WO2021232229A1

WO2021232229A1 - Virtual scene generation method and apparatus, computer device and storage medium

Info

Publication number: WO2021232229A1
Application number: PCT/CN2020/091014
Authority: WO
Inventors: 刘淙; 刘轩
Original assignee: 深圳元戎启行科技有限公司
Priority date: 2020-05-19
Filing date: 2020-05-19
Publication date: 2021-11-25
Also published as: CN113272838A

Abstract

A virtual scene generation method, comprising: establishing an initial virtual scene, the initial virtual scene comprising surrounding vehicles and a target vehicle; acquiring a multi-dimensional array, and inputting the multi-dimensional array to a trained trajectory generation model, so as to output driving trajectories of the surrounding vehicles, wherein the parameters of the trajectory generation model are obtained by performing training on the basis of a trained trajectory determination model, and the parameters of the trajectory determination model are obtained by performing training by inputting, to the trajectory determination model, real driving trajectory big data and virtual driving trajectory big data and expectedly outputting tags corresponding to the real driving trajectory big data and the virtual driving trajectory big data; and adding the driving trajectories of the surrounding vehicles to the initial virtual scene, so as to obtain a first target virtual scene, the first target virtual scene being used for testing a driving trajectory planning algorithm of the target vehicle.

Description

Virtual scene generation method, device, computer equipment and storage medium

Technical field

This application relates to the field of unmanned driving technology, in particular to a virtual scene generation method, device, computer equipment and storage medium.

Background technique

With the development of computer technology, unmanned driving technology has emerged. Unmanned driving technology is to automatically plan the driving route of the unmanned vehicle based on the unmanned driving algorithm, and control the unmanned vehicle based on the driving route, so that the unmanned vehicle can reach the predetermined target location. Unmanned driving technology can effectively improve the efficiency of the transportation system and the safety of people's travel, and bring convenience to people's lives.

In traditional technology, data collected by real road tests are mainly used to construct virtual scenes, and driverless algorithms are tested based on the constructed virtual scenes. However, the real scene is reproduced and played back based on the data collected during the real drive test. This kind of virtual scene generation method is single, and the data collected during the real drive test is only used to reproduce the real scene, and the data utilization rate is low.

Summary of the invention

The embodiments of the present application provide a virtual scene generation method, device, server, and non-volatile computer-readable storage medium, which can improve data utilization.

A method for generating a virtual scene, including:

Establish an initial virtual scene, which includes environmental vehicles and target vehicles;

Obtain a multi-dimensional array, input the multi-dimensional array into the trained trajectory generation model, and output the driving trajectory of the environmental vehicle. The parameters of the trajectory discrimination model take the real driving trajectory big data and the virtual driving trajectory big data as input, and the real driving trajectory big data The label corresponding to the virtual driving trajectory big data is used as the expected output, and the input trajectory discrimination model is trained;

The driving trajectory of the environmental vehicle is added to the initial virtual scene to obtain the first target virtual scene, and the first target virtual scene is used to test the driving trajectory planning algorithm of the target vehicle.

A virtual scene generating device, including:

The initial virtual scene establishment module is used to establish the initial virtual scene, the initial virtual scene includes the surrounding vehicle and the target vehicle;

The driving trajectory acquisition module is used to acquire a multi-dimensional array, input the multi-dimensional array into the trained trajectory generation model, and output the driving trajectory of the environmental vehicle. The parameters of the trajectory generation model are obtained by training based on the trained trajectory discrimination model. The trajectory discrimination model The parameters of is taking real driving trajectory big data and virtual driving trajectory big data as input, real driving trajectory big data and virtual driving trajectory big data corresponding labels as expected output, input trajectory discriminant model training;

The first target virtual scene determination module is used to add the driving trajectory of the environmental vehicle to the initial virtual scene to obtain the first target virtual scene, and the first target virtual scene is used to test the driving trajectory planning algorithm of the target vehicle.

In one aspect, a computer device is provided. The computer device includes a processor and a memory, and computer-readable instructions are stored in the memory. When the computer-readable instructions are executed by the processor, the processor Perform the following steps:

Obtain a multidimensional array, input the multidimensional array into the trained trajectory generation model, and output the driving trajectory of the environmental vehicle. The parameters of the trajectory generation model are obtained by training based on the trained trajectory discrimination model, and the parameters of the trajectory discrimination model are based on the actual driving trajectory Big data and virtual driving trajectory big data are used as input, and the labels corresponding to the real driving trajectory big data and virtual driving trajectory big data are used as the expected output, and the input trajectory discrimination model is trained;

In one aspect, one or more non-volatile storage media storing computer-readable instructions are provided. When the computer-readable instructions are executed by one or more processors, the one or more processors perform the following step:

The above-mentioned virtual scene generation method, device, computer equipment and storage medium establish an initial virtual scene, which includes environmental vehicles and target vehicles; obtain a multi-dimensional array, input the multi-dimensional array into the trained trajectory generation model, and output the environmental vehicle The parameters of the trajectory generation model are based on the trained trajectory discriminant model training. The parameters of the trajectory discriminant model are the real driving trajectory big data and virtual driving trajectory big data as input, real driving trajectory big data and virtual driving trajectory The label corresponding to the big data is used as the expected output, and the input trajectory discrimination model is trained; the driving trajectory of the environmental vehicle is added to the initial virtual scene to obtain the first target virtual scene. The first target virtual scene is used to test the driving trajectory planning algorithm of the target vehicle. Make full use of the data collected during the real road test to train the trajectory generation model. The trained trajectory generation model can generate a driving trajectory similar to but not exactly the same as the real driving trajectory. Input the generated driving trajectory into the virtual scene to establish a Target virtual scenes with similar but not identical real scenes can assist in testing the driving trajectory planning algorithm of unmanned vehicles.

Description of the drawings

In order to better describe and illustrate the embodiments and/or examples of those applications disclosed herein, one or more drawings may be referred to. The additional details or examples used to describe the drawings should not be considered as limiting the scope of any of the disclosed inventions, the currently described embodiments and/or examples, and the best mode of these applications currently understood.

