Law and regulation simulation test method, device and equipment for intelligent driving algorithm and storage medium

Abstract

The invention discloses a law and regulation simulation test method, device and equipment for an intelligent driving algorithm and a storage medium, and relates to the technical field of intelligent driving, and the method comprises the steps: carrying out the high-precision three-dimensional reconstruction of a law and regulation test site according to collected data, and obtaining a three-dimensional reconstruction scene, and the collected data comprises a laser radar point cloud and a multi-view camera image; obtaining simulation working condition test data based on the three-dimensional reconstruction scene; training a network model according to the simulation working condition test data and the real vehicle working condition test data to obtain a vehicle dynamics network model; and based on an automatic test program, constructing a closed-loop simulation link according to the three-dimensional reconstruction scene and the vehicle dynamics network model, and carrying out regulation simulation test on an intelligent driving algorithm. According to the invention, the efficiency and accuracy of intelligent driving regulation testing can be improved.

Description

Technical Field

[0001] The present application relates to the field of intelligent driving technology, and in particular to a regulatory simulation test method, device, equipment and storage medium for intelligent driving algorithms. Background Art

[0002] Smart driving regulatory testing is an important part of measuring the smart driving capabilities of cars. Car manufacturers usually conduct regulatory testing on the whole vehicle before the new car they develop is launched on the market. Regulatory tests such as CNCAP (China New Car Assessment Program) and ENCAP (European New Car Assessment Program) include a large number of tests related to the vehicle's smart driving functions (such as AEB, LKA, etc.). However, regulatory testing is costly, and the test results also have a great impact on the market performance of new cars. It is expensive for manufacturers to build simulation test sites, and the difference between the traditional simulation method and the actual test results is large. Therefore, there is an urgent need for a high-precision simulation test technology that can be used for smart driving regulatory testing. With the help of this technology, the regulatory test site and the smart driving functions of the whole vehicle can be simulated with high precision. The test results can be used to estimate the smart driving regulatory test score of the whole vehicle, and the smart driving algorithm can be optimized and iterated based on the test results. Summary of the invention

[0003] The main purpose of this application is to provide a regulatory simulation test method, device, equipment and storage medium for an intelligent driving algorithm, aiming to solve the technical problem of how to improve the efficiency and accuracy of intelligent driving regulatory testing.

[0004] To achieve the above objectives, the present application proposes a regulatory simulation test method for an intelligent driving algorithm, the method comprising:

[0005] Perform high-precision three-dimensional reconstruction of the regulatory test site based on the collected data to obtain a three-dimensional reconstructed scene, wherein the collected data includes a laser radar point cloud and a multi-view camera image;

[0006] Based on the three-dimensional reconstruction scene, obtaining simulation working condition test data;

[0007] The network model is trained according to the simulation working condition test data and the actual vehicle working condition test data to obtain a vehicle dynamics network model;

[0008] Based on the automated testing procedure, a closed-loop simulation link is constructed according to the three-dimensional reconstruction scenario and the vehicle dynamics network model to perform regulatory simulation testing on the intelligent driving algorithm.

[0009] In one embodiment, the step of performing high-precision three-dimensional reconstruction of the regulatory test site based on the collected data to obtain a three-dimensional reconstructed scene includes:

[0010] Get lidar point cloud and multi-view camera images;

[0011] Separating the foreground and background of the laser radar point cloud and stitching multiple frames of radar point clouds to obtain a stitched point cloud;

[0012] Separating the foreground and background of the multi-view camera image to obtain a dynamic foreground three-dimensional Gaussian and a static background three-dimensional Gaussian;

[0013] Through the laser radar point cloud, the dynamic foreground three-dimensional Gaussian and the static background three-dimensional Gaussian are subjected to point cloud prior to obtain a global three-dimensional Gaussian;

[0014] A three-dimensional Gaussian scene is reconstructed according to the global three-dimensional Gaussian, the laser radar point cloud and the spliced ​​point cloud to obtain a three-dimensional reconstructed scene, and the laser radar point cloud is used for supervised training in the scene reconstruction process.

[0015] In one embodiment, the step of training the network model according to the simulation working condition test data and the actual vehicle working condition test data to obtain the vehicle dynamics network model includes:

[0016] Obtain actual vehicle operating condition test data collected from actual vehicles;

[0017] Build a network model based on the preset network;

[0018] Pre-training the network model using the simulation working condition test data to obtain a pre-trained model;

[0019] The pre-trained model is trained again using the actual vehicle operating condition test data to obtain a vehicle dynamics network model.

[0020] In one embodiment, the step of performing secondary training on the pre-trained model using the actual vehicle operating condition test data to obtain a vehicle dynamics network model comprises:

[0021] Dividing the actual vehicle operating condition test data into training data and verification data according to a preset ratio;

[0022] The pre-trained model is trained and verified by using the training data and the verification data to obtain a vehicle dynamics network model.

[0023] In one embodiment, the step of constructing a closed-loop simulation link based on the three-dimensional reconstruction scene and the vehicle dynamics network model based on an automated testing program and performing a regulatory simulation test on the intelligent driving algorithm includes:

[0024] Obtain intelligent driving algorithms and simulated vehicle sensors to be tested;

[0025] Constructing a closed-loop simulation link through the three-dimensional reconstruction scene, the vehicle dynamics network model, the intelligent driving algorithm and the sensor of the vehicle to be tested;

[0026] Based on the automated testing procedure, the intelligent driving algorithm is subjected to regulatory simulation testing through the closed-loop simulation link.

[0027] In one embodiment, the step of constructing a closed-loop simulation link through the three-dimensional reconstructed scene, the vehicle dynamics network model, the intelligent driving algorithm and the sensor of the vehicle to be tested includes:

[0028] Based on the three-dimensional reconstructed scene, obtaining a perception result through image rendering;

[0029] Controlling the intelligent driving algorithm to obtain a vehicle control instruction according to the perception result;

[0030] Controlling the vehicle dynamics network model to obtain the vehicle posture and position change according to the vehicle control command;

[0031] Based on the posture and position change of the vehicle, the viewing angle of the sensor of the vehicle to be tested is adjusted, the three-dimensional reconstructed scene is updated, and a closed-loop simulation link is obtained.

[0032] In one embodiment, the step of performing a regulatory simulation test on the intelligent driving algorithm through the closed-loop simulation link based on an automated testing program includes:

[0033] Automatically monitor and record changes in simulation data of the intelligent driving algorithm;

[0034] Based on the test case results and the simulation data changes, determining whether the intelligent driving algorithm passes the regulatory simulation test;

[0035] After the regulatory simulation test is completed, the intelligent driving algorithm is quantitatively scored and optimized.

[0036] In addition, to achieve the above purpose, the present application also proposes a regulatory simulation test device for an intelligent driving algorithm, the device comprising:

[0037] A reconstruction module, used to perform high-precision three-dimensional reconstruction of the regulatory test site based on the collected data to obtain a three-dimensional reconstructed scene, wherein the collected data includes a laser radar point cloud and a multi-view camera image;

[0038] A simulation module, used to obtain simulation working condition test data based on the three-dimensional reconstruction scene;

[0039] A training module, used to train the network model according to the simulation working condition test data and the actual vehicle working condition test data to obtain a vehicle dynamics network model;

[0040] The test module is used to build a closed-loop simulation link based on the three-dimensional reconstruction scene and the vehicle dynamics network model based on an automated test program to perform regulatory simulation testing on the intelligent driving algorithm.

[0041] In addition, to achieve the above-mentioned objectives, the present application also proposes a regulatory simulation test device for an intelligent driving algorithm, the device comprising: a memory, a processor, and a computer program stored on the memory and executable on the processor, the computer program being configured to implement the steps of the regulatory simulation test method for the intelligent driving algorithm as described above.

[0042] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by the processor, the steps of the regulatory simulation test method of the intelligent driving algorithm as described above are implemented.

[0043] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, it implements the steps of the regulatory simulation test method for the intelligent driving algorithm as described above.

[0044] One or more technical solutions proposed in this application have at least the following technical effects:

[0045] By using 3DGS technology to reconstruct the test scenario and build a vehicle dynamics model based on a recurrent neural network, a closed-loop simulation link is formed with the intelligent driving algorithm to conduct regulatory testing on the intelligent driving algorithm. This solves the technical problems of high testing difficulty, high testing cost and low accuracy that exist in regulatory testing of the intelligent driving function of the whole vehicle. Compared with the existing technology, it achieves improved efficiency and accuracy in intelligent driving regulatory testing. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0047] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0048] Figure 1 A flowchart diagram of the first embodiment of the regulatory simulation test method for the intelligent driving algorithm of this application;

[0049] Figure 2A flow chart of the second embodiment of the regulation simulation test method for the intelligent driving algorithm of this application;

[0050] Figure 3 A flowchart of the reconstruction of regulatory test scenarios for the regulatory simulation test method for the intelligent driving algorithm provided in the second embodiment of the present application;

[0051] Figure 4 A flowchart diagram of the third embodiment of the regulatory simulation test method for the intelligent driving algorithm of this application;

[0052] Figure 5 A schematic diagram of a vehicle dynamics network model training for a regulatory simulation test method for an intelligent driving algorithm provided in Example 3 of the present application;

[0053] Figure 6 A flowchart diagram of the fourth embodiment of the regulation simulation test method for the intelligent driving algorithm of this application;

[0054] Figure 7 A schematic diagram of a closed-loop simulation link of a regulatory simulation test method for an intelligent driving algorithm provided in Embodiment 4 of the present application;

[0055] Figure 8 A schematic diagram of the overall scheme of the regulatory simulation test method for the intelligent driving algorithm provided in the embodiment of the present application;

[0056] Fig. 9 A schematic diagram of the module structure of a regulatory simulation test device for an intelligent driving algorithm according to an embodiment of the present application;

[0057] Fig.10 Schematic diagram of the equipment structure of the hardware operating environment involved in the regulatory simulation test method of the intelligent driving algorithm in the embodiment of the present application.

