability to lighting conditions and inaccurate noise distribution modeling in existing noise simulation methods, and provides an end-to-end generation method that can be embedded in a virtual simulation platform to dynamically generate noise images under different lighting scenarios.

[0067] In this embodiment, a public noise dataset (such as SIDD and PolyU) is used to construct a training set, which includes paired data of noise-free images and real noise images.

[0068] In this embodiment, estimating the illumination intensity of each noise-free image in the training set to obtain a calibrated noise-free image includes:

[0069] Each noise-free image is processed as follows:

[0070] Generate a brightness distribution map through grayscale histogram equalization processing;

[0071] Combine the camera ISO parameters and calibration factors to convert the average brightness into light intensity values;

[0072] The calibration images are collected under a constant illumination scene (such as 1000 Lux) and the calibration factors are optimized using linear regression.

[0073] In this embodiment, the noise generation model includes a generator and a discriminator, as follows:

[0074] Generator G design:

[0075] The input is a noise-free image x, a random noise variable z~N(0,σ 2 ), and the light intensity w, outputs a pseudo-noise image The goal is to approximate the true noise distribution p(y|x,w), that is, to satisfy:

[0076] p G(y∣x,z,w) →p(y∣x,w)pG(y∣x,z,w)→p(y∣x,w);

[0077] Discriminator D design:

[0078] The input is an image pair (x,y) or, The output is the probability of true or false discrimination, and the goal is to distinguish real noise images from generated images.

[0079] In this embodiment, the noise generation model is trained using the following method:

[0080] Through Bayesian inference, the distribution p will be generated G (x,y) is approximately the true joint distribution p(x,y), that is:

[0081]

[0082] where z i is the noise variable sampling, and L is the number of sampling times.

[0083] Adversarial training objective function:

[0084] min G max D L GAN (G,D)=

[0085] E {(x,y)～p(x,y)} [D(x,y)]-E {z～N(0,σ2),x～p(x)} [D(x,y^)];

[0086] Training optimization: An alternating update strategy is used to balance the generator and discriminator, and the gradient penalty technology of WGAN-GP is introduced to improve training stability.

[0087] The trained generator G is embedded in the virtual simulation engine, which receives the simulated light intensity signal w and the noise-free image x and outputs the noise image in real time.

[0088] The above method has the following advantages:

[0089] (1) Dynamic illumination adaptability: By introducing the light intensity variable, the model can simulate the noise degradation law under different lighting conditions; (2) High-fidelity noise generation: Utilizing the nonlinear fitting capability of GAN, the generated noise distribution is closer to the real sensor characteristics; (3) End-to-end scalability: It can be directly integrated into the virtual simulation platform and supports real-time noise rendering.

[0090] See Figure 3. Figure 3 To obtain the image result by the method of the present application, in this embodiment, the present application performs the following steps:

[0091] Use the SIDD dataset to calibrate the light intensity;

[0092] Build a U-Net structure generator and PatchGAN discriminator;

[0093] Set training parameters: batch size 1, learning rate 1e-4, and Adam optimizer;

[0094] The generator is updated once after every 5 training of the discriminator, and a gradient penalty coefficient λ = 10 is applied.

[0095] See also Figure 3 , Figure 3 For virtual driving simulation applications, a nighttime low-light scene (w = 50 Lux) is rendered in UE. The simulated images x and w are input into the generator, and a noisy image is output. Verify the consistency of the degradation effect of noisy images in target detection algorithms.

[0096] See also Figure 4 , Figure 4 (a) is the traditional camera dynamic imaging process, showing the light accumulation effect during the exposure time, and (b) is the blur generation architecture proposed in this invention, which includes 6-DoF motion decomposition, random jitter injection and multi-frame images.

[0097] In this embodiment, performing motion blur simulation on the data transmitted by the simulation camera includes:

[0098] Establish the position model and posture model of the vehicle at the end of exposure;

[0099] Get the total displacement;

[0100] Perform displacement decomposition and vibration simulation on the total displacement to correct the sensor position;

[0101] An intermediate image is rendered at each corrected sensor position, containing the scene characteristics within the sensor, and an exponential decay model is used to superimpose the intermediate images to generate the final blurred image.