Fig. 1 is a schematic diagram of an application environment of a method for generating a virtual scene in an embodiment.

Fig. 2 is a flowchart of a method for generating a virtual scene in an embodiment.

Fig. 3 is a flowchart of training a trajectory discrimination model in an embodiment.

Fig. 4 is a flowchart of a training trajectory generation model in an embodiment.

Fig. 5 is a flow chart of generating a target true driving trajectory in an embodiment.

Fig. 6 is a flowchart of a method for generating a virtual scene in another embodiment.

Fig. 7 is a flowchart of test driving trajectory planning in an embodiment.

Fig. 8 is a structural block diagram of an apparatus for generating a virtual scene in an embodiment.

Fig. 9 is a schematic diagram of the internal structure of a server in an embodiment.

Fig. 10 is a schematic diagram of the internal structure of a terminal in an embodiment.

Detailed ways

In order to make the purpose, technical solutions, and advantages of this application clearer and clearer, the following further describes the application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present application, and are not used to limit the present application.

It can be understood that the terms "first", "second", etc. used in the embodiments of the present application can be used herein to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from another element. For example, without departing from the scope of the present application, the first control may be referred to as the second control, and both the first control and the second control are controls, but they are not the same control.

Fig. 1 is an application environment diagram of a method for generating a virtual scene in an embodiment. As shown in FIG. 1, the application environment includes a terminal 110 and a server 120, where the terminal 110 may be a desktop terminal or a mobile terminal, and the mobile terminal may be at least one of a mobile phone, a tablet computer, and a notebook computer. The server 120 may be a single server or a server cluster, and the terminal 110 and the server 120 communicate through a network.

Specifically, the terminal 110 may send big data of the real driving trajectory to the server 120. Based on the big data of the real driving trajectory, the server 120 may train a trajectory generation model. Input a set of multi-dimensional arrays to the trained trajectory generation model, and the driving trajectory can be output. The server 120 may establish an initial virtual scene including the environmental vehicle and the target vehicle, and input the driving trajectory output by the trajectory generation model into the initial virtual scene as the driving trajectory of the environmental vehicle to obtain the target virtual scene. The target vehicle can drive in the target virtual scene based on the driving trajectory planning algorithm. The server 120 may send the target virtual scene to the terminal 110 for display. It is also possible to create an initial virtual scene including the environmental vehicle and the target vehicle on the terminal 110, input a set of multi-dimensional arrays to the trained trajectory generation model on the terminal 110 to obtain the driving trajectory of the environmental vehicle, and input the driving trajectory of the environmental vehicle to the initial The virtual scene is obtained, and the target virtual scene is obtained.

It can be understood that the above application scenario is only an example and does not constitute a limitation on the data processing method of this application. For example, the data processing method provided in this application may also be executed in a terminal.

Fig. 2 is a flowchart of a method for generating a virtual scene in an embodiment. As shown in Fig. 2, a method for generating a virtual scene is illustrated by taking the application on the server or terminal in Fig. 1 as an example, which specifically includes:

S202: Establish an initial virtual scene, where the initial virtual scene includes an environmental vehicle and a target vehicle.

Among them, the target vehicle refers to a virtual unmanned vehicle. Ambient vehicles refer to virtual vehicles located around the target vehicle. The initial virtual scene refers to a virtual environment including various models, including environmental vehicles, target vehicles, traffic lights, signs, and so on. In the initial virtual scene, the environmental vehicle can have no driving behavior and remain stationary, and the environmental vehicle can also have an initial driving behavior.

Specifically, drive test data can be acquired, and an initial virtual scene can be established based on the data at any time point in the drive test data. Drive test data refers to the sensor data collected in real time by the sensors of the driverless car in the real scene (real scene). The sensor of the driverless car can upload the collected data to the server or terminal. It is also possible to obtain an existing virtual scene and convert the existing virtual scene into an initial virtual scene. The existing virtual scene may be a virtual scene that has been used to test the driving trajectory planning algorithm of an unmanned vehicle. Converting the existing virtual scene into the initial virtual scene may specifically be resetting the surrounding vehicle to the state of the driving track to be set.

S204. Obtain a multi-dimensional array, input the multi-dimensional array into the trained trajectory generation model, and output the driving trajectory of the environmental vehicle. The big data of the driving trajectory and the big data of the virtual driving trajectory are used as input, the labels corresponding to the big data of the real driving trajectory and the virtual driving trajectory big data are used as the expected output, and the input trajectory discrimination model is trained.

Among them, the trajectory generation model is used to generate the driving trajectory of the environmental vehicle. The trajectory of the environmental vehicle refers to the trajectory of the environmental vehicle relative to the target vehicle. The driving trajectory is composed of several trajectory points arranged in sequence, and each trajectory point carries time information and position information. The location information includes longitude and latitude. The track points are arranged in chronological order. The time interval of each track point is the same.

The trajectory discrimination model is a machine learning model based on the pre-historic driving trajectory and the initial trajectory generation model training. When training the trajectory discriminant model, a supervised training method is used. Take the real driving trajectory big data as input, the label corresponding to the real driving trajectory big data as the expected output, carry out the trajectory discrimination model training, take the virtual driving trajectory big data as the input, and the label corresponding to the virtual driving trajectory big data as the expected output to perform the trajectory Discriminant model training. The virtual driving trajectory big data is the output data of the initial trajectory generation model. When the training completion condition is reached, a trained trajectory discrimination model is obtained. The training completion condition may be at least one of the training reaching the maximum number of iterations or the loss value of the model being less than a preset threshold. Among them, the label corresponding to the real driving trajectory indicates that the driving trajectory is true, and the label corresponding to the virtual driving trajectory indicates that the driving trajectory is false. When in use, the input data of the trajectory discrimination model is the driving trajectory, and the output data is the probability that the driving trajectory is true. The parameters of the trajectory generation model are machine learning models trained based on the trained trajectory discrimination model. The training goal is to input the generated virtual driving trajectory into the trajectory discrimination model, and the output probability is closer to the label corresponding to the real driving trajectory big data. When in use, the input data of the trajectory generation model is a multi-dimensional array, and the output data is the virtual driving trajectory.