[0058] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0059] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.

[0060] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0061] The main solution of the embodiment of the present application is: to perform high-precision three-dimensional reconstruction of the regulatory test site based on the collected data to obtain a three-dimensional reconstructed scene, and the collected data includes a lidar point cloud and a multi-view camera image; based on the three-dimensional reconstructed scene, simulated operating condition test data is obtained; based on the simulated operating condition test data and the actual vehicle operating condition test data, a network model is trained to obtain a vehicle dynamics network model; based on an automated testing program, a closed-loop simulation link is constructed according to the three-dimensional reconstructed scene and the vehicle dynamics network model to perform a regulatory simulation test on the intelligent driving algorithm.

[0062] In this embodiment, for ease of description, the following description is made taking the internal executor of the regulatory simulation test system as the execution subject.

[0063] The existing technology of vehicle intelligent driving function has technical problems such as high testing difficulty, high testing cost and low accuracy when conducting regulatory testing.

[0064] This application provides a solution that uses 3DGS technology to reconstruct test scenarios, builds a vehicle dynamics model based on a recurrent neural network, forms a closed-loop simulation link with the intelligent driving algorithm, and conducts regulatory testing on the intelligent driving algorithm, thereby improving the efficiency and accuracy of intelligent driving regulatory testing.

[0065] It can be seen from the above embodiments that the present application uses 3DGS technology to reconstruct the test scene, constructs a vehicle dynamics model based on a recurrent neural network, and forms a closed-loop simulation link with the intelligent driving algorithm to perform regulatory testing on the intelligent driving algorithm. This solves the technical problems of high testing difficulty, high testing cost, and low accuracy that exist when performing regulatory testing on the intelligent driving function of the whole vehicle, and achieves the goal of improving the efficiency and accuracy of intelligent driving regulatory testing.

[0066] It should be noted that the execution subject of this embodiment can be a computing service device with data processing, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or an electronic device capable of realizing the above functions, etc. The following takes the internal executor of the regulatory simulation test system as an example to illustrate this embodiment and the following embodiments.

[0067] Based on this, the embodiment of the present application provides a method for internal actuator of a regulatory simulation test system, referring to Figure 1 , Figure 1 This is a flow chart of the first embodiment of the internal actuator method of the regulatory simulation test system of the present application.

[0068] In this embodiment, the regulatory simulation test system internal actuator method includes steps S10 to S40:

[0069] Step S10, performing high-precision three-dimensional reconstruction of the regulatory test site based on the collected data to obtain a three-dimensional reconstructed scene, wherein the collected data includes a laser radar point cloud and a multi-view camera image.

[0070] It should be noted that the data collected is the regulatory site scene data collected by installing high-precision laser radar and 360° full-view HD camera on the data collection vehicle. The laser radar point cloud is collected by the high-precision laser radar, and the multi-view camera image is collected by the 360° full-view HD camera. The laser radar point cloud is determined by the laser radar emitting a laser beam and measuring the time from the laser emission to the return to the sensor to determine the distance between the object and the sensor. The laser radar scans and continuously emits laser pulses within a certain angle range, and records the return time to calculate the distance of each point on the surface of the surrounding object relative to it. The data composed of these points with three-dimensional coordinate information is often used to reflect the relevant situation of three-dimensional space and the position, shape and other characteristics of the object. Multi-view camera images refer to a group of images obtained by shooting the same scene from different angles simultaneously through multiple cameras or devices with multi-view shooting functions. The shooting equipment includes a multi-camera combination system composed of multiple monocular cameras, as well as light field cameras, panoramic cameras, and drones equipped with multiple cameras, etc., each with its own characteristics and applicable situations. It has obvious advantages and can provide richer information, enhance visual effects, improve measurement accuracy, and compensate for occlusion problems.

[0071] In addition, it should be noted that regulatory testing sites are specially established in accordance with relevant laws, regulations and standards to conduct specific tests to verify whether products, equipment, systems or personnel comply with regulatory requirements. 3D reconstruction scene refers to the process and results of converting objects or scenes in the real world into 3D digital models that can be processed by computers through various technical means.

[0072] In addition, it should be noted that 3D reconstruction is a process that uses a variety of technical means to obtain the 3D geometric information of an object or scene and construct its 3D model. Its data sources are very diverse. For example, point cloud data can be obtained through lidar. These point cloud data contain the 3D spatial positions of many points on the surface of the object. Based on these data, appropriate algorithms can be used to reconstruct the 3D shape of the object; images taken by the camera can also be used. Through computer vision technology, such as multi-view geometry principles, the relationship between corresponding points in different images can be found, and the positions of these points in 3D space can be calculated, and then the 3D model can be gradually constructed. 3DGS, or three-dimensional Gaussian splashing, is a cutting-edge technology for real-time radiation field rendering, which plays a significant role in the fields of 3D reconstruction and rendering. It represents the scene as a set of three-dimensional Gaussian bodies, and optimizes its properties to fit the geometric and visual characteristics of the scene to approximate the surface and volume of the object, so as to achieve efficient rendering and scene manipulation and editing. It has high rendering efficiency, can synthesize new views in real time at 1080p resolution, and its training speed is nearly 100 times faster than the neural radiation field method. It has good visual effects and can handle complex visual effects such as view-related lighting and transparency, presenting realistic effects. It has high model accuracy. Some products combine related algorithms with it to obtain centimeter-level precision three-dimensional models. It has efficient data processing, and the underlying engine design helps large-scale data processing. Some of the patents containing data compression can alleviate storage and other pressures. This embodiment and other embodiments are described by taking the 3DGS technology to reconstruct the regulatory test scene as an example.

[0073] Step S20, obtaining simulation working condition test data based on the three-dimensional reconstructed scene.

[0074] After completing the 3D reconstruction of the regulatory test scene, the sensors of the actual vehicle to be tested can be arranged in the 3D reconstruction scene according to their positions and parameters. LiDAR, millimeter-wave radar, etc. can also be simulated based on their working principles. In the 3D reconstruction scene, dynamic simulation software is used to perform simulation tests based on the acceleration, braking, steering and other working conditions of the vehicle to obtain simulation test data.

[0075] Step S30, training the network model according to the simulation working condition test data and the actual vehicle working condition test data to obtain a vehicle dynamics network model.

[0076] It should be noted that the actual vehicle operating condition test data is based on the test data of various operating conditions such as acceleration, braking, steering, etc. collected from the actual vehicle. The network model is a network model established with the help of the branch network LSTM (Long Short-Term Memory Network) in RNN (Recurrent Neural Network). After training the network model, a high-precision vehicle dynamics network model can be obtained.

[0077] In addition, it should be noted that recurrent neural network (RNN) is a type of neural network that is extremely important in the field of deep learning. It is mainly used to process sequence data, such as text, speech, time series, etc. The characteristic of RNN is that the connection between its neurons forms a directed cycle, which enables it to consider the information of previous input when processing each new input. This memory characteristic is its key advantage. RNN also has certain limitations. For example, when processing long sequence data, there may be problems of gradient vanishing or gradient explosion, which makes it difficult for the model to effectively learn long-distance dependencies. In order to solve these problems, improved versions such as long short-term memory network (LSTM) and gated recurrent unit (GRU) have emerged. They add some control gates on the basis of RNN, so that information can flow and be better preserved in the network, thereby improving the processing ability of long sequence data. Long short-term memory network (LSTM) is a special recurrent neural network, which is mainly designed to solve the problems of gradient vanishing and gradient explosion that occur in traditional RNN when processing long sequence data. LSTM controls the flow of information by introducing a gating mechanism, including forget gate, input gate and output gate. The forget gate determines which information is discarded from the cell state. It generates a value between 0 and 1 based on the current input and the hidden state at the previous moment to determine how much past information is forgotten. The input gate is used to determine which new information to update to the cell state. It combines the current input and the hidden state at the previous moment to filter information. Only allowed information can enter the cell state. The output gate controls which information is output from the cell state as the hidden state at the current moment, so that information useful for subsequent sequence processing can be selectively output. The existence of these gates enables LSTM to better capture long-term dependencies in long sequence data.

[0078] Step S40, based on the automated testing program, a closed-loop simulation link is constructed according to the three-dimensional reconstruction scene and the vehicle dynamics network model to perform a regulatory simulation test on the intelligent driving algorithm.