[0102] In the existing technology, during the camera imaging process, the high-speed movement of the vehicle causes the continuous accumulation of scene light on the sensor, forming a dynamic blur effect. Traditional image simulation methods are based on static scene rendering and cannot simulate the continuous displacement of objects during the exposure time, resulting in distorted blur effects. Existing blur simulation methods mostly use post-processing filtering (such as Gaussian blur) or interpolation based on adjacent frames, but such methods cannot accurately reflect the physical process of sensor imaging, especially in high-speed motion scenes, where the errors are significant. In addition, the existing technology does not take into account the random displacement caused by vehicle vibration, resulting in differences between the blur morphology and the actual sensor data.

[0103] The core of this application is to achieve physically realistic dynamic blur generation by accurately tracking the vehicle's motion trajectory and decomposing the displacement sub-steps, combined with sensor vibration simulation.

[0104] In this embodiment, establishing the position model and posture model of the vehicle at the end of exposure includes:

[0105] Real-time acquisition of 6-DoF information of the vehicle in the virtual environment, including its position and attitude angle and the instantaneous linear velocity v i =(v x ,v y ,v z ) and angular velocity\omega t =(P x ,Y y ,R z ). According to the exposure time Predicting the vehicle's position and posture at the end of exposure; wherein the position model includes:

[0106]

[0107] The attitude angle model includes:

[0108]

[0109] in,

[0110] For location information, is the attitude angle, instantaneous linear velocity v i =(v x ,v y ,v z ) and angular velocity\omega t =(Px ,Y y ,R z ).

[0111] In this embodiment, performing displacement decomposition and vibration simulation on the total displacement to correct the sensor position includes:

[0112] The total displacement Decompose into N sub-steps and calculate the displacement gain of each step Introducing random jitter in each displacement step Simulate vehicle vibration effects and correct sensor positions:

[0113] In this embodiment, at each step, the decomposition position x n Rendering the intermediate image I n , including sensor internal parameters (focal length F, photosensitive element height H) and scene characteristics. The exponential decay model is used to superimpose the intermediate images to generate the final blurred image I b :

[0114] in,

[0115] I b is the final blurred image, I n is the intermediate image, T e is the exposure time and N is the number of decomposition steps.

[0116] This model can simulate the asymptotic saturation characteristics of the sensor's photosensitive elements.

[0117] It is understandable that the present application can provide exposure time T e , decomposition steps N, jitter intensity J n Adjustment interface to adapt to different sensor models and sports scene requirements.

[0118] The above method has the following advantages:

[0119] (1) High-precision motion modeling: 6-DoF trajectory decomposition accurately restores the continuous changes in sensor position during exposure. (2) Realistic vibration effects: Random jitter is introduced to make the blurred edges exhibit non-uniform diffusion characteristics, closer to real data. (3) Computational efficiency optimization: By controlling the number of decomposition steps N, the rendering quality and real-time requirements are balanced.

[0120] See also Figure 4 , Figure 5 The two figures above show the dynamic exposure control mechanism: Capture the initial image I and analyze its histogram H I And calculate the current exposure value The following two figures show the exposure compensation process: Calculating the exposure difference And adjust the exposure compensation value EC according to the dynamic curve to generate the corrected image I c .

[0121] In this embodiment, dynamically exposing the data transmitted by the simulation camera includes:

[0122] Get the initial image;

[0123] Obtaining a current exposure value according to a histogram of the initial image;

[0124] Get exposure compensation value according to exposure value;

[0125] Apply the exposure compensation value to the original image to produce a corrected image.