Specifically, a multi-dimensional array with a preset length can be randomly generated, and the multi-dimensional array can be input into the trained trajectory generation model to output the driving trajectory of the environmental vehicle.

In one embodiment, each segment of the driving trajectory can be represented by a two-dimensional matrix, the rows of the matrix represent different time points, and the columns of the matrix represent the position information of the vehicle at each time point, such as longitude and latitude. The driving trajectory can also be represented by a one-dimensional sequence. The first and second numbers in the sequence represent the longitude and latitude of the first trajectory point, and the third and fourth numbers in the sequence represent the longitude and latitude of the second trajectory point. ,And so on. When designing the initial trajectory generation model, you can set the output format of the driving trajectory. The dimensions of the multidimensional array input to the trained trajectory generation model during use and the dimension of the multidimensional training array input to the trajectory generation model during training and the length of each dimension need to be consistent.

S206: Add the driving trajectory of the environmental vehicle to the initial virtual scene to obtain a first target virtual scene, and the first target virtual scene is used to test the driving trajectory planning algorithm of the target vehicle.

Specifically, after generating the driving trajectory of the environmental vehicle through the trained trajectory generation model, the generated effective driving trajectory can be added to the environmental vehicle in the initial virtual scene. Since the generated trajectory of the environmental vehicle is the trajectory of the environmental vehicle relative to the target vehicle, it is necessary to obtain the position information of the target vehicle in the initial virtual scene, and convert the driving trajectory generated by the model into Adapting to the driving trajectory of the initial virtual scene, adding the converted driving trajectory to the environmental vehicle in the initial virtual scene, and then obtaining the target virtual scene. The driving trajectory planning algorithm can dynamically plan the driving trajectory of the target vehicle according to the status information of each model in the target virtual scene, and verify and test the driving trajectory planning algorithm according to the driving trajectory of the target vehicle. The target virtual scene can assist the development and improvement of the driving trajectory planning algorithm, which can replace the method of algorithm testing in the real environment to a certain extent, and reduce the cost of algorithm development and improvement.

Since a large number of multi-dimensional arrays can be generated, and a large number of driving trajectories can be generated through the trained trajectory generation model, a large number of target virtual scenes can be obtained. Then the trained trajectory generation model can be used to generate the target virtual scene at any time, verify and test the driving trajectory planning algorithm at any time, and improve the algorithm development and efficiency.

In one embodiment, a trajectory generation model dedicated to generating a driving trajectory corresponding to a normal driving behavior can be trained, or a trajectory generation model dedicated to generating a driving trajectory corresponding to an abnormal driving behavior can be trained. Specifically, real trajectory big data can be classified into real driving trajectory big data corresponding to normal driving behavior and real driving trajectory big data corresponding to abnormal driving behavior. The real driving trajectory big data corresponding to normal driving behavior is the real driving trajectory big data without traffic accidents or traffic violations. The real driving trajectory big data corresponding to the normal driving behavior is the real driving trajectory big data of traffic accidents or violations of traffic rules. The real driving trajectory big data corresponding to the normal driving behavior is used to train the trajectory generation model used to generate the driving trajectory corresponding to the normal driving behavior. The real driving trajectory big data corresponding to the abnormal driving behavior is used to train the trajectory generation model used to generate the driving trajectory corresponding to the abnormal driving behavior. When in use, you can select the corresponding trajectory generation model according to your needs, test the driving trajectory planning algorithm in a targeted manner, and further improve the algorithm development and improve the efficiency.

The above virtual scene generation method is to establish an initial virtual scene, which includes the environmental vehicle and the target vehicle; obtain a multi-dimensional array, input the multi-dimensional array into the trained trajectory generation model, and output the driving trajectory of the environmental vehicle and the trajectory generation model The parameters are obtained based on the training of the trained trajectory discrimination model. The parameters of the trajectory discrimination model take the real driving trajectory big data and the virtual driving trajectory big data as input, and the labels corresponding to the real driving trajectory big data and the virtual driving trajectory big data are the expected output. , The input trajectory discrimination model is trained; the driving trajectory of the environmental vehicle is added to the initial virtual scene to obtain the first target virtual scene, and the first target virtual scene is used to test the driving trajectory planning algorithm of the target vehicle. Make full use of the road test data to train the trajectory generation model. The trained trajectory generation model can generate a driving trajectory that is similar but not exactly the same as the real driving trajectory. Input the generated driving trajectory into the virtual scene to create a similar but real scene. The target virtual scene is not exactly the same, which can assist in testing the driving trajectory planning algorithm of the unmanned vehicle.

In one embodiment, as shown in FIG. 3, step S202, that is, before establishing the initial virtual scene, further includes:

S302: Establish an initial trajectory generation model, the input of the trajectory generation model is a multi-dimensional array, and the output is a virtual driving trajectory.

Among them, the virtual driving trajectory refers to the driving trajectory generated on a computer device. The virtual driving trajectory is relative to the real driving trajectory. The real driving trajectory refers to the driving trajectory of the real environment vehicle relative to the real unmanned vehicle in the real environment, which can be collected by relevant equipment.

Specifically, an initial trajectory generation model can be designed, the input data of the initial trajectory generation model is a multi-dimensional array, and the output data is a virtual driving trajectory. The parameters of the initial trajectory generation model can be set randomly. The parameters of the initial trajectory generation model can also be set by technicians based on the prior knowledge of machine learning, so as to shorten the training time of the initial trajectory generation model.

S304: Obtain multiple multi-dimensional training arrays, input each multi-dimensional training array to the initial trajectory generation model, and output a virtual driving trajectory corresponding to each multi-dimensional training array, and the virtual driving trajectory corresponds to the first label.

S306: Acquire multiple sets of target real driving trajectories, and the target real driving trajectories correspond to the second label.

Among them, the multi-dimensional training array is the input data used to train the initial trajectory generation model. The target real driving trajectory is the real driving trajectory in the real scene where there is more interaction between real vehicles. For example, car A suddenly decelerates and car B has to slow down to prevent rear-end collision, and car A in the side lane merges in front of car B, causing B The car slowed down.