[0079] It should be noted that an automated testing program is a program that uses software tools and scripts to replace manual testing tasks, so as to efficiently and accurately test the functions, performance, compatibility and other aspects of various products such as software, systems, and websites. It can automatically simulate user operations based on pre-set test cases and rules, such as automatically clicking interface buttons, entering text, and selecting menu options in software testing, so as to verify whether the software runs as expected and whether the functions are fully available. In terms of performance testing, automated testing programs can simulate the situation where a large number of concurrent users access the system at the same time, collect key indicator data such as response time, throughput, and resource utilization, and then judge the performance of the system under high load. Automated testing programs can be written in a variety of programming languages. For example, Python is often used to build such programs due to its concise syntax and rich test framework library. Java is also one of the common choices. Its powerful object-oriented features are conducive to creating stable and reliable test codes. In addition, tools such as Selenium, which are specifically used for automated testing of Web applications, can easily interact with browsers to locate and operate web page elements; Appium is suitable for automated testing of mobile applications, and can operate applications on Android and Apple phones for testing.

[0080] In addition, it should be noted that the closed-loop simulation link is a complex system simulation architecture, which is mainly used to simulate the feedback control mechanism during the operation of the actual system. In this link, there will first be an input signal, which can represent physical quantities, instructions or other types of data, and it will be sent to a simulation model. This simulation model is built based on the mathematical model or physical principle of the actual system to simulate the behavior of the actual system. When the input signal enters the simulation model, the model will process it according to its own rules and algorithms to generate an output result. This output result will be compared with the expected target or reference signal, and the difference between the two will be obtained by a feedback controller. Based on this difference, the feedback controller will use a specific control strategy, such as proportional-integral-differential (PID) control strategy or a more complex adaptive control strategy, to generate an adjustment signal. This adjustment signal will be fed back to the input end of the simulation model, thereby changing the input of the model so that the output of the model can be closer to the expected target. This closed-loop structure can be continuously cycled and continuously adjust the output of the model until a satisfactory result is achieved or certain convergence conditions are met. In the field of autonomous driving, it can simulate the control of the vehicle's power, steering and other systems, providing strong support for the development and testing of autonomous driving algorithms.

[0081] In addition, it should be noted that the intelligent driving algorithm is the core part of the intelligent driving system, which covers many complex calculations and decision-making processes. At the perception level, sensor data fusion algorithms are used, such as combining camera images, lidar point cloud data, and millimeter-wave radar detection information. Camera images can provide rich visual details, such as identifying traffic signs, lane lines, and the appearance features of other vehicles; lidar point cloud data can accurately construct the three-dimensional structure of the vehicle's surroundings and clearly present the location and distance of obstacles; millimeter-wave radar performs well in detecting the speed and position of distant objects. By fusing this data, the intelligent driving algorithm can perceive the surrounding environment more comprehensively and accurately.

[0082] In addition, it should be noted that regulatory simulation testing is a method that uses computer technology and simulation environments to verify regulatory compliance. It builds test scenarios based on real regulatory terms and requirements. In this scenario, the system or product is placed in various simulated scenarios, which may involve different legal conditions, business scenarios, and compliance boundary conditions. In the field of transportation, regulatory simulation testing can simulate various traffic scenarios, such as normal traffic flow, traffic accidents, traffic violations, etc., to verify whether the autonomous driving algorithm complies with traffic regulations. For example, in a simulated intersection scenario, test whether the vehicle can correctly obey traffic lights, avoid pedestrians, and pass in the correct order, etc., to ensure that the autonomous driving system can drive safely in accordance with laws and regulations in complex traffic environments.

[0083] This embodiment provides a regulatory simulation test method for an intelligent driving algorithm. By using 3DGS technology to reconstruct the test scene, a vehicle dynamics model based on a recurrent neural network is constructed, and a closed-loop simulation link is formed with the intelligent driving algorithm to perform regulatory testing on the intelligent driving algorithm. This solves the technical problems of high testing difficulty, high testing cost, and low accuracy that exist when performing regulatory testing on the intelligent driving function of the entire vehicle, and improves the efficiency and accuracy of intelligent driving regulatory testing.

[0084] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as those in the above-mentioned embodiment 1 can be referred to the above introduction, and will not be repeated in the following. Figure 2 , step S10 includes steps S11 to S15:

[0085] Step S11, obtaining the laser radar point cloud and multi-view camera images.

[0086] A large part of the difficulty in evaluating intelligent driving functions in advance according to regulatory test standards lies in the restoration of the test site. If a simulation test site is built based on regulatory test standards, a large amount of funds will be required for construction and maintenance, and there will inevitably be errors with the official test site. The use of traditional simulation methods will result in large differences between the simulation scene and the real site. Therefore, it is necessary to use the 3DGS (three-dimensional Gaussian splash) technology in three-dimensional reconstruction technology to reconstruct and restore the regulatory test site with high precision. The regulatory test of intelligent driving requires strict reference to the test standards to test the vehicle. In order to ensure sufficient accuracy in restoring the regulatory test scene, a high-precision laser radar and a 360° full-view HD camera are installed on the data collection vehicle. The laser radar point cloud is collected by a high-precision laser radar, and the multi-view camera images are collected by a 360° full-view HD camera. After processing these data, three-dimensional Gaussian data can be obtained for three-dimensional scene reconstruction.

[0087] Before collecting the basic data for 3D reconstruction of the scene, you can first select a site, collect the laser radar point cloud and multi-view camera images of this site through the data collection vehicle and perform 3D reconstruction, compare the reconstructed scene with the actual scene, and adaptively adjust the parameters and viewing angles of the high-precision laser radar and 360° full-view HD camera installed on the data collection vehicle based on the comparison results to ensure the accuracy of the formal 3D scene reconstruction. After completing the preparations, the data collection vehicle operates according to regulatory test standards and collects scene data at the same time.

[0088] Step S12, separating the foreground and background of the laser radar point cloud and stitching multiple frames of radar point clouds to obtain a stitched point cloud.

[0089] It should be noted that foreground and background separation is a technology widely used in image processing, computer vision and other fields. Its main purpose is to distinguish the main part of the image (i.e., foreground) from the part outside the main part (i.e., background). This separation can be achieved in many ways. One is a threshold-based method, which divides image pixels into foreground and background by setting one or more grayscale value or color value thresholds. For example, in a simple black and white image, if the foreground is a black object and the background is white, a suitable grayscale threshold can be used to classify black pixels as foreground and white pixels as background. The stitched point cloud is a complete point cloud data that can more accurately depict the scene, which is integrated from multiple independent point cloud data sets.

[0090] LiDAR point cloud is a collection of points in three-dimensional space, which contain the location information of objects, etc. A common method for separating the foreground and background of LiDAR point cloud is to separate based on distance information. Due to the working principle of LiDAR, points closer to the sensor are usually foreground objects of interest, such as vehicles, pedestrians, etc., while points far away from the sensor may be background, such as distant buildings or mountains. By setting a distance threshold, close points can be preliminarily classified into the foreground category, and farther points into the background category. Another method is to use clustering algorithms. LiDAR point cloud data can be regarded as a cluster of points generated by multiple different objects. Through clustering algorithms such as DBSCAN (density-based spatial clustering application), the point cloud is divided into multiple clusters. Then, the size, shape, and location of the cluster are used to determine whether it is the foreground or the background. For example, a small, regular-shaped cluster located in front of the sensor may be a pedestrian, which can be judged as the foreground; while a huge, wide-coverage, and distant cluster may be a building background. Semantic segmentation technology can also be applied to the separation of foreground and background of LiDAR point cloud. With the help of point cloud semantic segmentation models in deep learning, these models can learn point cloud feature patterns of different objects. For example, after being trained with a large amount of labeled data, the model can identify the part of the point cloud representing the vehicle as the foreground, and the part representing the road or the sky as the background. At the same time, the model can also use the spatial relationship and contextual information of the objects to separate more accurately. This technology is particularly important in the field of autonomous driving because it can help vehicles quickly identify important targets in the surrounding environment, such as vehicles and obstacles on the driving path as the foreground, and trees on both sides of the road as the background, so as to better support autonomous driving decisions. This embodiment and other embodiments do not limit the specific method of separating lidar point clouds.

[0091] Multi-frame radar point cloud stitching is a very critical technology in the field of lidar data processing. It mainly integrates multi-frame radar point cloud data acquired at different times or different perspectives into a unified coordinate system. When stitching, the first thing to consider is the point cloud registration problem. Point cloud registration is to find the spatial transformation relationship between point clouds of different frames, which usually involves estimating rotation and translation parameters. A common method is to use feature point matching. Representative feature points are extracted from each frame of point cloud, such as edge points and corner points of the point cloud. These feature points have relatively stable geometric features between different frames. By calculating the geometric relationship between these feature points, such as distance, angle, etc., a matching relationship between feature point pairs can be established. Then, these matching relationships are used to solve the spatial transformation matrix, which can align point clouds of different frames in space. In practical applications, multi-frame radar point cloud stitching can construct a more complete environmental model around the vehicle by stitching multi-frame point clouds acquired during vehicle driving.

[0092] Step S13, separating the foreground and background of the multi-view camera image to obtain a dynamic foreground three-dimensional Gaussian and a static background three-dimensional Gaussian.