[0126] During high-speed train operation, complex and changing lighting conditions (such as day-night transitions and tunnel entrances and exits) often lead to rapid changes in ambient light intensity, which in turn causes delays in adjusting camera exposure parameters. This delay can result in over- or underexposure, affecting image quality and the accuracy of perception algorithms. Furthermore, the dynamic range of the scene also affects exposure accuracy. A large dynamic range can lead to loss of image detail, especially in high-contrast scenes. Therefore, precise dynamic exposure control is crucial in dynamic lighting environments.

[0127] This application analyzes the image histogram in real time, dynamically calculates the exposure compensation value, and adjusts the dynamic range in combination with the nonlinear characteristics of ISO to achieve precise exposure control under complex lighting conditions.

[0128] In this embodiment, obtaining the current exposure value according to the histogram of the initial image includes:

[0129] Camera in time Capture an initial image I; analyze the histogram H of image I I , calculate the current exposure value

[0130] In this embodiment, obtaining the exposure compensation value according to the exposure value includes:

[0131] Calculate target exposure value With the current exposure value The difference: Adjust the exposure compensation value EC according to the dynamic curve: EC=f(ΔT e );in,

[0132] The dynamic curve adopts the following formula:

[0133]

[0134] Among them, E(t) is the dynamic curve, Es 、E k 、E b are the compensation values of saturation state, transition state and initial state respectively, t k is the transition time, α is the decay rate, and H is the step function.

[0135] In this embodiment, applying the exposure compensation value to the initial image to generate the corrected image includes:

[0136] Apply the exposure compensation value EC to the initial image I to generate the corrected image I c ; in time Output corrected image I c .

[0137] Based on the nonlinear characteristics of ISO, it simulates the dynamic range limitation of the sensor; by adjusting the ISO setting, it limits the brightness of the highlight area while preserving the details in the shadow area.

[0138] The above method has the following advantages:

[0139] (1) Real-time exposure adjustment: Through histogram analysis and dynamic compensation, rapid exposure adjustment is achieved under complex lighting conditions; (2) Dynamic range optimization: Based on the nonlinear characteristics of ISO, the dynamic range of the image is effectively controlled to avoid overexposure or underexposure; (3) High adaptability: Suitable for dynamic lighting scenes such as high-speed railways and autonomous driving, ensuring image quality and algorithm accuracy.

[0140] In this embodiment, the simulated vehicle-mounted perception device of the present application also includes a radar sensor.

[0141] In the prior art, the existing radar system is easily disturbed by its own motion when the vehicle is moving at high speed, resulting in translation, rotation and deformation distortion of the point cloud data. After the radar point cloud is collected, the ray is emitted through the emission point, and then reflected to the receiving point after colliding with the detection object. An optical path is formed between the emission point and the receiving point, and it takes time for the light to be transmitted on the optical path. When the train is moving at high speed, the received point cloud will usually have micro-distortion compared with the stationary state. The method of the radar simulation system based on computing power usually calculates the point cloud situation in a static scene in a certain running frame, which does not contain the above-mentioned non-idealized features. The data generated in this mode is different from the real data. In the simulation environment, this type of sensor has the situation of no motion distortion and point cloud loss. This application takes into account the dynamic characteristics of the high-speed movement of the train, and introduces a simulated radar motion distortion model and a simulated radar point cloud loss model in combination with speed information to correct the radar sensor data.

[0142] This application makes the following improvements to the above situation:

[0143] Correction of radar sensor data in a simulation environment includes:

[0144] Dynamic generation of radar point cloud motion distortion through IMU data;

[0145] Achieve high-fidelity radar point cloud loss simulation through mathematical modeling.

[0146] See also Figure 6 In this embodiment, the dynamic generation of radar point cloud motion distortion using IMU data includes:

[0147] Get the original point cloud data;

[0148] Get the coordinate offset value at a fixed time interval;

[0149] Get the radar scanning time interval;

[0150] Point cloud correction data is obtained according to the original point cloud data, the coordinate offset value and the radar scanning time interval.

[0151] Specifically, first collect the IMU raw data, including yaw angle, pitch angle, heading angle, vehicle linear velocity and angular velocity; obtain the radar initial point cloud, including spatial coordinates (x, y, z) and reflection intensity values.