Specifically, multiple multi-dimensional training arrays whose dimensions used for training are the same as the length of each dimension can be generated. Each multi-dimensional training array is input to the initial trajectory generation model, and the initial trajectory generation model outputs the virtual driving trajectory corresponding to each of the multi-dimensional training arrays. Since the parameters of the initial trajectory generation model are not optimal, the virtual driving trajectory output by the initial trajectory generation model is obviously far from the real driving trajectory. Therefore, the label corresponding to the virtual driving trajectory and the label corresponding to the real driving trajectory can be set as mutually exclusive labels. For example, the label corresponding to the real driving trajectory is 1, and the label corresponding to the virtual driving trajectory is 0.

S308, each group of target real driving trajectories and each group of virtual driving trajectories are respectively input to the initial trajectory discrimination model, and the trajectory discrimination results of each group of target real driving trajectories and the trajectory discrimination results of each group of virtual driving trajectories are output.

Specifically, an initial trajectory discrimination model can be designed. The parameters of the initial trajectory discrimination model can be set randomly. The parameters of the initial trajectory discrimination model can also be set by technicians based on the prior knowledge of machine learning to shorten the training time of the initial trajectory discrimination model. Input the real driving trajectories of each group of targets into the initial trajectory discrimination model, and the initial trajectory discrimination model outputs the trajectory discrimination results of the real driving trajectories of each group of targets, and input the virtual driving trajectories of each group into the initial trajectory discrimination model and the initial trajectory discrimination model. The trajectory discrimination results of each group of virtual driving trajectories are respectively output. The trajectory discrimination result may be the probability that the driving trajectory is true.

S310: Obtain a trained trajectory discrimination model when the trajectory discrimination results of each group of target real driving trajectories and the trajectory discrimination results of each group of virtual driving trajectories reach a preset training completion condition.

Specifically, the trajectory discrimination result of the virtual driving trajectory can be compared with the first label to obtain the first comparison result, and the trajectory discrimination result of the target real driving trajectory can be compared with the second label to obtain the second comparison result. The comparison result and the second comparison result adjust the parameters of the initial trajectory discrimination model. For example, the trajectory discrimination result of the virtual driving trajectory is subtracted from the first label to obtain the first difference calculation result, and the trajectory discrimination result of the target real driving trajectory is subtracted from the second label to obtain the second difference calculation result, and then The parameters of the initial trajectory discrimination model are adjusted according to the first difference calculation result and the second difference result. For example, the label value of the first label is 0, and the trajectory discrimination result of the virtual driving trajectory is 0.4, then the first difference calculation result is -log(1-0.4)=0.510, the label value of the second label is 1, and the target real travel The trajectory discrimination result of the trajectory is 0.9, then the second difference calculation result is -log(0.9)=0.105, and the average value of the first difference calculation result and the second difference calculation result 0.306 can be used to adjust the parameters of the initial trajectory discrimination model. The parameters of the initial trajectory discrimination model can also be adjusted by the stochastic gradient descent method. Among them, the training completion condition is used to judge whether the training of the model is completed. The training completion condition is set in advance as required, and includes at least one of the iteration when the number of training reaches the maximum or the difference calculation result is less than the set value.

In the foregoing embodiment, the model training is performed through the real driving trajectory and the corresponding label, and the virtual driving trajectory and the corresponding label, and when the training completion condition is reached, the trained trajectory discrimination model is obtained.

In one embodiment, as shown in FIG. 4, step S310, that is, when the trajectory discrimination result of each group of target real driving trajectory and the trajectory discrimination result of each group of virtual driving trajectory reach the preset training completion condition, obtain the trained The trajectory discrimination model includes:

S402: Input the trajectory discrimination results of each group of target real driving trajectories and the trajectory discrimination results of each group of virtual driving trajectories into a first target loss function for calculation to obtain a first target loss value.

S404: When the first target loss value does not exceed the first preset threshold, the preset training completion condition is reached, and the trained trajectory discrimination model is obtained.

Among them, the target loss function is used to measure the trajectory of the model to judge whether the training result is good or bad, which is reflected by the loss value. Generally speaking, the smaller the loss value, the better the trajectory discrimination training result, and the more accurate the trained model.

Specifically, the first preset threshold is a preset loss value, which can be set as required. When the first target loss value exceeds the preset threshold, the first target loss value can be used to adjust the model parameters of the initial trajectory discrimination model to obtain an updated initial trajectory discrimination model. Then the target real trajectory big data and virtual trajectory big data are acquired again to train the updated initial trajectory discrimination model. Until the first target loss value obtained by the model training does not exceed the first preset threshold, the trajectory discrimination model at the last training is used as the trained trajectory discrimination model.

In an embodiment, the first objective loss function is a loss function set according to needs, for example, it may be a cross-entropy loss function, and the cross-entropy loss function may be as follows:

Among them, y _i represents the label corresponding to the i-th driving trajectory. When the i-th driving trajectory is a virtual driving trajectory, the label corresponding to the i-th driving trajectory is 0. When the i-th driving trajectory is the target real driving trajectory, the label corresponding to the i-th driving trajectory is 1. p _i represents the trajectory discrimination result corresponding to the i-th driving trajectory.

In the above embodiment, the target value of the first target loss function is used to determine whether the training completion condition is met. When the first target loss value does not exceed the first preset threshold, the preset training completion condition is reached, which can improve the training track. The accuracy of the discriminant model.