[0093] It should be noted that the dynamic foreground 3D Gaussian refers to the modeling and representation of the foreground objects in motion in the scene through a series of 3D Gaussian distributions. It uses a 3D Gaussian function to fit the information of the dynamic foreground objects in space, such as the shape, position and motion state. Each dynamic foreground object is usually represented as a combination of one or more 3D Gaussian distributions with specific parameters, which is dynamic, complex and individual. The static background 3D Gaussian is aimed at the relatively static background parts in the scene, such as buildings, roads, trees, etc., and uses 3D Gaussian distribution to model the geometric shape, position and surface features of the background in 3D space. The entire static background scene is represented by the superposition and combination of multiple 3D Gaussian distributions, which is stable, detailed and holistic. 3D Gaussian is a mathematical model. It is an extension of 2D Gaussian distribution in 3D space. From an intuitive geometric perspective, the shape of 3D Gaussian distribution is similar to a "hill". The mean vector determines the center position of this "hill". The covariance matrix affects the shape of the "hill", including its "fatness" and "tilt". If the variance of the covariance matrix is ​​equal in each dimension and there is no correlation between different dimensions, then the shape of this three-dimensional Gaussian distribution is a symmetrical, sphere-like "hill"; if the variances in different dimensions are different, then it may be a "hill" stretched or compressed in a certain direction; when there is covariance between different dimensions, the "hill" will also appear tilted.

[0094] To separate the foreground and background of multi-view camera images to obtain dynamic foreground 3D Gaussians and static background 3D Gaussians, it is first necessary to align and correct the images collected by the multi-view camera to ensure that the images from different perspectives can accurately correspond in space and geometric relationships. This can be achieved using feature point matching and related geometric transformation algorithms. For example, by extracting feature points such as corners and edges in the image, the correspondence between feature points between images from different perspectives is found, and then the transformation parameters such as rotation and translation between them are determined, so that the image is in a unified coordinate system. Then, conventional methods for separating foreground and background are used, such as pixel-based threshold segmentation and machine learning-based classification algorithms. Pixel-based threshold segmentation can set appropriate thresholds based on the grayscale value, color value and other features of the pixels in the image to preliminarily distinguish the foreground and background areas, but this method may not be accurate enough in complex scenes. The classification algorithm based on machine learning, such as using convolutional neural network (CNN), is trained by a large amount of image data with annotated foreground and background, so that the model can learn the characteristic performance of different objects in the image, and then accurately judge the foreground and background parts of the image. The foreground here includes those dynamically changing objects, such as pedestrians and moving vehicles, and the background is the relatively static environment, such as buildings and fixed road facilities. After separating the foreground and background, it is necessary to convert them into a three-dimensional space representation and construct the corresponding Gaussian model. For the dynamic foreground part, since the object is in motion and has its own position, shape and other attributes in the three-dimensional space, the disparity information in the multi-view image and the triangulation principle are used to calculate the coordinate position and other data of the foreground object in the three-dimensional space. Then, according to the distribution characteristics and motion trajectory of these objects in the three-dimensional space, the Gaussian mixture model (GMM) and other methods are used to fit the three-dimensional Gaussian representation of the dynamic foreground, so that it can accurately describe the probability distribution of the dynamic foreground in the three-dimensional space. The same is true for the static background. Based on its presentation in the multi-view image and the corresponding three-dimensional spatial information, the key information such as the background coordinates in the three-dimensional space is calculated in a similar way, and then the three-dimensional Gaussian of the static background is constructed using appropriate Gaussian modeling methods to reflect the stable characteristics and probability distribution state of the static background in the three-dimensional space.

[0095] Step S14, performing point cloud prior analysis on the dynamic foreground three-dimensional Gaussian and the static background three-dimensional Gaussian through the laser radar point cloud to obtain a global three-dimensional Gaussian.

[0096] It should be noted that point cloud prior refers to some information, features or constraints about point clouds that are known or assumed in advance when processing point cloud data. This prior knowledge can help to more effectively perform subsequent operations such as point cloud processing, analysis and understanding. Point cloud priors include geometric shape priors. If it is known that the object represented by the point cloud has a specific shape, such as buildings usually have regular geometric shapes, and cars have roughly rectangular or streamlined shapes, this shape information is a prior. In point cloud processing, this shape prior can be used to perform operations such as model fitting, segmentation or recognition. For example, by matching the point cloud with a known car shape model, it is easier to identify the car point cloud in the scene. It also includes depth priors. Depth priors are pre-set knowledge or constraints when processing data related to depth information (such as depth images, point clouds). In the context of depth information, it can be an assumption about the depth distribution law of objects in the scene. For example, in indoor scenes, we usually assume that the depth of the wall is relatively stable, and the depth changes of the floor and ceiling are also within a certain range. This common sense about depth is depth priors.

[0097] The collected lidar point cloud image has high precision and can be used to perform point cloud prior on dynamic foreground 3D Gaussian and static background 3D Gaussian, including geometric shape prior, depth prior, etc. The data after point cloud prior can be used as global 3D Gaussian for 3D scene reconstruction.

[0098] Step S15, performing a three-dimensional Gaussian scene reconstruction according to the global three-dimensional Gaussian, the lidar point cloud and the spliced ​​point cloud to obtain a three-dimensional reconstructed scene, and the lidar point cloud is used for supervised training in the scene reconstruction process.

[0099] The dynamic foreground 3D Gaussian and static background 3D Gaussian obtained through point cloud prior can be directly used as global 3D Gaussian for 3D scene reconstruction. At the same time, the lidar point cloud map has high precision and can be supervised for training during the 3D reconstruction process. The spliced ​​point cloud obtained by the lidar point cloud can be used as a supplement to the global 3D Gaussian, and then with the help of 3DGS technology, high-precision regulatory scene reconstruction results can be obtained.

[0100] High-precision LiDAR point clouds contain rich three-dimensional spatial information, reflecting the position and shape of objects in real scenes, so they can be used for supervised training of three-dimensional reconstruction. For the supervised training process, an example can be given. For example, the three-dimensional reconstruction model to be trained is input with relevant feature information. These features can be extracted from the LiDAR point cloud data, such as the coordinates, normal vectors, local geometric structures, etc. of the points. The model attempts to generate a three-dimensional reconstruction result based on these inputs. This result may be a three-dimensional model, voxel representation, or other form to describe the scene. Then, the generated reconstruction result is compared with the real scene represented by the accurate LiDAR point cloud that has been annotated. By calculating the differences between the two in terms of spatial position, shape, object category, etc., such differences can be measured by using appropriate loss functions such as mean square error and intersection-over-union ratio. Based on the difference calculated by the loss function, the optimization algorithm is used to back-propagate the error and adjust the parameters in the three-dimensional reconstruction model, so that the reconstruction result generated by the model next time can be closer to the real scene reflected by the LiDAR point cloud. The above process of input, comparison, and parameter adjustment is repeated in this way. As the number of iterations increases, the model continues to learn and optimize, and gradually becomes able to more accurately reconstruct a three-dimensional structure that matches the real scene based on the input feature information, thereby ultimately improving the three-dimensional reconstruction capability and enabling it to be applied in many fields such as high-precision map construction for autonomous driving, architectural surveying and digital archiving, and virtual reality scene creation, providing people with more accurate and realistic three-dimensional scene presentations.

[0101] like Figure 3 The figure shows the reconstruction flow chart of the regulatory test scene, which separates the foreground and the background of the lidar point cloud and splices the multi-frame radar point cloud to obtain a spliced ​​point cloud; separates the foreground and the background of the multi-view camera image to obtain a static background 3D Gaussian and a dynamic foreground 3D Gaussian, and performs point cloud prior on the lidar point cloud to obtain a global 3D Gaussian; 3DGS scene reconstruction is performed based on the global 3D Gaussian, the spliced ​​point cloud and the lidar point cloud used for supervised training to obtain a three-dimensional reconstructed scene.

[0102] The present embodiment provides a regulation simulation test method for an intelligent driving algorithm, which obtains a laser radar point cloud and a multi-view camera image; separates the foreground and background of the laser radar point cloud and splices multiple frames of radar point clouds to obtain a spliced ​​point cloud; separates the foreground and background of the multi-view camera image to obtain a dynamic foreground three-dimensional Gaussian and a static background three-dimensional Gaussian; uses the laser radar point cloud to perform point cloud prior on the dynamic foreground three-dimensional Gaussian and the static background three-dimensional Gaussian to obtain a global three-dimensional Gaussian; performs three-dimensional Gaussian scene reconstruction based on the global three-dimensional Gaussian, the laser radar point cloud and the spliced ​​point cloud to obtain a three-dimensional reconstructed scene, and the laser radar point cloud is used for supervised training in the scene reconstruction process, thereby improving the accuracy of intelligent driving regulation testing.

[0103] Based on the first embodiment of the present application, in the third embodiment of the present application, the same or similar contents as those in the above-mentioned embodiment 1 can be referred to the above introduction, and will not be repeated in the following. Figure 4 , step S30 includes steps S31 to S34:

[0104] Step S31, obtaining actual vehicle operating condition test data collected from an actual vehicle.