[0152] In this embodiment, this application needs to unify the relationship between the world coordinate system, the vehicle coordinate system, and the radar coordinate system. The specific method is as follows:

[0153] 1) Define the reference coordinate system. The origin of the world coordinate system (WCS) is the radar installation base point. The origin of the vehicle coordinate system (VCS) is the center point of the vehicle. The x-axis is to the right of the vehicle's direction of travel, the y-axis is in the direction of travel, and the z-axis is toward the ground. The radar coordinate system (SCS) is consistent with the radar hardware definition. 2) Calculate the initial rotation matrix R rot , perform the following rotation transformations in order: generate R around the z, y, and x axes respectively Zrot 、R Yrot 、R Xrot , the final combination is the total rotation matrix R rot =R Zrot ×R Yrot ×R Xrot .

[0154] In this embodiment, the Euler angle sequence (heading → pitch → yaw) is used to achieve three-axis rotation superposition:

[0155]

[0156] The rotation matrices of each component (yaw angle, pitch angle, heading angle) are defined as follows:

[0157]

[0158] In this embodiment, it is necessary to obtain the motion information of the object in the simulation environment, specifically, to obtain the motion speed of the object under the WCS. and calculate the unit velocity vector Calculate V from IMU data egoW To estimate the vehicle's motion state, the relative speed is synthesized based on the above information. The specific process is as follows:

[0159]

[0160] In this embodiment, the moving speed of the object under WCS is Obtain it through the following methods:

[0161]

[0162] In this embodiment, the combined velocity field is calculated as follows:

[0163]

[0164]

[0165] In this embodiment, the point cloud correction data is obtained according to the original point cloud data, the coordinate offset value and the radar scanning time interval by the following formula:

[0166] in,

[0167] R rot For the transformation matrix, the three types of coordinate axes are unified interpersonal relationships, The ideal yaw angle, pitch angle, and heading angle output by the onboard IMU, and S egoV are the unit vectors of the velocity direction of the environment objects except the vehicle in the world coordinate system, the velocity direction of the vehicle in the world coordinate system, and the direction of vehicle movement, respectively. and They are the speed and direction of the objects in the vehicle's external environment in the world coordinate system, the speed and direction of the vehicle in the world coordinate system, and the speed and direction of the vehicle's moving direction. is the coordinate offset value at a fixed time interval, τ is the radar scanning time interval, D S and D S ′ is the original point cloud coordinate data and the corrected data.

[0168] This method dynamically generates radar point cloud motion distortion using IMU data, making it suitable for radar point cloud acquisition scenarios in high-speed train environments. By incorporating spatial and temporal information, this method addresses the problem of overly idealized simulated point clouds.

[0169] Existing technologies, given the large number of point clouds, are difficult to implement point-by-point correction within the constraints of existing hardware. Furthermore, traditional interpolation methods cannot accurately simulate the point cloud sparsification caused by factors such as the Doppler effect and signal occlusion in real-world scenarios.

[0170] In this embodiment, high-fidelity radar point cloud loss simulation is achieved through mathematical modeling, including:

[0171] Get the original point cloud data;

[0172] Processing the original point cloud data by randomly discarding the original point cloud data, thereby obtaining point cloud data after random discarding;

[0173] The original point cloud data after random discarding is filtered by intensity filtering to obtain filtered point cloud data.

[0174] This section proposes a dual-factor probabilistic dropout mechanism, achieving high-fidelity point cloud loss simulation through mathematical modeling. This model incorporates a probabilistic strategy of randomized dropout and threshold control based on intensity decay to simulate point cloud loss. Randomized dropout randomly discards a certain percentage of the point cloud (e.g., 10%) to simulate factors such as the Doppler effect. The intensity decay threshold, based on the intensity decay factor in the engine's internal raycasting model, discards points that are too far from the radar.

[0175] In this embodiment, the original point cloud data includes spatial coordinates (x, y, z) and reflection intensity values.