In one embodiment, it further includes: gradually adjusting the parameters of the initial trajectory generation model, inputting each multi-dimensional training array to the initial trajectory generation model after each adjustment, and inputting the output data of the initial trajectory model after each adjustment to the The trained trajectory discrimination model outputs the trajectory discrimination results corresponding to each multi-dimensional training array; the trajectory discrimination results corresponding to each multi-dimensional training array are input into the second target loss function to calculate, and the second target loss value is obtained; when the second target loss value When the second preset threshold is not exceeded, the last adjusted parameter is used as the target parameter of the trajectory generation model to obtain the trained trajectory discrimination model. Specifically, after the trajectory discrimination model is trained, the initial trajectory generation model can be trained based on the trained trajectory discrimination model. The parameters of the initial trajectory discrimination model can be adjusted by the stochastic gradient descent method. Input each multi-dimensional training array to the initial trajectory generation model after each adjustment, input the output data of the initial trajectory model after each adjustment to the trained trajectory discrimination model, and output the trajectory discrimination results corresponding to each multi-dimensional training array; The trajectory discrimination result corresponding to each multi-dimensional training array is input into the second target loss function for calculation, and the second target loss value is obtained. When the second target loss value exceeds the second preset threshold, the second target loss value may be used to adjust the model parameters of the initial trajectory generation model to obtain an updated initial trajectory generation model. Then re-acquire the multi-dimensional training array to train the updated initial trajectory generation model. Until the second target loss value obtained by the model training does not exceed the second preset threshold, the parameters adjusted during the last training are used as the target parameters of the trajectory generation model to obtain the trained trajectory discrimination model.

In one embodiment, the second objective loss function is a loss function set according to needs, for example, it may be a cross-entropy loss function, the cross-entropy loss function may be as shown below, and its purpose is to make the virtual driving corresponding to the multi-dimensional training array The probability that the trajectory is judged to be true p _j becomes larger:

Among them, M represents the total number of multi-dimensional training arrays. p _j represents the trajectory discrimination result corresponding to the j-th multi-dimensional training array.

In the above embodiment, the trajectory generation model is trained according to the trained trajectory discrimination model, so that the virtual driving trajectory generated by the trained trajectory generation model can be determined as a true driving trajectory with a probability close to 1, which is false.

In one embodiment, as shown in FIG. 5, step S306, that is, before obtaining multiple sets of target real driving trajectories, further includes:

S502: Acquire driving data, and divide the driving data to obtain multiple sets of candidate driving data.

Specifically, an unmanned vehicle will collect and record a large amount of data in real time through multiple sensors with different functions during actual road testing, including the state of the unmanned vehicle itself and the state of the surrounding environment, including the state of the unmanned vehicle itself. Including its own shape, size, location, and speed. The state of the surrounding environment includes information such as the shape, size, location, and speed of the surrounding vehicles, and the location of the surrounding traffic lights. The data collected in real time can be integrated according to time to obtain driving data. The driving data includes the state of the unmanned vehicle at each time point and the state of the surrounding environment. After the driving data is obtained, the driving data needs to be filtered and processed to filter out the fragments that affect the trajectory behavior between the unmanned vehicle and the surrounding real vehicles. For example, the sudden deceleration of the real vehicle in front causes the unmanned vehicle to also decelerate. In order to prevent rear-end collision, another example is that the real vehicle in the side lane merges from the front of the unmanned vehicle to cause the main vehicle to decelerate. Before screening, the driving data can be preprocessed, and the driving data can be divided into multiple groups of candidate driving data to improve the efficiency of screening. The driving data division may specifically be divided into groups of candidate driving data with the same length of time according to time information, or divided into groups of candidate driving data with the same driving distance according to location information. After the division, the candidate driving data can also be filtered out to further improve the screening efficiency. The candidate driving data filtering may specifically be filtering candidate driving data where there is no environmental vehicle.

In S504, multiple sets of candidate driving data are screened according to preset conditions to obtain multiple sets of target driving data, and the preset conditions are set based on behaviors affecting trajectories between real vehicles.

S506: Extract the driving trajectory of the real vehicle from each group of target driving data, determine the target real driving trajectory according to the extracted driving trajectory, and obtain multiple sets of target real driving trajectories.

Specifically, an unmanned vehicle will collect and record a large amount of data in real time during actual road testing, including the state of the unmanned vehicle itself and the state of the surrounding environment. The state of the surrounding environment includes the shape, size, location, and location of the surrounding vehicle. Speed, location of surrounding traffic lights and other information. You can set the filter conditions as needed and write the filter conditions as a mathematical function that returns the logical judgment value. Just input the candidate driving data. If the function returns true, then the candidate driving data is the target driving data. The driving trajectory of the unmanned vehicle and the driving trajectory of the environmental vehicle can be extracted from the target driving data, and the actual driving trajectory of the target can be determined according to the extracted driving trajectory. Specifically, it may be that the driving trajectory of the driving trajectory of the environmental vehicle is subtracted from the driving trajectory of the unmanned vehicle to obtain the target real driving trajectory.

In the above embodiment, the driving data is obtained, and the driving data is divided to obtain multiple sets of candidate driving data; the multiple sets of candidate driving data are filtered according to preset conditions to obtain multiple sets of target driving data, and the preset conditions are based on the actual vehicle Influencing the trajectory behavior setting; extract the driving trajectory of the real vehicle from each group of target driving data, determine the target real driving trajectory according to the extracted driving trajectory, and obtain multiple sets of target real driving trajectories. Obtain the target's true driving trajectory from the driving data, so that it is convenient to use the target's true driving trajectory as training data to train the trajectory generation model.

In an embodiment, as shown in FIG. 6, the virtual scene generation method further includes:

S602: Acquire the current state of the environmental vehicle and the target vehicle.

S604: Based on a preset algorithm, determine the target state of the environmental vehicle according to the current state of the environmental vehicle and the target vehicle.

S606: Update the target state of the environmental vehicle to the initial virtual scene to obtain a second target virtual scene.

Specifically, using real data to construct an initial virtual scene can add some intelligent behaviors to the surrounding vehicles on the basis of real data. For example, in real data, a certain real environment vehicle is driving forward at a constant speed in the left lane of an unmanned vehicle. However, in the target virtual environment, it is necessary to change and debug the driving trajectory planning algorithm of the target vehicle. For example, at this time, the target vehicle starts to merge into the left lane from the front right of the environment vehicle. If the environmental vehicle drives completely according to the behavior in the real data, it is very likely that the environmental vehicle will rear-end the target vehicle during the merging process. Therefore, the intelligent behavior to prevent rear-end collision can be added to the environmental vehicle, so that the environmental vehicle will actively decelerate according to the relative position and speed during the merging of the target vehicle, thereby avoiding the possibility of rear-end collision. Adding intelligent behaviors to environmental vehicles can specifically be based on preset algorithms, such as adaptive cruise algorithm and rear-end collision prevention algorithm. According to the current state of the environmental vehicle and the current state of the target vehicle, the target state of the environmental vehicle at the next time point is calculated in real time. The target state of the environmental vehicle obtained by real-time calculation is updated to the initial virtual scene to obtain another target virtual scene. In this way, based on the same piece of real data, technicians can flexibly adjust the planning algorithm of the main vehicle, thereby greatly improving the utilization of data.