[0105] To obtain the real vehicle working condition test data collected by the real vehicle, we must first determine the test vehicle and select the model that meets the test requirements. For example, if the city's autonomous driving function is tested, a vehicle with a corresponding autonomous driving assistance system must be selected. At the same time, it is necessary to ensure that the vehicle is in good mechanical condition, and that various equipment and sensors can work normally. For example, the vehicle's engine, braking system, steering system, etc. have no hidden faults, and the laser radar, camera, millimeter-wave radar and other sensors on the vehicle can accurately collect data. After the vehicle is ready, it is necessary to plan the test route and working condition scenarios. The test route should cover a variety of road conditions, including congested sections of the city, open highways, rural roads, etc., so that data under different driving conditions can be fully collected. The working condition scenario should take into account different driving conditions, such as constant speed driving, accelerated overtaking, emergency braking, turning, etc. In order to ensure the accuracy and comprehensiveness of the data, it may be necessary to repeat the test on the same route and under the same working conditions many times. Then, choose the appropriate data acquisition equipment. In addition to the sensors that come with the vehicle, you can also install some professional data recording instruments, such as high-precision GPS positioning devices to record the precise location and driving trajectory of the vehicle, and data acquisition cards to integrate and store data from different sensors. At the same time, the data acquisition frequency of these devices should be set. Too high a frequency may result in too much stored data and troublesome subsequent processing. Too low a frequency may miss some key data change details. Generally, it is set reasonably according to specific test requirements and sensor characteristics. When the actual vehicle starts to drive for testing, pay close attention to the data acquisition situation to ensure that the equipment is operating normally and the data can be recorded and stored smoothly. Once a device failure or data abnormality is found, take timely measures to solve it, such as restarting the device and checking the line connection. After the test, the collected data is exported from the device and stored in a suitable storage medium, such as a large-capacity hard disk or a professional data storage server. The data should be preliminarily sorted and classified, and divided according to different test routes, working conditions, time and other factors to facilitate further analysis, processing and application, so as to obtain valuable real vehicle working condition test data.

[0106] Step S32: constructing a network model based on the preset network.

[0107] It should be noted that the preset network is a branch network LSTM (Long Short-Term Memory Network) in RNN (Recurrent Neural Network).

[0108] The vehicle dynamics model is an important part of the closed-loop simulation link of intelligent driving. The traditional simulation test link mainly uses dynamics simulation software such as Carsim, or models based on vehicle dynamics theory. However, these methods have problems such as accuracy error and high cost. Therefore, a high-precision vehicle dynamics network model can be established with the help of the branch network LSTM (Long Short-Term Memory Network) in RNN (Recurrent Neural Network). The LSTM network has the advantages of high computational efficiency, the ability to extract the time series characteristics of data, and nonlinear modeling capabilities, and can be used to establish a high-precision vehicle dynamics model.

[0109] Step S33, pre-training the network model using the simulation working condition test data to obtain a pre-trained model.

[0110] Since there are a large number of different test scenarios in regulatory testing, involving longitudinal and lateral movements of the vehicle, in order to ensure the accuracy of the vehicle dynamics model, it is necessary to collect a large amount of vehicle operation data to train the network model. The method of simply using real vehicle data is costly and time-consuming. Therefore, the network model can be pre-trained based on a large amount of simulation data from dynamics software, and then the network model can be retrained with a small amount of real vehicle data, thereby obtaining a high-precision vehicle dynamics network model at a lower cost.

[0111] When pre-training the LSTM network model, the pre-training method includes unsupervised pre-training, such as using the autoencoder architecture combined with LSTM, allowing the input sequence to be encoded by the LSTM encoder, and then restored by the decoder, minimizing the difference between the two to learn the data structure. During the training process, the data should be divided into appropriate batches, and hyperparameters, including learning rate, hidden layer size, etc., should be set. The data is calculated and output in the LSTM through forward propagation, and the gap between the predicted output and the actual output is measured by the loss function (such as cross entropy loss, mean square error, etc.), and then the LSTM parameters are updated through back propagation and optimization algorithms (such as Adam). After multiple rounds of training, the model gradually learns the characteristics of the data, completes pre-training, and prepares for subsequent fine-tuning or direct application.

[0112] Step S34, performing secondary training on the pre-trained model using the actual vehicle operating condition test data to obtain a vehicle dynamics network model.

[0113] The pre-trained LSTM network model is retrained through real vehicle operating test data. First, the real vehicle operating test data must be sorted and preprocessed. Real vehicle operating test data often contains multiple types of information, such as vehicle speed, acceleration, steering wheel angle, detection data of various sensors, and corresponding operating environment conditions. These data need to be uniformly sorted in a suitable format to remove outliers and noise data to ensure data quality. For some missing data parts, appropriate interpolation methods can be used to supplement and improve them. At the same time, according to the characteristics of the data and the model input requirements, the data is standardized or normalized to be in a suitable numerical range to facilitate subsequent calculations and learning of the model.

[0114] In a feasible implementation manner, step S34 includes steps S341 to S342:

[0115] Step S341, dividing the actual vehicle operating condition test data into training data and verification data according to a preset ratio.

[0116] The actual vehicle working condition test data is divided into training data and validation data according to the preset ratio, that is, 8:2, of which the training data accounts for 80% and the validation data accounts for 20%. The reasonable division ratio should be determined according to the total amount of data and specific task requirements. Usually, the training set accounts for a larger proportion and is used for model parameter adjustment. The validation set is used to evaluate the performance of the model during the training process and assist in determining hyperparameters such as the optimal number of training rounds.

[0117] Step S342: training and verifying the pre-trained model using the training data and the verification data to obtain a vehicle dynamics network model.

[0118] Input the training data into the pre-trained LSTM network model, perform forward propagation calculations according to the established structure and operation rules of the model, and obtain the output results of the model for these actual vehicle operating conditions, such as the predicted driving status of the vehicle at the next moment. Then compare the output results with the corresponding actual results in the actual vehicle operating condition test data, and use a suitable loss function (such as mean square error, cross entropy, etc., which are selected according to the output type) to measure the degree of difference between the two, so as to reflect the current prediction accuracy of the model. Based on the difference calculated by the loss function, the back propagation algorithm is used to propagate the error from the output layer back to each layer of the LSTM network model. According to the gradient of each layer, the model parameters are updated and adjusted in combination with the optimization algorithm (such as the commonly used Adam optimizer, etc.), so that the model can make more accurate responses and predictions when facing actual vehicle operating conditions. The above process of inputting data, forward propagation, calculating loss, back propagation and parameter updating is repeated continuously. After multiple rounds of iterative training, the performance of the model on the verification data is continuously observed. When the performance is no longer significantly improved or the preset training stop condition is reached, the training is stopped. Finally, the LSTM network model after secondary training with actual vehicle working condition test data is obtained, that is, the vehicle dynamics network model, which can better adapt to actual vehicle application scenarios and provide more reliable support for related tasks such as autonomous driving decision-making and vehicle fault prediction.

[0119] like Figure 5 The figure shows a schematic diagram of vehicle dynamics network model training. During training, vehicle status data such as steering angle, vehicle speed, and yaw rate are used. After passing through an encoder and decoder with N modules, output prediction data such as longitudinal acceleration, lateral acceleration, and yaw rate acceleration are obtained. After a round of training is completed, the vehicle parameters are updated and a new round of output prediction of vehicle status data is performed. This cycle is repeated multiple times to achieve training of the network model.

[0120] By training and verifying the network model, the accuracy of the network model can be improved.

[0121] This embodiment provides a regulatory simulation test method for an intelligent driving algorithm, which obtains actual vehicle operating condition test data collected from an actual vehicle; constructs a network model based on a preset network; pre-trains the network model through the simulation operating condition test data to obtain a pre-trained model; and re-trains the pre-trained model through the actual vehicle operating condition test data to obtain a vehicle dynamics network model, thereby improving the accuracy of the vehicle dynamics network model.

[0122] Based on the first embodiment of the present application, in the fourth embodiment of the present application, the same or similar contents as those in the first embodiment can be referred to the above description, and will not be described in detail later. Figure 6 , step S40 includes steps S41 to S43:

[0123] Step S41, obtaining an intelligent driving algorithm and simulated sensors of the vehicle to be tested.

[0124] It should be noted that the simulated sensor of the vehicle to be tested is obtained by arranging the sensor of the vehicle to be tested in the scene and simulating the sensor after completing the three-dimensional scene reconstruction.

[0125] Step S42, constructing a closed-loop simulation link through the three-dimensional reconstructed scene, the vehicle dynamics network model, the intelligent driving algorithm and the sensor of the vehicle to be tested.

[0126] After completing the reconstruction of high-precision regulatory test scenarios, the construction of high-precision vehicle dynamics network models, and the simulation of vehicle sensors to be tested, a simulation test closed-loop link can be formed with the intelligent driving algorithm to be tested.

[0127] In a feasible implementation manner, step S42 includes steps S421 to S424:

[0128] Step S421, obtaining a perception result through image rendering based on the three-dimensional reconstructed scene.

