Methods, devices and electronic equipment for testing autonomous vehicles

Abstract

This invention discloses a method, apparatus, and electronic device for testing autonomous vehicles. The method includes: acquiring target motion parameters corresponding to the autonomous vehicle under multiple collision avoidance test conditions; predicting the collision avoidance behavior of the autonomous vehicle based on the target motion parameters using a target collision avoidance behavior prediction model to obtain target collision avoidance behavior parameters corresponding to the target motion parameters, wherein the target collision avoidance behavior prediction model is obtained through machine learning using a training sample set, and the target collision avoidance behavior parameters are used to characterize the collision avoidance behavior of the autonomous vehicle under multiple collision avoidance test conditions; and generating test results for the autonomous vehicle using the target motion parameters and the target collision avoidance behavior parameters, wherein the test results are used to characterize the emergency collision avoidance performance of the autonomous vehicle. This invention solves the technical problems of low intelligence and poor accuracy in related methods for predicting vehicle collision avoidance behavior.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving, and more specifically, to a method, apparatus, and electronic device for testing autonomous vehicles. Background Technology

[0002] As research into autonomous driving deepens, the safety and reliability of intelligent driving vehicles have received widespread attention from the industry. Facing complex road traffic flows, addressing potential traffic hazards poses a significant challenge to autonomous driving. In many traffic scenarios, the sudden appearance of other road users within the vehicle's blind spots poses a substantial threat to the safety of autonomous driving. Therefore, to improve the active safety of intelligent driving vehicles, predictive research on their driving behavior during collision avoidance is essential. However, current collision avoidance behavior prediction methods are primarily based on control logic described by traditional mathematical models. These collision avoidance systems have low levels of intelligence and cannot accurately predict vehicle driving behavior, resulting in difficulties in making flexible control decisions when dealing with complex and ever-changing traffic scenarios, and hindering the guarantee of driving safety in dangerous situations.

[0003] There is currently no effective solution to the problems of low intelligence and poor accuracy in the methods for predicting vehicle collision avoidance behavior using the aforementioned technologies. Summary of the Invention

[0004] This invention provides a method, apparatus, and electronic device for testing autonomous vehicles, to at least address the technical problems of low intelligence and poor accuracy in related technologies for predicting vehicle collision avoidance behavior.

[0005] According to one aspect of the present invention, a method for testing an autonomous vehicle is provided, comprising:

[0006] The system acquires target motion parameters for autonomous vehicles under multiple collision avoidance test conditions. Based on these parameters, it uses a target collision avoidance behavior prediction model to predict the collision avoidance behavior of the autonomous vehicle, obtaining target collision avoidance behavior parameters corresponding to the target motion parameters. The target collision avoidance behavior prediction model is obtained through machine learning using a training sample set, and the target collision avoidance behavior parameters characterize the collision avoidance behavior of the autonomous vehicle under multiple collision avoidance test conditions. Using the target motion parameters and target collision avoidance behavior parameters, the system generates test results for the autonomous vehicle, which characterize the emergency collision avoidance performance of the autonomous vehicle.

[0007] Optionally, the method for testing autonomous vehicles further includes: conducting simulation tests on autonomous vehicles using simulation test drive, constructing a training sample set, wherein each training sample in the training sample set includes: simulation motion parameters and simulation collision avoidance behavior parameters corresponding to the simulation motion parameters; and training an initial collision avoidance behavior prediction model using the training sample set to obtain a target collision avoidance behavior prediction model.

[0008] Optionally, the simulation test-driven simulation test of the autonomous vehicle is used to construct the training sample set, which includes: conducting the simulation test of the autonomous vehicle using the simulation test-driven simulation test to obtain the test log; selecting target test data from the test log based on multiple simulated driving segments during the simulation test; and constructing the training sample set using the target test data.