[0176] In this embodiment, the raw point cloud data is processed by random discarding to obtain point cloud data after random discarding, which includes:

[0177] 1) Calculate the discard probability: Dynamically generate the discard probability P based on preset parameters (such as scene type, vehicle speed) drop ; 2) Random judgment: Generate a random r~U(0,1), if r <P drop , then discard the point cloud and enter the end process; otherwise, keep it and go to the next step.

[0178] In this embodiment, a dynamic random drop algorithm is used for random drop, wherein the dynamic random drop algorithm is specifically as follows:

[0179] Introducing speed-related parameters v ego and wavelength λ, construct the Doppler frequency shift probability model:

[0180]

[0181] where f doppler =2v ego v obj cosθ / λ,v obj is the target speed, θ is the angle between the radar and the target, and k is the scene adaptation coefficient, which defaults to 0.1.

[0182] Implementation steps: (a) Calculate the Doppler frequency deviation of all point clouds; (b) Generate random numbers r~Beta (α=2, β=5) that obey Beta distribution; (c) When r <P drop Mark the point as lost.

[0183] In this embodiment, the original point cloud data after random discarding is filtered by intensity filtering, so that the filtered point cloud data obtained includes:

[0184] 1) Calculate the distance threshold: Based on the radar operating parameters (transmit power, wavelength, antenna gain) and the environmental propagation model, calculate the intensity attenuation threshold for the current scenario. 2) Intensity comparison: If the point cloud reflection intensity (η is the signal-to-noise ratio compensation factor) is high, the point cloud is retained; otherwise, it is discarded and the process ends.

[0185] In this embodiment, a distance-dependent threshold function is established to calculate the intensity attenuation threshold in the current scene. The distance-dependent threshold function is as follows:

[0186]

[0187] Where d is the distance from the point cloud to the radar, G is the antenna gain, L includes propagation loss terms such as atmospheric absorption and Rayleigh scattering, and C env is the environmental coefficient (the value range for open road scenes is 0.6-0.9).

[0188] The present application further includes: generating a valid point cloud library, storing the point cloud data after double screening into the valid point cloud library for subsequent algorithm processing.

[0189] The present application also provides a simulated vehicle-mounted perception device data correction device, which includes a camera sensor data correction module and a radar sensor data correction module, wherein:

[0190] The camera sensor data correction module is used to correct the camera sensor data in the simulation environment;

[0191] The radar sensor data correction module is used to correct radar sensor data in a simulation environment.

[0192] In this embodiment, the method of the present application is used for a virtual simulation platform for the rail transit industry. The virtual simulation platform for the rail transit industry mainly includes the following modules:

[0193] (1) Physics engine: It has models such as physical motion, raycasting, object collision, and image rendering, and can simulate vehicle motion with high precision. The engine simulates train driving based on the time and evolution rate built into the engine. It combines the environment editor module, sensor editor module, and train operation control module to respectively realize the construction of high-realistic driving scenes, the establishment of sensor models and data acquisition, and the real-time speed curve control of trains. In addition, users can directly interact with the engine through the Python language on the local side. The corresponding instructions will be compiled into C++ instructions in the engine and transmitted to the platform user end (Client), and then communicate with the physics engine in real time through the TCP protocol and plug-in script files. After receiving the instructions and executing the corresponding tasks, the engine can feedback the results to the server and user end through the reverse process. The following modules all communicate with the platform in two directions through the above method.

[0194] (2) Environment editor module: The environment editor module focuses on urban rail transit scenarios and high-speed railway train operation scenarios, including basic components such as train bodies, railway tracks, and trackside facilities. It also introduces weather factors such as wind, frost, rain, snow, fog, and dust, illumination factors such as morning, noon, and night, special working conditions such as tunnels and bridges, and edge events such as foreign object intrusion and track damage. The above factors can be freely combined by users to form driving scenarios.