[0129] It should be noted that image rendering is a technical process that converts a virtual three-dimensional scene or model into a two-dimensional image, using computer graphics methods to simulate visual effects such as lighting, materials, textures, and geometric shapes in the real world. From a technical perspective, in the rendering process, the geometric objects in the scene are processed first. These geometric objects are composed of various basic geometric shapes (such as triangles, quadrilaterals, etc.), and the computer needs to determine the position, direction, and relationship of these geometric shapes in three-dimensional space. For example, in a building model, the walls, roofs, and other parts are composed of a large number of geometric facets, and the rendering system must know how they are stitched together. Lighting calculation is a key part of image rendering. It simulates the propagation, reflection, refraction, and scattering of light in the scene. Different lighting models are used to achieve different lighting effects. For example, the simple Phong lighting model takes into account ambient light, diffuse light, and specular light. Ambient light simulates the uniform lighting that is prevalent in the scene; diffuse light makes the surface of the object produce uniform light and dark changes according to its material properties and lighting direction; specular light is used to generate highlight effects, such as the bright reflections seen on metal surfaces. More complex physically based rendering (PBR) models simulate the interaction between light and matter more accurately, taking into account factors such as the roughness and metallicity of the material, making the rendered objects look more realistic. Texture is also an important factor in image rendering. Texture can be a pattern, color change, or other detailed information on the surface of an object, such as the wood grain on the surface of wood, the graininess of stone, etc. Through texture mapping technology, pre-made texture images are associated with the surface of the 3D model to make the surface of the model have richer details. For example, when rendering an ancient building model, mapping the brick and stone texture with traces of time to the wall can enhance the realism of the building. In the later stages of the rendering pipeline, some image post-processing operations are also performed. For example, anti-aliasing can reduce jagged defects on the edges of the image and make the lines smoother; the depth of field effect simulates the focusing characteristics of a real camera, making some parts of the image clear and other parts blurred, adding a sense of hierarchy; color correction can adjust the overall hue, saturation, and contrast of the image to make the image more in line with visual aesthetics or a specific artistic style.

[0130] In addition, it should be noted that the perception result is a two-dimensional test scene formed after image rendering of the three-dimensional reconstructed scene, which can be used as perception input into the intelligent driving algorithm.

[0131] In order to obtain the perception result through image rendering based on the 3D reconstructed scene, it is necessary to select a suitable rendering engine, such as Unity, Unreal Engine or Blender, import the reconstructed 3D scene model into it, configure the lighting conditions, and assign the corresponding material properties to the object. During rendering, the rendering engine processes the geometric model, projects it onto the 2D image plane according to the camera's perspective, calculates the lighting effect, performs texture mapping, and performs post-processing to improve the image quality and realism. After that, the rendered image is analyzed using algorithms such as target detection, semantic segmentation, and depth estimation. It can also be optimized and supplemented by combining other sensor data such as lidar or prior knowledge, and finally obtains a 2D scene to be tested, which is the perception result.

[0132] Step S422, controlling the intelligent driving algorithm to obtain vehicle control instructions according to the perception results.

[0133] The two-dimensional test scene contains rich information obtained through various sensors after perception and processing, such as road conditions (such as road width, curve conditions, presence or absence of obstacles, etc.) obtained from image rendering analysis, other traffic participants (such as the location, movement direction and speed of vehicles and pedestrians, etc.) and traffic signs (such as the location of speed limit signs and parking signs, etc.). These perception results are presented in the form of two-dimensional images or processed data, providing basic information for the intelligent driving algorithm to fully understand the vehicle's surrounding environment. After the two-dimensional test scene, that is, the perception results, are input into the intelligent driving algorithm, the algorithm will analyze and process this information based on existing rules and learned patterns. For the trajectory planning part, it will comprehensively consider factors such as the vehicle's current position, speed and target location, combined with the direction and traffic conditions of the road in the perception results, and use path search algorithms (such as the A* algorithm to search for feasible paths on the map) and trajectory optimization algorithms (which will adjust the trajectory to make it more reasonable according to constraints such as vehicle dynamics) to generate a safe and efficient driving trajectory from the current position to the target position. This trajectory will clarify the coordinates of the points that the vehicle will pass through later and key parameters such as speed and steering angle at each point. Then, based on the planned trajectory, the intelligent driving algorithm further outputs vehicle control instructions. For example, if the trajectory requires the vehicle to accelerate to pass the intersection, the algorithm will output an instruction to increase the throttle; if a turn is required, the corresponding steering angle instruction will be output to the vehicle's steering system; and when obstacles such as pedestrians or vehicles are detected in front, and the trajectory planning requires deceleration to avoid them, a braking instruction will be output to control the vehicle's brakes, so that the vehicle can accurately drive on the road according to the planned trajectory, ensuring driving safety and achieving efficient travel goals.

[0134] Step S423, controlling the vehicle dynamics network model to obtain the vehicle posture and position change according to the vehicle control instruction.

[0135] When the vehicle control command is input into the vehicle dynamics network model, the model will first parse the key information contained in the command, such as the acceleration command corresponding to the throttle opening, the steering wheel rotation angle command, and the braking intensity command. Based on these specific commands, the model will consider many physical characteristic parameters of the vehicle itself for calculation, such as the vehicle's mass, moment of inertia, friction coefficient of the tires, parameters of the suspension system, etc. These parameters determine the actual response performance of the vehicle under force. For the self-vehicle posture, that is, the change in the vehicle's own posture angle, when receiving the steering command, the model will accurately calculate the vehicle's yaw rate based on factors such as the vehicle's steering system transmission ratio, the steering angle of the front wheels and the vehicle's moment of inertia, through dynamics-related calculation formulas, such as the use of Ackerman steering geometry principles and considering the vehicle's roll, yaw and other movements during the steering process, and then derive the vehicle's heading angle change in the plane. At the same time, it can also analyze the changes in the vehicle's roll angle and other posture angles over time, so as to clearly show how the vehicle's posture changes with the control command. In terms of position change, according to the longitudinal force of the vehicle determined by the acceleration or braking command, combined with factors such as the vehicle's mass and current speed, and using physical laws such as Newton's second law, the vehicle's speed change in the longitudinal direction (that is, the front and rear direction of the vehicle) and the corresponding displacement change can be calculated. At the same time, combined with information such as yaw angular velocity and the force generated by steering and other operations in the lateral direction of the vehicle (perpendicular to the direction of travel), the lateral position movement of the vehicle is comprehensively calculated. By continuously integrating and calculating these longitudinal and lateral position changes, as time accumulates, the position change trajectory of the vehicle in space can be accurately determined, showing the complete process of the vehicle moving from one position to another after receiving the control command. In general, the vehicle dynamics network model uses the working conditions in the control command as the vehicle state data, and after encoding and decoding in the model, it outputs the predicted data, which is the self-vehicle posture and position change data.

[0136] Step S424, based on the posture and position change of the vehicle, adjust the viewing angle of the sensor of the vehicle to be tested, update the three-dimensional reconstructed scene, and obtain a closed-loop simulation link.

[0137] The sensors on the data collection vehicle adjust the sensor viewing angle according to the changes in the vehicle's posture and position, and can obtain new three-dimensional scene data, which is equivalent to updating the three-dimensional reconstructed scene, rendering the new three-dimensional scene data, obtaining new scene data to be tested, and inputting it into the intelligent driving algorithm through sensors. In this way, the construction of a closed-loop simulation link can be completed, and high-precision regulatory simulation tests can be performed on the intelligent driving algorithm to be tested.

[0138] like Figure 7The figure shows a schematic diagram of a closed-loop simulation link. The reconstructed scene is input into the intelligent driving algorithm as a perception result after passing through the simulation sensor. The intelligent driving algorithm plans the trajectory based on the input perception result and outputs the vehicle control command. The high-precision vehicle dynamics model obtains the bus data and control command, outputs the vehicle posture and position change, and adjusts the sensor perspective. The new rendered scene is input into the intelligent driving algorithm through the sensor.

[0139] By building a closed-loop simulation link, regulatory simulation testing of intelligent driving algorithms can be achieved.

[0140] Step S43: Based on the automated testing procedure, the intelligent driving algorithm is subjected to a regulatory simulation test through the closed-loop simulation link.

[0141] Regulatory testing of vehicle intelligent driving functions is a complex project. For example, there are hundreds of test cases for AEB functions in CNCAP. In order to ensure the accuracy and reliability of simulation tests, simulation tests can be implemented with the help of automated testing methods. Therefore, automated test programs can be written with the help of programming languages ​​such as Python. The test program can call the required scenarios according to the test cases, and perform simulation tests according to the expected settings through closed-loop simulation links based on the test case test process.

[0142] In a feasible implementation manner, step S43 includes steps S431 to S433:

[0143] Step S431, automatically monitor and record the changes in simulation data of the intelligent driving algorithm.

[0144] In the simulation regulation test, first select a suitable simulation platform such as CarSim, PreScan, and SUMO, and configure the test scenarios including various road conditions, weather conditions, and traffic participants according to the regulations and actual scenario requirements. Then analyze the key data of the intelligent driving algorithm to clarify the monitoring indicators such as speed, acceleration, steering angle, distance to surrounding objects, scene recognition results, decision output, trajectory planning parameters, etc., and embed the code modules that obtain these indicator data into the appropriate position in the simulation platform code framework. Use the interface function to obtain data accurately and timely. At the same time, configure the recording function to store the data in a specific format to the local disk or data storage server at a reasonable frequency. Then start the simulation regulation test, let the intelligent driving algorithm run in the set scenario, and the monitoring and recording modules work synchronously to capture and record data changes in real time, so as to realize automatic monitoring and recording of simulation data changes of the intelligent driving algorithm.