[0009] Optionally, training the initial collision avoidance behavior prediction model using the training sample set to obtain the target collision avoidance behavior prediction model includes: predicting collision avoidance behavior using the initial collision avoidance behavior prediction model based on the simulated motion parameters in the training sample set, and obtaining the collision avoidance behavior prediction vector corresponding to the simulated motion parameters; calculating the training loss based on the simulated collision avoidance behavior parameters and the collision avoidance behavior prediction vector corresponding to the simulated motion parameters; and adjusting the network parameters of the initial collision avoidance behavior prediction model based on the training loss to obtain the target collision avoidance behavior prediction model.

[0010] Optionally, the target collision avoidance behavior prediction model includes a long short-term memory sub-model and an attention mechanism sub-module. The initial collision avoidance behavior prediction model has the same network structure as the target collision avoidance behavior prediction model. Based on the simulated motion parameters in the training sample set, the initial collision avoidance behavior prediction model is used to predict collision avoidance behavior, obtaining the collision avoidance behavior prediction vector corresponding to the simulated motion parameters. This includes: using the long short-term memory sub-model in the initial collision avoidance behavior prediction model to perform temporal correlation learning on the simulated motion parameters in the training sample set, generating a first feature vector corresponding to the simulated motion parameters, wherein the weights of multiple feature values ​​in the first feature vector are evenly distributed; using the attention mechanism sub-module in the initial collision avoidance behavior prediction model to adjust the weights of multiple feature values ​​in the first feature vector, obtaining a second feature vector; and performing fully connected regression processing on the second feature vector to obtain the collision avoidance behavior prediction vector corresponding to the simulated motion parameters.

[0011] Optionally, the Long Short-Term Memory (LSTM) sub-model in the initial collision avoidance behavior prediction model is used to perform temporal correlation learning on the simulated motion parameters in the training sample set to generate the first feature vector corresponding to the simulated motion parameters. This includes: performing temporal correlation learning on the simulated motion parameters in the training sample set at multiple consecutive time points in the multiple layer structures of the LSM sub-model to generate the first feature vector; wherein each layer structure includes at least: a forgetting gate component, an input gate component, and an output gate component; the forgetting gate component is used to: during the temporal correlation learning process, based on the simulated motion parameters output by the previous layer structure at the current time point and the output feature vector of the current layer structure at the previous time point, determine the portion of network state parameters to be retained at the current time point from the output network state parameters at the previous time point in the current layer structure; the input gate component... The first component is used for: during temporal correlation learning, based on the simulated motion parameters output by the previous layer at the current time and the output feature vector of the current layer at the previous time, determining the second part of the network parameters to be updated at the current time in the current layer, creating candidate network state parameters at the current time in the current layer, using the candidate network state parameters to determine the second part of the network state parameters to be updated at the current time, and determining the output network state parameters at the current time based on the first part of the network state parameters and the second part of the network state parameters; the output gating component is used for: during temporal correlation learning, based on the simulated motion parameters output by the previous layer at the current time and the output feature vector of the current layer at the previous time, determining the output feature vector at the current time in the current layer using the output network state parameters at the current time.

[0012] Optionally, the second feature vector is obtained by adjusting the weights of multiple feature values ​​in the first feature vector using the attention mechanism submodule in the initial collision avoidance behavior prediction model. This includes: using the fully connected network parameters and preset activation function in the attention mechanism submodule to perform dependency analysis on multiple feature values ​​in the first feature vector to generate a target weight set corresponding to the first feature vector; and using the target weight set to update the weights of multiple feature values ​​to obtain the second feature vector.

[0013] Optionally, based on the target motion parameters, the collision avoidance behavior prediction model is used to predict the collision avoidance behavior of the autonomous vehicle, obtaining the target collision avoidance behavior parameters corresponding to the target motion parameters. This includes: using the long short-term memory sub-model in the target collision avoidance behavior prediction model to perform temporal correlation learning on the target motion parameters, generating a third feature vector corresponding to the target motion parameters, wherein the weights of multiple feature values ​​in the third feature vector are evenly distributed; using the attention mechanism sub-module in the target collision avoidance behavior prediction model to adjust the weights of multiple feature values ​​in the third feature vector, obtaining a fourth feature vector; performing fully connected regression processing on the fourth feature vector to obtain the collision avoidance behavior target vector corresponding to the target motion parameters; and using the collision avoidance behavior target vector to determine the target collision avoidance behavior parameters corresponding to the target motion parameters.