[0195] (3) Sensor Editor Module: Train active perception mainly relies on sensor devices such as cameras, radars, and IMUs. This module simulates the data collection of sensor devices by building relevant sensor models and provides interfaces for modifying various device parameters (such as camera resolution, aperture, ISO, exposure factor, radar detection distance, scanning frequency, number of points, number of channels, IMU noise level, etc.).

[0196] (4) Train operation control module: This module obtains sensor data and makes corresponding decisions, outputs the corresponding train operation speed curve, and controls the train operation speed.

[0197] This application is mainly used to correct the data of various sensors (cameras, radars) in the sensor editor module.

[0198] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for correcting data of a simulated vehicle-mounted sensing device, used in a virtual simulation platform, characterized in that: The automatic scene generation method based on railway topological relationships includes: Correct the camera sensor data in the simulation environment; Correct radar sensor data in a simulation environment.

2. The method for correcting data of a simulated vehicle-mounted sensing device according to claim 1, wherein: The correction of the camera sensor data in the simulation environment includes: Perform noise correction on the data transmitted by the simulated camera; Perform motion blur simulation on the data transmitted by the simulated camera; Dynamically expose the data transmitted by the simulated camera.

3. The method for correcting data of a simulated vehicle-mounted sensing device according to claim 2, wherein: The noise correction of the data transmitted by the simulation camera includes: Acquire an original data set, wherein the original data set includes paired data of a noise-free image and a real noise image; Performing illumination intensity estimation on each noise-free image in the training set to obtain a calibrated noise-free image, wherein the calibrated noise-free image and the real noise image constitute the training set; Obtain a noise generation model; Training the noise generation model using the training set to obtain a trained noise generation model; The generator in the trained noise generation model is embedded in the virtual simulation engine, which receives the noise-free image of the simulated light intensity signal and outputs the noise image in real time.

4. The method for correcting data of a simulated vehicle-mounted sensing device according to claim 2, wherein: The motion blur simulation of the data transmitted by the simulation camera includes: Establish the position model and posture model of the vehicle at the end of exposure; Get the total displacement; Perform displacement decomposition and vibration simulation on the total displacement to correct the sensor position; An intermediate image is rendered at each corrected sensor position, containing the scene characteristics within the sensor, and an exponential decay model is used to superimpose the intermediate images to generate the final blurred image.

5. The method for correcting data of a simulated vehicle-mounted sensing device according to claim 2, wherein: The dynamic exposure of the data transmitted by the simulation camera includes: Get the initial image; Obtaining a current exposure value according to a histogram of the initial image; Get exposure compensation value according to exposure value; Apply the exposure compensation value to the original image to produce a corrected image.

6. The method for correcting data of a simulated vehicle-mounted sensing device according to claim 1, wherein: The correction of radar sensor data in the simulation environment includes: Dynamic generation of radar point cloud motion distortion through IMU data; Achieve high-fidelity radar point cloud loss simulation through mathematical modeling.

7. The method for correcting data of a simulated vehicle-mounted sensing device according to claim 6, wherein: The dynamic generation of radar point cloud motion distortion using IMU data includes: Get original point cloud data; Get the coordinate offset value at a fixed time interval; Get the radar scanning time interval; Point cloud correction data is obtained according to the original point cloud data, the coordinate offset value and the radar scanning time interval.

8. The method for correcting data of a simulated vehicle-mounted sensing device according to claim 7, wherein: The high-fidelity radar point cloud loss simulation achieved through mathematical modeling includes: Get original point cloud data; Processing the original point cloud data by randomly discarding the original point cloud data, thereby obtaining point cloud data after random discarding; The original point cloud data after random discarding is filtered by intensity filtering to obtain filtered point cloud data.

9. A simulated vehicle-mounted sensing device data correction device, characterized in that: The simulated vehicle-mounted sensing device data correction device includes: A camera sensor data correction module, wherein the camera sensor data correction module is used to correct the camera sensor data in the simulation environment; A radar sensor data correction module is used to correct radar sensor data in a simulation environment.

Images