In an embodiment, as shown in FIG. 7, the method further includes:

S702: Obtain a target driving trajectory of the target vehicle in a target virtual scene based on a driving trajectory planning algorithm, where the target virtual scene includes at least one of a first target virtual scene and a second target virtual scene;

S704: Acquire a target driving trajectory of the environmental vehicle in the target virtual scene.

S706: When the target driving trajectory of the target vehicle intersects the target driving trajectory of the environmental vehicle, generate feedback information and send it to a preset terminal.

S708: Receive an adjusted driving trajectory planning algorithm sent by a preset terminal.

S710: Re-plan the target driving trajectory of the target vehicle based on the adjusted driving trajectory planning algorithm.

Wherein, the intersection of the target driving trajectory of the target vehicle and the target driving trajectory of the environmental vehicle means that the target driving trajectory of the target vehicle and the target driving trajectory of the environmental vehicle coincide with the trajectory points corresponding to the same time point.

Specifically, the target virtual scene can be used to test the driving trajectory planning algorithm. The target driving trajectory of the target vehicle in the first target virtual scene or the second target virtual scene based on the driving trajectory planning algorithm can be acquired, and the target driving trajectory of the environmental vehicle in the first target virtual scene or the second target virtual scene can be acquired. When it is detected that the target driving trajectory of the target vehicle and the target driving trajectory of the environmental vehicle intersect in the same target virtual scene, it indicates that the target vehicle and the environmental vehicle send a collision in the first target virtual scene or the second target virtual scene. Obviously the driving trajectory There are loopholes in the planning algorithm. At this time, feedback information can be generated based on the intersecting target driving trajectory and sent to the corresponding terminal of the technician, so that the technician can debug the driving trajectory planning algorithm in time. Further, it is possible to receive the adjusted driving trajectory planning algorithm sent by the technician through the terminal, and re-plan the target driving trajectory of the target vehicle based on the adjusted driving trajectory planning algorithm in the first target virtual scene or the second target virtual scene, and test Whether the adjusted driving trajectory planning algorithm solves the previous loopholes.

It should be understood that although the various steps in the above flowchart are displayed in sequence as indicated by the arrows, these steps are not necessarily performed in sequence in the order indicated by the arrows. Unless specifically stated in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least part of the steps in the above flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The order of execution is not necessarily performed sequentially, but may be performed alternately or alternately with other steps or at least a part of sub-steps or stages of other steps.

Fig. 8 is a structural block diagram of an apparatus for generating a virtual scene in an embodiment. As shown in FIG. 8, a virtual scene generating device includes an initial virtual scene establishing module 802, a driving track acquisition module 804, and a first target virtual scene determining module 806. in:

The initial virtual scene establishment module 802 is used to establish an initial virtual scene, and the initial virtual scene includes an environmental vehicle and a target vehicle.

The driving trajectory acquisition module 804 is used to acquire a multi-dimensional array, input the multi-dimensional array into the trained trajectory generation model, and output the driving trajectory of the environmental vehicle. The parameters of the trajectory generation model are obtained by training based on the trained trajectory discrimination model, and the trajectory discrimination The parameters of the model take the real driving trajectory big data and the virtual driving trajectory big data as input, the labels corresponding to the real driving trajectory big data and the virtual driving trajectory big data as the expected output, and the input trajectory discrimination model is trained.

The first target virtual scene determination module 806 is configured to add the driving trajectory of the environmental vehicle to the initial virtual scene to obtain the first target virtual scene, and the first target virtual scene is used to test the driving trajectory planning algorithm of the target vehicle.

In an embodiment, the virtual scene generating apparatus further includes:

The training module is used to obtain multiple multi-dimensional training arrays, input each multi-dimensional training array to the initial trajectory generation model, and output the virtual driving trajectory corresponding to each multi-dimensional training array. The virtual driving trajectory corresponds to the first label; to obtain multiple sets of target real driving trajectories , The target real driving trajectory corresponds to the second label; each group of target real driving trajectory and each group of virtual driving trajectory are respectively input to the initial trajectory discrimination model, and the trajectory discrimination results of each group of target real driving trajectories and the trajectory of each group of virtual driving trajectories are output Discrimination results: When the trajectory discrimination results of each group of target real driving trajectories and the trajectory discrimination results of each group of virtual driving trajectories reach the preset training completion conditions, the trained trajectory discrimination model is obtained.

In one embodiment, the training module is also used to input the trajectory discrimination results of each group of target real driving trajectories and the trajectory discrimination results of each group of virtual driving trajectories into the first target loss function for calculation to obtain the first target loss value; When the first target loss value does not exceed the preset threshold, the preset training completion condition is reached, and the trained trajectory discrimination model is obtained.

In an embodiment, the virtual scene generating apparatus further includes:

The screening module is used to obtain drive test data, divide the drive test data to obtain multiple drive test data segments; filter multiple drive test data segments according to preset conditions to obtain multiple target drive test data segments, and preset conditions It is set based on the behavior of influencing trajectory between real vehicles; extracting the driving trajectory of the real vehicle from each target drive test data segment, determining the target real driving trajectory according to the extracted driving trajectory, and obtaining multiple sets of target real driving trajectories.

In an embodiment, the virtual scene generating apparatus further includes:

The second target virtual scene determination module is used to obtain the current state of the environmental vehicle and the target vehicle; based on a preset algorithm, determine the target state of the environmental vehicle according to the current state of the environmental vehicle and the target vehicle; update the target state of the environmental vehicle to the initial The virtual scene is obtained, and the second target virtual scene is obtained.