[0014] According to another aspect of the present invention, an apparatus for testing autonomous vehicles is also provided, comprising:

[0015] The system comprises the following modules: an acquisition module for acquiring target motion parameters corresponding to the autonomous vehicle under multiple collision avoidance test conditions; a prediction module for predicting the collision avoidance behavior of the autonomous vehicle based on the target motion parameters using a target collision avoidance behavior prediction model, thereby obtaining target collision avoidance behavior parameters corresponding to the target motion parameters. The target collision avoidance behavior prediction model is obtained through machine learning using a training sample set, and the target collision avoidance behavior parameters are used to characterize the collision avoidance behavior of the autonomous vehicle under multiple collision avoidance test conditions; and a generation module for generating test results for the autonomous vehicle using the target motion parameters and target collision avoidance behavior parameters. The test results are used to characterize the emergency collision avoidance performance of the autonomous vehicle.

[0016] Optionally, the method for testing autonomous vehicles further includes: a simulation module for conducting simulation tests on autonomous vehicles using simulation test drive, constructing a training sample set, wherein each training sample in the training sample set includes: simulation motion parameters and simulation collision avoidance behavior parameters corresponding to the simulation motion parameters; and training an initial collision avoidance behavior prediction model using the training sample set to obtain a target collision avoidance behavior prediction model.

[0017] Optionally, the above simulation module is also used to: conduct simulation tests on autonomous vehicles using simulation test drive, and construct a training sample set, including: conducting simulation tests on autonomous vehicles using simulation test drive to obtain test logs; selecting target test data from the test logs based on multiple simulated driving segments during the simulation test; and constructing a training sample set using the target test data.

[0018] Optionally, the simulation module is further configured to: train the initial collision avoidance behavior prediction model using the training sample set to obtain the target collision avoidance behavior prediction model, including: predicting collision avoidance behavior using the initial collision avoidance behavior prediction model based on the simulated motion parameters in the training sample set, and obtaining the collision avoidance behavior prediction vector corresponding to the simulated motion parameters; calculating the training loss based on the simulated collision avoidance behavior parameters and the collision avoidance behavior prediction vector corresponding to the simulated motion parameters; and adjusting the network parameters of the initial collision avoidance behavior prediction model based on the training loss to obtain the target collision avoidance behavior prediction model.

[0019] Optionally, the simulation module is further configured to: The target collision avoidance behavior prediction model includes a long short-term memory sub-model and an attention mechanism sub-module. The initial collision avoidance behavior prediction model has the same network structure as the target collision avoidance behavior prediction model. Based on the simulated motion parameters in the training sample set, the initial collision avoidance behavior prediction model is used to predict collision avoidance behavior, obtaining the collision avoidance behavior prediction vector corresponding to the simulated motion parameters. This includes: using the long short-term memory sub-model in the initial collision avoidance behavior prediction model to perform temporal correlation learning on the simulated motion parameters in the training sample set, generating a first feature vector corresponding to the simulated motion parameters, wherein the weights of multiple feature values ​​in the first feature vector are evenly distributed; using the attention mechanism sub-module in the initial collision avoidance behavior prediction model to adjust the weights of multiple feature values ​​in the first feature vector, obtaining a second feature vector; and performing fully connected regression processing on the second feature vector to obtain the collision avoidance behavior prediction vector corresponding to the simulated motion parameters.