[0145] Step S432: Based on the test case results and the simulation data changes, determine whether the intelligent driving algorithm passes the regulatory simulation test.

[0146] It should be noted that the test case results, also known as the expected results of the test cases, are the ideal outputs or behavioral performances that are pre-set based on regulatory requirements, functional specifications and specific test scenarios before testing the intelligent driving algorithm.

[0147] To determine whether the intelligent driving algorithm has passed the regulatory simulation test based on the expected results of the test cases and the changes in simulation data, it is necessary to first sort out the correspondence between the key points of the regulations and the expected results of the test cases, and clarify the requirements of the regulations for the algorithm behavior in each scenario. Then analyze the changes in simulation data to check the speed, such as whether the speed in the specified speed limit area meets the standard, whether the acceleration and deceleration process meets the maximum and minimum acceleration limits; in terms of distance, whether the distance between the vehicle and other traffic participants meets the corresponding safety distance regulations in scenarios such as following, meeting, and overtaking; in terms of decision-making and control, whether the decisions such as steering, lane changes, and traffic signal responses are reasonable and meet the operating procedures required by regulations. These simulation data changes are fully compared with the regulatory requirements and the expected results of the test cases. If they are all in line, the algorithm passes the regulatory simulation test. If there are any discrepancies, it means that it has not passed, and the algorithm needs to be further improved and optimized and retested.

[0148] Step S433: after the regulatory simulation test is completed, the intelligent driving algorithm is quantitatively scored and optimized.

[0149] After the simulation test is completed, the intelligent driving algorithm can be quantitatively scored according to the regulatory testing standards. The regulatory testing standards cover many aspects, such as the vehicle's control of speed under different road conditions, such as whether it can stably maintain the legal speed limit on the highway, and whether it can accurately reduce the speed to the specified value in special areas such as school areas. This part can be given a corresponding score based on the degree of fit between the actual speed and the standard speed; there is also the accuracy of traffic sign recognition. The higher the proportion of accurate recognition and correct response, the higher the score; as well as the timeliness of response and rationality of decision-making when dealing with various emergencies (such as sudden intrusion of pedestrians, emergency braking of the vehicle in front, etc.). By comprehensively considering the performance of these different dimensions and using the established scoring criteria, a quantitative score is given to the intelligent driving algorithm to intuitively show its overall level. At the same time, for the data recorded by the test cases that failed, these data record in detail the specific situations in which the intelligent driving algorithm has problems in specific scenarios. For example, in a test scenario where a pedestrian crosses the road, the recorded data may show that the algorithm failed to detect the pedestrian in time. From the data, we can further analyze whether it is a problem in the sensor data acquisition link, or a fault in the algorithm's recognition module, or the decision-making part fails to make the correct braking or avoidance instructions; or in the lane change test case, the algorithm operation causes the vehicle to cross the line illegally. Analysis of relevant data can check whether it is caused by inaccurate judgment of the lane line or unreasonable setting of steering control parameters. By carefully analyzing these data, we can accurately locate the weak links and defects of the intelligent driving algorithm, and then take targeted improvement measures to optimize and iterate the algorithm, such as adjusting relevant parameters, updating the recognition algorithm, and improving the decision logic, so that the intelligent driving algorithm can better meet regulatory requirements in subsequent tests and actual applications, improve its performance, and ensure driving safety and efficient driving.

[0150] By testing and optimizing the scoring of intelligent driving algorithms, the efficiency of regulatory simulation testing of intelligent driving algorithms can be improved.

[0151] This embodiment provides a regulatory simulation test method for an intelligent driving algorithm, which obtains an intelligent driving algorithm and a simulated vehicle sensor to be tested; constructs a closed-loop simulation link through the three-dimensional reconstruction scene, the vehicle dynamics network model, the intelligent driving algorithm and the vehicle sensor to be tested; and based on an automated testing program, performs a regulatory simulation test on the intelligent driving algorithm through the closed-loop simulation link, thereby improving the efficiency and accuracy of the regulatory simulation test.

[0152] For example, to help understand the implementation process of the regulatory simulation test method of the intelligent driving algorithm obtained by combining the first, second, third, and fourth embodiments, please refer to Figure 8 , Figure 8A schematic diagram of the overall scheme of a regulatory simulation test method for an intelligent driving algorithm is provided, specifically:

[0153] Collect lidar point cloud data and multi-view camera images as high-precision data to construct a 3DGS reconstruction scene; render the 3DGS reconstruction scene to obtain the scene to be tested; use the scene to be tested as perception input to the intelligent driving algorithm to be tested; the dynamics model obtains the control output data of the intelligent driving algorithm to be tested, and outputs the vehicle state adjustment data; the vehicle adjusts the view according to the vehicle state adjustment data and bus data to obtain a new 3DGS reconstruction scene.

[0154] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the regulatory simulation test method of the intelligent driving algorithm of the present application. More simple transformations based on this technical concept are all within the scope of protection of the present application.

[0155] This application also provides a regulatory simulation test device for intelligent driving algorithms, please refer to Fig. 9 , the device comprises:

[0156] The reconstruction module 10 is used to perform high-precision three-dimensional reconstruction of the regulatory test site based on the collected data to obtain a three-dimensional reconstructed scene. The collected data includes a laser radar point cloud and multi-view camera images.

[0157] The simulation module 20 is used to obtain simulation working condition test data based on the three-dimensional reconstruction scene.

[0158] The training module 30 is used to train the network model according to the simulation working condition test data and the actual vehicle working condition test data to obtain a vehicle dynamics network model.

[0159] The testing module 40 is used to build a closed-loop simulation link based on the three-dimensional reconstruction scene and the vehicle dynamics network model based on an automated testing program to perform regulatory simulation testing on the intelligent driving algorithm.

[0160] In one embodiment, the reconstruction module 10 is also used to obtain a lidar point cloud and a multi-view camera image; perform foreground and background separation on the lidar point cloud and splice multiple frames of lidar point clouds to obtain a spliced ​​point cloud; perform foreground and background separation on the multi-view camera image to obtain a dynamic foreground three-dimensional Gaussian and a static background three-dimensional Gaussian; perform point cloud prior on the dynamic foreground three-dimensional Gaussian and the static background three-dimensional Gaussian through the lidar point cloud to obtain a global three-dimensional Gaussian; perform three-dimensional Gaussian scene reconstruction based on the global three-dimensional Gaussian, the lidar point cloud and the spliced ​​point cloud to obtain a three-dimensional reconstructed scene, and the lidar point cloud is used for supervised training during the scene reconstruction process.

[0161] In one embodiment, the training module 30 is also used to obtain actual vehicle operating condition test data collected from an actual vehicle; construct a network model based on a preset network; pre-train the network model through the simulation operating condition test data to obtain a pre-trained model; and perform secondary training on the pre-trained model through the actual vehicle operating condition test data to obtain a vehicle dynamics network model.

[0162] In one embodiment, the training module 30 is further used to divide the actual vehicle operating condition test data into training data and verification data according to a preset ratio; the pre-trained model is trained and verified by the training data and the verification data to obtain a vehicle dynamics network model.

[0163] In one embodiment, the test module 40 is also used to obtain the intelligent driving algorithm and the simulated vehicle sensors to be tested; construct a closed-loop simulation link through the three-dimensional reconstruction scene, the vehicle dynamics network model, the intelligent driving algorithm and the vehicle sensors to be tested; and perform regulatory simulation testing on the intelligent driving algorithm through the closed-loop simulation link based on an automated testing program.

[0164] In one embodiment, the test module 40 is also used to obtain a perception result through image rendering based on the three-dimensional reconstructed scene; control the intelligent driving algorithm to obtain a vehicle control instruction according to the perception result; control the vehicle dynamics network model to obtain the vehicle posture and position change according to the vehicle control instruction; based on the vehicle posture and position change, adjust the viewing angle of the sensor of the vehicle to be tested, update the three-dimensional reconstructed scene, and obtain a closed-loop simulation link.

[0165] In one embodiment, the test module 40 is also used to automatically monitor and record the changes in simulation data of the intelligent driving algorithm; based on the test case results and the changes in the simulation data, determine whether the intelligent driving algorithm has passed the regulatory simulation test; after the regulatory simulation test is completed, quantitatively score and optimize the intelligent driving algorithm.

[0166] The regulation simulation test device for the intelligent driving algorithm provided in the present application adopts the regulation simulation test method for the intelligent driving algorithm in the above-mentioned embodiment, which can solve the technical problem of how to improve the efficiency and accuracy of the regulation test for intelligent driving. Compared with the prior art, the beneficial effects of the regulation simulation test device for the intelligent driving algorithm provided in the present application are the same as the beneficial effects of the regulation simulation test method for the intelligent driving algorithm provided in the above-mentioned embodiment, and the other technical features in the regulation simulation test device for the intelligent driving algorithm are the same as the features disclosed in the above-mentioned embodiment method, which will not be repeated here.

[0167] The present application provides a regulatory simulation test device for an intelligent driving algorithm, and the regulatory simulation test device for an intelligent driving algorithm includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the regulatory simulation test method for the intelligent driving algorithm in the above-mentioned embodiment one.