In an embodiment, the virtual scene generating apparatus further includes:

The algorithm test module is used to obtain the target driving trajectory of the target vehicle in the target virtual scene based on the driving trajectory planning algorithm. The target virtual scene includes at least one of the first target virtual scene and the second target virtual scene; The target driving trajectory of the virtual scene; when the target driving trajectory of the target vehicle intersects the target driving trajectory of the environmental vehicle, the feedback information is generated and sent to the preset terminal; the adjusted driving trajectory planning algorithm sent by the preset terminal is received; based on the adjusted The proposed driving trajectory planning algorithm re-plans the target driving trajectory of the target vehicle.

Regarding the specific limitation of the virtual scene generating device, please refer to the above limitation of the virtual scene generating method, which will not be repeated here. Each module in the above virtual scene generating device may be implemented in whole or in part by software, hardware, and a combination thereof. The above-mentioned modules may be embedded in the form of hardware or independent of the processor in the computer equipment, or may be stored in the memory of the computer equipment in the form of software, so that the processor can call and execute the operations corresponding to the above-mentioned modules.

In some embodiments, a computer device is provided. The computer device may be the server 120 in FIG. 1, and its internal structure diagram may be as shown in FIG. 9. The computer equipment includes a processor, a memory, and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide calculation and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. The computer program is executed by the processor to realize a virtual scene generation method.

In some embodiments, a computer device is provided. The computer device may be the terminal 110 in FIG. 1, and its internal structure diagram may be as shown in FIG. 10. The computer equipment includes a processor, a memory, a network interface, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide calculation and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. The computer program is executed by the processor to realize a virtual scene generation method. The input device of the computer equipment can be a touch layer covered on the display screen, it can also be a button, a trackball or a touchpad provided on the housing of the computer equipment, and it can also be an external keyboard, a touchpad or a mouse.

Those skilled in the art can understand that the structures shown in FIG. 9 and FIG. 10 are only block diagrams of part of the structure related to the solution of the present application, and do not constitute a limitation on the computer equipment to which the solution of the present application is applied. The computer device may include more or fewer components than shown in the figures, or combine certain components, or have a different component arrangement.

In one embodiment, a computer device is provided, including a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the virtual scene generation method described above. Here, the steps of the method for generating a virtual scene may be the steps in the method for generating a virtual scene in each of the foregoing embodiments.

In one embodiment, a computer-readable storage medium is provided, which stores a computer program, and when the computer program is executed by a processor, the processor causes the processor to execute the steps of the virtual scene generation method described above. Here, the steps of the method for generating a virtual scene may be the steps in the method for generating a virtual scene in each of the foregoing embodiments.

A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage. In the medium, when the computer program is executed, it may include the processes of the above-mentioned method embodiments. Wherein, any reference to memory, storage, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (Read-Only Memory, ROM), magnetic tape, floppy disk, flash memory, or optical storage. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not a limitation, RAM can be in various forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), etc.

The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be It is considered as the range described in this specification.

The above-mentioned embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the scope of the invention patent. It should be pointed out that for those of ordinary skill in the art, without departing from the concept of this application, several modifications and improvements can be made, and these all fall within the protection scope of this application. Therefore, the scope of protection of the patent of this application shall be subject to the appended claims.

Claims

A method for generating a virtual scene, which is characterized in that it includes:

Establishing an initial virtual scene, the initial virtual scene including an environmental vehicle and a target vehicle;

Obtain a multi-dimensional array, input the multi-dimensional array into the trained trajectory generation model, and output the driving trajectory of the environmental vehicle. The parameters of the trajectory generation model are obtained by training based on the trained trajectory discrimination model, and the trajectory discrimination The parameters of the model take the real driving trajectory big data and the virtual driving trajectory big data as input, the labels corresponding to the real driving trajectory big data and the virtual driving trajectory big data are used as the expected output, and the input trajectory discrimination model is trained;

The driving trajectory of the environmental vehicle is added to the initial virtual scene to obtain a first target virtual scene, and the first target virtual scene is used to test the driving trajectory planning algorithm of the target vehicle.
The method according to claim 1, characterized in that, before the establishment of the initial virtual scene, the method further comprises:

Establish an initial trajectory generation model, the input of the trajectory generation model is a multi-dimensional array, and the output is a virtual driving trajectory;

Acquiring a plurality of multi-dimensional training arrays, inputting each multi-dimensional training array to the initial trajectory generation model, and outputting a virtual driving trajectory corresponding to each multi-dimensional training array, the virtual driving trajectory corresponding to the first label;

Acquiring multiple sets of target real driving trajectories, where the target real driving trajectories correspond to the second label;

Each group of target real driving trajectory and each group of virtual driving trajectory are respectively input to the initial trajectory discrimination model, and the trajectory discrimination result of each group of target real driving trajectory and the trajectory discrimination result of each group of virtual driving trajectory are output;

When the trajectory discrimination results of each group of target real driving trajectories and the trajectory discrimination results of each group of virtual driving trajectories reach a preset training completion condition, the trained trajectory discrimination model is obtained.
The method according to claim 2, characterized in that, when the trajectory discrimination result of each group of target real driving trajectory and the trajectory discrimination result of each group of virtual driving trajectory reach a preset training completion condition, the obtained The trained trajectory discrimination model includes:

Inputting the trajectory discrimination result of each group of target real driving trajectory and the trajectory discrimination result of each group of virtual driving trajectory into the first target loss function for calculation to obtain the first target loss value;

When the first target loss value does not exceed the first preset threshold, the preset training completion condition is reached, and the trained trajectory discrimination model is obtained.
The method according to claim 2, wherein the method further comprises:

The parameters of the initial trajectory generation model are gradually adjusted, each multi-dimensional training array is input to the initial trajectory generation model after each adjustment, and the output data of the initial trajectory model after each adjustment is input to the trained trajectory discrimination model , Output the trajectory discrimination result corresponding to each multi-dimensional training array;

Inputting the trajectory discrimination result corresponding to each of the multi-dimensional training arrays into the second target loss function for calculation to obtain the second target loss value;