[0020] Optionally, the simulation module is further configured to: employ the Long Short-Term Memory (LSTM) sub-model in the initial collision avoidance behavior prediction model to perform temporal correlation learning on the simulated motion parameters in the training sample set, generating a first feature vector corresponding to the simulated motion parameters. This includes: performing temporal correlation learning on the simulated motion parameters in the training sample set at multiple consecutive time points in the multiple layer structures of the LSM sub-model to generate a first feature vector; wherein each layer structure includes at least: a forgetting gate component, an input gate component, and an output gate component; the forgetting gate component is configured to: during the temporal correlation learning process, based on the simulated motion parameters output by the previous layer structure at the current time point and the output feature vector of the current layer structure at the previous time point, determine the portion of network state parameters to be retained at the current time point from the output network state parameters at the previous time point in the current layer structure; The input gating component is used to: during temporal correlation learning, based on the simulated motion parameters output by the previous layer at the current time and the output feature vector of the current layer at the previous time, determine the second part of the network parameters to be updated at the current time in the current layer, create candidate network state parameters at the current time in the current layer, use the candidate network state parameters to determine the second part of the network state parameters to be updated at the current time, and determine the output network state parameters at the current time based on the first part of the network state parameters and the second part of the network state parameters; the output gating component is used to: during temporal correlation learning, based on the simulated motion parameters output by the previous layer at the current time and the output feature vector of the current layer at the previous time, use the output network state parameters at the current time in the current layer to determine the output feature vector at the current time.

[0021] Optionally, the simulation module described above is further configured to: adjust the weights of multiple feature values ​​in the first feature vector using the attention mechanism submodule in the initial collision avoidance behavior prediction model to obtain the second feature vector, including: using the fully connected network parameters and preset activation function in the attention mechanism submodule to perform dependency analysis on multiple feature values ​​in the first feature vector to generate a target weight set corresponding to the first feature vector; and using the target weight set to update the weights of multiple feature values ​​to obtain the second feature vector.

[0022] Optionally, the prediction module is further configured to: predict the collision avoidance behavior of the autonomous vehicle based on the target motion parameters using a target collision avoidance behavior prediction model, and obtain the target collision avoidance behavior parameters corresponding to the target motion parameters, including: using the long short-term memory sub-model in the target collision avoidance behavior prediction model to perform temporal correlation learning on the target motion parameters, generating a third feature vector corresponding to the target motion parameters, wherein the weights of multiple feature values ​​in the third feature vector are evenly distributed; using the attention mechanism sub-module in the target collision avoidance behavior prediction model to adjust the weights of multiple feature values ​​in the third feature vector to obtain a fourth feature vector; performing fully connected regression processing on the fourth feature vector to obtain the collision avoidance behavior target vector corresponding to the target motion parameters; and using the collision avoidance behavior target vector to determine the target collision avoidance behavior parameters corresponding to the target motion parameters.

[0023] According to another aspect of the present invention, an electronic device is also provided, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method of testing an autonomous vehicle as described in any of the preceding embodiments.

[0024] In this embodiment of the invention, firstly, target motion parameters corresponding to the autonomous vehicle under multiple collision avoidance test conditions are obtained. Then, based on the target motion parameters, the collision avoidance behavior of the autonomous vehicle is predicted using a target collision avoidance behavior prediction model to obtain target collision avoidance behavior parameters corresponding to the target motion parameters. The target collision avoidance behavior prediction model is obtained through machine learning using a training sample set. The target collision avoidance behavior parameters are used to characterize the collision avoidance behavior of the autonomous vehicle under multiple collision avoidance test conditions. Finally, the test results of the autonomous vehicle are generated using the target motion parameters and the target collision avoidance behavior parameters. The test results are used to characterize the emergency collision avoidance performance of the autonomous vehicle.

[0025] It is easy to understand that the method provided by the present invention obtains target collision avoidance behavior parameters by predicting the collision avoidance behavior of autonomous vehicles based on the target motion parameters corresponding to the autonomous vehicle and using a target collision behavior prediction model. Then, it uses the target motion parameters and target collision avoidance behavior parameters to determine the emergency collision avoidance performance of the autonomous vehicle. This achieves the purpose of intelligently and accurately predicting the target collision avoidance behavior of autonomous vehicles to determine their emergency collision avoidance performance. Thus, it realizes the technical effect of improving the intelligence level of vehicle collision avoidance behavior prediction methods and improving the accuracy of target collision avoidance behavior. In turn, it solves the technical problems of low intelligence level and poor accuracy of related technologies for predicting vehicle collision avoidance behavior. Attached Figure Description

[0026] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0027] Figure 1 This is a structural block diagram of a vehicle terminal for testing an autonomous vehicle, according to an embodiment of the present invention.