[0168] Reference below Fig.10 , which shows a schematic diagram of the structure of a regulation simulation test device suitable for implementing the intelligent driving algorithm of the embodiment of the present application. The regulation simulation test device of the intelligent driving algorithm in the embodiment of the present application may include but is not limited to mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Fig.10 The regulatory simulation test equipment for the intelligent driving algorithm shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0169] like Fig.10As shown, the regulation simulation test equipment for intelligent driving algorithms may include a processing device 1001 (e.g., a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 to a random access memory (RAM: Random Access Memory) 1004. In RAM1004, various programs and data required for the operation of the regulation simulation test equipment for intelligent driving algorithms are also stored. The processing device 1001, ROM1002, and RAM1004 are connected to each other via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, a touch screen, a touchpad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; storage devices 1003 including, for example, a magnetic tape, a hard disk, etc.; and communication devices 1009. The communication device 1009 can allow the regulatory simulation test equipment for the intelligent driving algorithm to communicate wirelessly or wired with other devices to exchange data. Although the figure shows a regulatory simulation test equipment for an intelligent driving algorithm with various systems, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems may be implemented or provided alternatively.

[0170] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.

[0171] The regulation simulation test equipment for intelligent driving algorithms provided in the present application adopts the regulation simulation test method for intelligent driving algorithms in the above-mentioned embodiments, which can solve the technical problem of how to improve the efficiency and accuracy of intelligent driving regulation testing. Compared with the prior art, the beneficial effects of the regulation simulation test equipment for intelligent driving algorithms provided in the present application are the same as the beneficial effects of the regulation simulation test method for intelligent driving algorithms provided in the above-mentioned embodiments, and the other technical features in the regulation simulation test equipment for intelligent driving algorithms are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.

[0172] It should be understood that the various parts disclosed in this application can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0173] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0174] The present application provides a computer-readable storage medium having computer-readable program instructions (i.e., computer programs) stored thereon, and the computer-readable program instructions are used to execute the regulatory simulation test method for the intelligent driving algorithm in the above-mentioned embodiment.

[0175] The computer-readable storage medium provided in the present application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, systems or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination of the above.

[0176] The above-mentioned computer-readable storage medium may be included in the regulatory simulation test device for the intelligent driving algorithm; or it may exist independently without being assembled into the regulatory simulation test device for the intelligent driving algorithm.

[0177] The above-mentioned computer-readable storage medium carries one or more programs. When the above-mentioned one or more programs are executed by the regulatory simulation test equipment of the intelligent driving algorithm, the regulatory simulation test equipment of the intelligent driving algorithm: performs high-precision three-dimensional reconstruction of the regulatory test site according to the collected data to obtain a three-dimensional reconstructed scene, and the collected data includes a lidar point cloud and a multi-view camera image; based on the three-dimensional reconstructed scene, obtains simulation working condition test data; trains a network model according to the simulation working condition test data and the actual vehicle working condition test data to obtain a vehicle dynamics network model; based on an automated test program, constructs a closed-loop simulation link according to the three-dimensional reconstruction scene and the vehicle dynamics network model to perform a regulatory simulation test on the intelligent driving algorithm.

[0178] Computer program code for performing the operations of the present application may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0179] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0180] The modules involved in the embodiments described in this application may be implemented by software or hardware, wherein the name of the module does not constitute a limitation on the unit itself in some cases.

[0181] The readable storage medium provided in this application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned intelligent driving algorithm regulation simulation test method, which can solve the technical problem of how to improve the efficiency and accuracy of intelligent driving regulation testing. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as the beneficial effects of the regulation simulation test method for the intelligent driving algorithm provided in the above-mentioned embodiment, and will not be repeated here.

[0182] The present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the regulatory simulation test method for the intelligent driving algorithm as described above.

[0183] The computer program product provided in this application can solve the technical problem of how to improve the efficiency and accuracy of intelligent driving regulatory testing. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as the beneficial effects of the regulatory simulation testing method for the intelligent driving algorithm provided in the above embodiment, which will not be repeated here.

[0184] The above descriptions are only some embodiments of the present application, and are not intended to limit the patent scope of the present application. All equivalent structural changes made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect applications in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A regulatory simulation test method for an intelligent driving algorithm, characterized in that: The method comprises: Perform high-precision three-dimensional reconstruction of the regulatory test site based on the collected data to obtain a three-dimensional reconstructed scene, wherein the collected data includes a laser radar point cloud and a multi-view camera image; Based on the three-dimensional reconstruction scene, obtaining simulation working condition test data; The network model is trained according to the simulation working condition test data and the actual vehicle working condition test data to obtain a vehicle dynamics network model; Based on the automated testing procedure, a closed-loop simulation link is constructed according to the three-dimensional reconstruction scenario and the vehicle dynamics network model to perform regulatory simulation testing on the intelligent driving algorithm.

2. The method according to claim 1, characterized in that The step of performing high-precision three-dimensional reconstruction of the regulatory test site based on the collected data to obtain a three-dimensional reconstructed scene includes: Get lidar point cloud and multi-view camera images; Separating the foreground and background of the laser radar point cloud and stitching multiple frames of radar point clouds to obtain a stitched point cloud; Separating the foreground and background of the multi-view camera image to obtain a dynamic foreground three-dimensional Gaussian and a static background three-dimensional Gaussian; Through the laser radar point cloud, the dynamic foreground three-dimensional Gaussian and the static background three-dimensional Gaussian are subjected to point cloud prior to obtain a global three-dimensional Gaussian; A three-dimensional Gaussian scene is reconstructed according to the global three-dimensional Gaussian, the laser radar point cloud and the spliced ​​point cloud to obtain a three-dimensional reconstructed scene, and the laser radar point cloud is used for supervised training in the scene reconstruction process.

3. The method according to claim 1, characterized in that The step of training the network model according to the simulation working condition test data and the actual vehicle working condition test data to obtain the vehicle dynamics network model comprises: Obtain actual vehicle operating condition test data collected from actual vehicles; Build a network model based on the preset network; Pre-training the network model using the simulation working condition test data to obtain a pre-trained model; The pre-trained model is trained again using the actual vehicle operating condition test data to obtain a vehicle dynamics network model.

4. The method according to claim 3, characterized in that The step of performing secondary training on the pre-trained model through the actual vehicle working condition test data to obtain a vehicle dynamics network model comprises: Dividing the actual vehicle operating condition test data into training data and verification data according to a preset ratio; The pre-trained model is trained and verified by using the training data and the verification data to obtain a vehicle dynamics network model.

5. The method according to claim 1, characterized in that The step of constructing a closed-loop simulation link based on the three-dimensional reconstruction scene and the vehicle dynamics network model based on the automated testing program and performing a regulatory simulation test on the intelligent driving algorithm comprises: Obtain intelligent driving algorithms and simulated vehicle sensors to be tested; Constructing a closed-loop simulation link through the three-dimensional reconstruction scene, the vehicle dynamics network model, the intelligent driving algorithm and the sensor of the vehicle to be tested; Based on the automated testing procedure, the intelligent driving algorithm is subjected to regulatory simulation testing through the closed-loop simulation link.

6. The method according to claim 5, characterized in that The step of constructing a closed-loop simulation link through the three-dimensional reconstruction scene, the vehicle dynamics network model, the intelligent driving algorithm and the sensor of the vehicle to be tested includes: Based on the three-dimensional reconstructed scene, obtaining a perception result through image rendering; Controlling the intelligent driving algorithm to obtain a vehicle control instruction according to the perception result; Controlling the vehicle dynamics network model to obtain the vehicle posture and position change according to the vehicle control command; Based on the posture and position change of the vehicle, the viewing angle of the sensor of the vehicle to be tested is adjusted, the three-dimensional reconstructed scene is updated, and a closed-loop simulation link is obtained.

7. The method according to claim 5, characterized in that The step of performing a regulatory simulation test on the intelligent driving algorithm through the closed-loop simulation link based on the automated testing program includes: Automatically monitor and record changes in simulation data of the intelligent driving algorithm; Based on the test case results and the simulation data changes, determining whether the intelligent driving algorithm passes the regulatory simulation test; After the regulatory simulation test is completed, the intelligent driving algorithm is quantitatively scored and optimized.

8. A regulation simulation test device for intelligent driving algorithm, characterized in that: The device comprises: A reconstruction module, used to perform high-precision three-dimensional reconstruction of the regulatory test site based on the collected data to obtain a three-dimensional reconstructed scene, wherein the collected data includes a laser radar point cloud and a multi-view camera image; A simulation module, used to obtain simulation working condition test data based on the three-dimensional reconstruction scene; A training module, used to train the network model according to the simulation working condition test data and the actual vehicle working condition test data to obtain a vehicle dynamics network model; The test module is used to build a closed-loop simulation link based on the three-dimensional reconstruction scene and the vehicle dynamics network model based on an automated test program to perform regulatory simulation testing on the intelligent driving algorithm.

9. A regulatory simulation test device for intelligent driving algorithms, characterized in that: The device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the regulatory simulation test method for the intelligent driving algorithm as described in any one of claims 1 to 7.

10. A storage medium, characterized in that: The storage medium is a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the regulatory simulation test method for the intelligent driving algorithm as described in any one of claims 1 to 7 are implemented.

Images