When the second target loss value does not exceed the second preset threshold, the last adjusted parameter is used as the target parameter of the trajectory generation model to obtain the trained trajectory discrimination model.
The method according to claim 2, characterized in that, before said obtaining the multiple sets of target real driving trajectories, the method further comprises:

Acquiring driving data, dividing the driving data to obtain multiple sets of candidate driving data;

Screening the multiple sets of candidate driving data according to preset conditions to obtain multiple sets of target driving data, where the preset conditions are set based on behaviors affecting trajectories between real vehicles;

The driving trajectory of the real vehicle is extracted from each group of target driving data respectively, and the target real driving trajectory is determined according to the extracted driving trajectory, and multiple sets of target real driving trajectory are obtained.
The method according to claim 1, wherein the method further comprises:

Acquiring the current state of the environmental vehicle and the target vehicle;

Based on a preset algorithm, determine the target state of the environmental vehicle according to the current state of the environmental vehicle and the target vehicle;

The target state of the environmental vehicle is updated to the initial virtual scene to obtain a second target virtual scene.
The method according to any one of claims 1-6, wherein the method further comprises:

Acquiring a target driving trajectory of the target vehicle in a target virtual scene based on a driving trajectory planning algorithm, where the target virtual scene includes at least one of a first target virtual scene and a second target virtual scene;

Acquiring the target driving trajectory of the environmental vehicle in the target virtual scene;

When the target driving trajectory of the target vehicle intersects the target driving trajectory of the environmental vehicle, generate feedback information and send it to a preset terminal;

Receiving the adjusted driving trajectory planning algorithm sent by the preset terminal;

Re-planning the target driving trajectory of the target vehicle based on the adjusted driving trajectory planning algorithm.
A virtual scene generating device, characterized in that it comprises:

An initial virtual scene establishment module, used to establish an initial virtual scene, the initial virtual scene including an environmental vehicle and a target vehicle;

The driving trajectory acquisition module is used to acquire a multi-dimensional array, input the multi-dimensional array into the trained trajectory generation model, and output the driving trajectory of the environmental vehicle. The parameters of the trajectory generation model are based on the trained trajectory discrimination model After training, the parameters of the trajectory discrimination model take the real driving trajectory big data and the virtual driving trajectory big data as input, and the labels corresponding to the real driving trajectory big data and the virtual driving trajectory big data are used as the expected output, and the trajectory discrimination model is input Trained

The first target virtual scene determination module is configured to add the driving trajectory of the environmental vehicle to the initial virtual scene to obtain a first target virtual scene, and the first target virtual scene is used to test the driving trajectory planning algorithm of the target vehicle.
The device according to claim 8, wherein the device further comprises:

The training module is used to establish an initial trajectory generation model, the input of the trajectory generation model is a multi-dimensional array, and the output is a virtual driving trajectory; multiple multi-dimensional training arrays are obtained, and each multi-dimensional training array is input to the initial trajectory generation model, and the output is The virtual driving trajectory corresponding to each multi-dimensional training array, the virtual driving trajectory corresponding to the first label; obtaining multiple sets of target real driving trajectories, the target real driving trajectory corresponding to the second label; combining each group of target real driving trajectory with each group of virtual The driving trajectories are respectively input to the initial trajectory discrimination model, and the trajectory discrimination results of the real driving trajectories of each group of targets and the trajectory discrimination results of each group of virtual driving trajectories are output; When the trajectory discrimination result of the virtual driving trajectory reaches the preset training completion condition, the trained trajectory discrimination model is obtained.
The device according to claim 9, wherein the training module is further configured to input the trajectory discrimination results of the respective groups of target real driving trajectories and the trajectory discrimination results of each group of virtual driving trajectories into the first target loss function When the first target loss value does not exceed the first preset threshold, the preset training completion condition is reached, and the trained trajectory discrimination model is obtained.
The device according to claim 9, wherein the device further comprises:

Gradually adjust the parameters of the initial trajectory generation model, input each multi-dimensional training array to the initial trajectory generation model after each adjustment, and input the output data of the initial trajectory model after each adjustment to the trained trajectory discrimination model, and output each multi-dimensional The trajectory discrimination result corresponding to the training array;

Input the trajectory discrimination result corresponding to each multi-dimensional training array into the second target loss function for calculation to obtain the second target loss value;

When the second target loss value does not exceed the second preset threshold, the last adjusted parameter is used as the target parameter of the trajectory generation model to obtain the trained trajectory discrimination model.
The device according to claim 9, wherein the device further comprises:

The screening module is used to obtain driving data, divide the driving data to obtain multiple sets of candidate driving data; filter the multiple sets of candidate driving data according to preset conditions to obtain multiple sets of target driving data, the preset conditions It is set based on the behavior of influencing the trajectory between real vehicles; extracts the driving trajectory of the real vehicle from each group of target driving data, determines the target real driving trajectory according to the extracted driving trajectory, and obtains multiple sets of target real driving trajectories.
The device according to claim 8, wherein the device further comprises:

The second target virtual scene determination module is used to obtain the current state of the environmental vehicle and the target vehicle; based on a preset algorithm, determine the target state of the environmental vehicle according to the current state of the environmental vehicle and the target vehicle; The target state of the environmental vehicle is updated to the initial virtual scene, and the second target virtual scene is obtained.
The device according to any one of claims 8-13, wherein the device further comprises:

The algorithm test module is used to obtain the target driving trajectory of the target vehicle in a target virtual scene based on a driving trajectory planning algorithm, the target virtual scene including at least one of a first target virtual scene and a second target virtual scene; The target driving trajectory of the environmental vehicle in the target virtual scene; when the target driving trajectory of the target vehicle and the target driving trajectory of the environmental vehicle intersect, feedback information is generated and sent to a preset terminal; the preset is received The adjusted driving trajectory planning algorithm sent by the terminal; and re-planning the target driving trajectory of the target vehicle based on the adjusted driving trajectory planning algorithm.
A computer device comprising a memory and a processor, the memory storing a computer program, wherein the processor implements the steps of the method according to any one of claims 1 to 7 when the computer program is executed by the processor.
A computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of the method according to any one of claims 1 to 7 when the computer program is executed by a processor.