[0028] Figure 2 This is a flowchart of a method for testing an autonomous vehicle according to an embodiment of the present invention;

[0029] Figure 3 This is a schematic diagram of an optional collision avoidance behavior prediction model for an autonomous vehicle according to an embodiment of the present invention;

[0030] Figure 4 This is a schematic diagram of another optional collision avoidance behavior prediction model for testing autonomous vehicles according to an embodiment of the present invention;

[0031] Figure 5 This is a schematic diagram of another optional collision avoidance behavior prediction model for an autonomous vehicle according to an embodiment of the present invention;

[0032] Figure 6 This is a flowchart of another optional process for testing an autonomous vehicle according to an embodiment of the present invention;

[0033] Figure 7 This is a structural block diagram of an optional apparatus for testing autonomous vehicles according to an embodiment of the present invention;

[0034] Figure 8 This is a structural block diagram of another optional apparatus for testing autonomous vehicles according to an embodiment of the present invention. Detailed Implementation

[0035] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0036] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0037] According to an embodiment of the present invention, a method embodiment for testing autonomous vehicles is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0038] Figure 1 This is a structural block diagram of a vehicle terminal for an optional method of testing an autonomous vehicle according to an embodiment of the present invention, such as... Figure 1 As shown, the vehicle terminal 10 (or a mobile device 10 that communicates with the vehicle) may include one or more processors 102 (processors 102 may include, but are not limited to, processing devices such as microprocessors (MCUs) or field-programmable gate arrays (FPGAs),) a memory 104 for storing data, and a transmission device 106 for communication functions. In addition, it may also include: a display device 110, an input / output device 108 (i.e., I / O devices), a Universal Serial Bus (USB) port (which may be included as one of the ports of a computer bus, not shown in the figure), a network interface (not shown in the figure), a power supply (not shown in the figure), and / or a camera (not shown in the figure). Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the vehicle terminal 1 described above. For example, the vehicle terminal 10 may also include components that are more... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0039] It should be noted that the aforementioned one or more processors 102 and / or other data processing circuits may be embodied, in whole or in part, as software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuitry may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the vehicle terminal 10 (or mobile device).

[0040] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the method for testing autonomous vehicles in this embodiment of the invention. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby implementing the aforementioned method for testing autonomous vehicles. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the vehicle terminal 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0041] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the vehicle terminal 10. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0042] Under the above operating environment, the embodiments of the present invention provide as follows: Figure 2 The method shown is for testing autonomous vehicles. Figure 2 This is a flowchart of a method for testing an autonomous vehicle according to an embodiment of the present invention, such as... Figure 2 As shown above, Figure 2 The embodiments shown may include at least the following implementation steps, namely, the technical solutions implemented by steps S21 to S23.

[0043] Step S21: Obtain the target motion parameters corresponding to the autonomous vehicle under multiple collision avoidance test conditions;

[0044] In the optional solutions provided in step S21 above, the multiple collision avoidance test conditions may include, but are not limited to: collision avoidance targets being one or more other vehicles traveling normally around the autonomous vehicle; collision avoidance targets being one or more pedestrians moving normally around the autonomous vehicle; and collision avoidance targets being one or more other vehicles traveling abnormally (such as exceeding the current road speed limit) around the autonomous vehicle. The target motion parameters may include, but are not limited to: the speed of the main vehicle (i.e., the aforementioned autonomous vehicle), the speed of pedestrians around the main vehicle, the deceleration of the main vehicle, the deceleration of the pedestrians, and the relative speed between the main vehicle and the pedestrians. It should also be noted that the target motion parameters may further include: the speeds of other vehicles around the main vehicle, the deceleration of other vehicles, and the relative speed between the main vehicle and other vehicles.

[0045] Step S22: Based on the target motion parameters, the collision avoidance behavior prediction model is used to predict the collision avoidance behavior of the autonomous vehicle, and the target collision avoidance behavior parameters corresponding to the target motion parameters are obtained. The target collision avoidance behavior prediction model is obtained by machine learning using a training sample set, and the target collision avoidance behavior parameters are used to characterize the collision avoidance behavior of the autonomous vehicle under multiple collision avoidance test conditions.

[0046] In the optional solutions provided in step S22 above, the target collision avoidance behavior prediction model can be, but is not limited to, a Long Short-Term Memory (LSTM) network model. It is understood that the LSTM network model is a specific form of Recurrent Neural Network (RNN) network model. The LSTM network model addresses the short-term memory problem of RNN network models by adding a threshold. The target collision avoidance behavior parameters can include, but are not limited to: the activation state of the vehicle's automatic emergency braking function, the collision time at the moment the automatic emergency braking function is activated, and the longitudinal distance between the vehicle and other traffic participants (surrounding vehicles, pedestrians, etc.).

[0047] The technical solution provided by this invention predicts the collision avoidance behavior of an autonomous vehicle based on target motion parameters using a target collision avoidance behavior prediction model. Specifically, the target collision avoidance behavior prediction model is an LSTM two-layer network model. The target motion parameters and the performance of the autonomous vehicle are input into the LSTM model. Through chain-like loops within the model's neurons and recursive computation, the potential temporal correlation of the target motion parameters at different times is learned, and a feature vector is output. The output feature vector is then filtered to obtain the target collision avoidance behavior parameters that significantly affect the collision avoidance behavior of the autonomous vehicle. This improves the intelligence level of autonomous vehicle collision avoidance behavior prediction and the accuracy of the predicted target collision avoidance behavior parameters. Furthermore, based on the target collision avoidance behavior parameters, the autonomous vehicle can be controlled to take collision avoidance measures to improve its active safety.

[0048] Step S23: Using the target motion parameters and target collision avoidance behavior parameters, generate test results for the autonomous vehicle, wherein the test results are used to characterize the emergency collision avoidance performance of the autonomous vehicle.

[0049] In the optional solutions provided in step S23 above, the test results may include, but are not limited to, the adjustment values ​​of the parameters of the automatic emergency braking function of the autonomous vehicle. Based on the adjustment values, the defects of the current automatic emergency braking function of the autonomous vehicle (such as excessively long response time, excessively long active distance, and excessively fast collision speed) can be determined.

[0050] In this embodiment of the invention, firstly, target motion parameters corresponding to the autonomous vehicle under multiple collision avoidance test conditions are obtained. Then, based on the target motion parameters, the collision avoidance behavior of the autonomous vehicle is predicted using a target collision avoidance behavior prediction model to obtain target collision avoidance behavior parameters corresponding to the target motion parameters. The target collision avoidance behavior prediction model is obtained through machine learning using a training sample set. The target collision avoidance behavior parameters are used to characterize the collision avoidance behavior of the autonomous vehicle under multiple collision avoidance test conditions. Finally, the test results of the autonomous vehicle are generated using the target motion parameters and the target collision avoidance behavior parameters. The test results are used to characterize the emergency collision avoidance performance of the autonomous vehicle.

[0051] It is easy to understand that the method provided by the present invention obtains target collision avoidance behavior parameters by predicting the collision avoidance behavior of autonomous vehicles based on the target motion parameters corresponding to the autonomous vehicle and using a target collision behavior prediction model. Then, it uses the target motion parameters and target collision avoidance behavior parameters to determine the emergency collision avoidance performance of the autonomous vehicle. This achieves the purpose of intelligently and accurately predicting the target collision avoidance behavior of autonomous vehicles to determine their emergency collision avoidance performance. Thus, it realizes the technical effect of improving the intelligence level of vehicle collision avoidance behavior prediction methods and improving the accuracy of target collision avoidance behavior. In turn, it solves the technical problems of low intelligence level and poor accuracy of related technologies for predicting vehicle collision avoidance behavior.

[0052] The methods described in the embodiments of the present invention will be further described below.

[0053] In an optional embodiment, in step S21, obtaining the target motion parameters corresponding to the autonomous vehicle under multiple collision avoidance test conditions includes:

[0054] Step S211: Determine the target motion parameters based on the configuration files of multiple collision avoidance test conditions, wherein the target motion parameters include the driving parameters of the autonomous vehicle and the action parameters of the collision object.

[0055] In the optional solution provided by this invention, the collision avoidance test condition is assumed to be a CPNCO (Car-to-Pedestraian Nearside Child Obstructed) scenario. This scenario is described as follows: the main vehicle is traveling forward and collides with a child pedestrian crossing its path from behind an obstacle near the main vehicle. Without applying braking, the front structure of the main vehicle impacts the child pedestrian at 50% of the vehicle's width. The data in the aforementioned configuration file may include, but is not limited to: the driving speed of the main vehicle, the speed of the child pedestrian, the trajectory of the child pedestrian, and the distance between the main vehicle and the child pedestrian. It is understood that the driving parameters of the aforementioned autonomous vehicle may include the driving speed of the main vehicle and the distance between the main vehicle and the child pedestrian, and the driving parameters of the vehicle collision object may include the speed of the child pedestrian and the trajectory of the child pedestrian.

[0056] In an optional embodiment, the method for testing autonomous vehicles further includes:

[0057] Step S24: Conduct simulation tests on autonomous vehicles using simulation test drive to construct a training sample set. Each training sample in the training sample set includes: simulation motion parameters and simulation collision avoidance behavior parameters corresponding to the simulation motion parameters.

[0058] Step S25: Train the initial collision avoidance behavior prediction model using the training sample set to obtain the target collision avoidance behavior prediction model.

[0059] In the optional solutions provided in steps S24 to S25 above, the simulation experiment can be conducted based on simulation software, which may include, but is not limited to, VTD (Virtual Test Drive), PreScan, CarSim, and Vissim. The simulation motion parameters can be the target motion parameters corresponding to the autonomous vehicle determined using simulation software. The simulation collision avoidance behavior parameters can be the collision avoidance behavior parameters obtained through simulation testing using simulation software. The initial collision avoidance behavior prediction model can be a collision avoidance behavior prediction model that has not yet been trained using a training sample set.

[0060] In an optional embodiment, in step S24, the autonomous vehicle is simulated using a simulation test drive to construct a training sample set, including:

[0061] Step S241: Conduct a simulation test on the autonomous vehicle using the simulation test driver to obtain the test log;

[0062] Step S242: Based on multiple simulated driving segments during the simulation test, target test data is obtained by filtering from the test log;

[0063] Step S243: Construct a training sample set using the target experimental data.

[0064] In the optional solutions provided by steps S241 to S243 above, the data in the aforementioned test log may include, but is not limited to: the motion parameters of the autonomous vehicle and the parameter values ​​collected by various sensors of the vehicle. Each of the aforementioned multiple simulated driving segments may start at the moment when the vehicle speed is greater than zero and end at the moment when the vehicle's position no longer changes after the vehicle brakes to a stop. The aforementioned target experimental data may be the data in the test log corresponding to the aforementioned multiple simulated driving segments (including but not limited to: the motion parameters of the aforementioned autonomous vehicle and the parameter values ​​collected by various sensors of the aforementioned vehicle). It should also be noted that the selected target experimental data can be used to construct a training sample set, a validation set, and a test set. Specifically, for example, 80% of the target experimental data can be used as the training sample set, 10% as the validation set, and 10% as the test set.

[0065] In the above optional embodiments, the technical effects that can be achieved are: based on the data contained in the test log, the simulated driving segments are divided, the target driving segments are determined from the simulated driving segments, and then the target test data corresponding to the target driving segments are obtained from the test log. Invalid test data (such as driving segments in which the main vehicle is always in a normal state) can be eliminated, thereby avoiding the problem of network training being stuck or terminated due to redundancy of target test data, and improving the efficiency of network training and the accuracy of training results.

[0066] In an optional embodiment, in step S25, the initial collision avoidance behavior prediction model is trained using a training sample set to obtain the target collision avoidance behavior prediction model, including:

[0067] Step S251: Based on the simulated motion parameters in the training sample set, the initial collision avoidance behavior prediction model is used to predict the collision avoidance behavior, and the collision avoidance behavior prediction vector corresponding to the simulated motion parameters is obtained.

[0068] Step S252: Calculate the training loss based on the simulated collision avoidance behavior parameters and collision avoidance behavior prediction vectors corresponding to the simulated motion parameters;

[0069] Step S253: Adjust the network parameters of the initial collision avoidance behavior prediction model based on the training loss to obtain the target collision avoidance behavior prediction model.

[0070] In the optional solutions provided in steps S251 to S253 above, the aforementioned collision avoidance behavior prediction vector may include, but is not limited to: the feature vector output by the collision avoidance behavior prediction model (including but not limited to: the initial collision avoidance behavior prediction model, the intermediate collision avoidance behavior prediction model updated during training), and the weighted feature vector obtained after reweighting the aforementioned feature vector. The aforementioned training loss may be the Mean-Square Error (MSE) loss function. It should also be noted that the MSE loss function can be used to determine whether the collision avoidance behavior model has converged. Specifically, for example, during training, the changes in the MSE loss function of the training sample set and the validation set with the number of training iterations are observed. When the function value of the MSE loss function tends to zero and remains stable, it can be determined that the collision avoidance behavior model has been sufficiently trained. Furthermore, the test set can be input into the collision avoidance behavior model to verify the model's collision avoidance behavior prediction capability.

[0071] In the above optional embodiments, the technical effect that can be achieved is: the training loss is calculated based on the simulated collision avoidance behavior parameters and the collision avoidance behavior prediction vector, and the network parameters of the collision avoidance behavior prediction model are continuously adjusted according to the training loss, which can improve the accuracy of the network parameters of the collision avoidance behavior prediction model, and thus improve the collision avoidance behavior prediction capability of the collision avoidance behavior prediction model.

[0072] In an optional embodiment, in step S251, the target collision avoidance behavior prediction model includes a long short-term memory sub-model and an attention mechanism sub-module. The initial collision avoidance behavior prediction model has the same network structure as the target collision avoidance behavior prediction model. Based on the simulated motion parameters in the training sample set, the initial collision avoidance behavior prediction model is used to predict the collision avoidance behavior, and the resulting collision avoidance behavior prediction vector corresponding to the simulated motion parameters includes:

[0073] Step S2511: Using the long short-term memory sub-model in the initial collision avoidance behavior prediction model, the simulation motion parameters in the training sample set are subjected to time correlation learning to generate the first feature vector corresponding to the simulation motion parameters, wherein the weights of multiple feature values ​​in the first feature vector are evenly distributed.

[0074] Step S2512: Using the attention mechanism submodule in the initial collision avoidance behavior prediction model, the weights of multiple feature values ​​in the first feature vector are adjusted to obtain the second feature vector;

[0075] Step S2513: Perform fully connected regression processing on the second feature vector to obtain the collision avoidance behavior prediction vector corresponding to the simulated motion parameters.

[0076] The following combination Figure 3 , Figure 4 The above methods will be further explained.

[0077] Figure 3 This is a schematic diagram of an optional collision avoidance behavior prediction model for an autonomous vehicle according to an embodiment of the present invention. Figure 4 This is a schematic diagram of another optional collision avoidance behavior prediction model for an autonomous vehicle according to an embodiment of the present invention, such as... Figure 3 , Figure 4 As shown, the collision avoidance behavior prediction model for autonomous vehicles can include: a two-layer LSTM network unit (including multiple LSTM networks), an attention mechanism unit, and a fully connected regression layer. The two-layer LSTM network unit can process simulated motion parameters in the training sample set (…). Figure 3 The matrix shown is [x1, x2, ..., x... n Temporal correlation learning is performed, and the first feature vector corresponding to the simulated motion parameters is generated. Figure 4 h shown t Furthermore, the attention mechanism unit is used to adjust the weights of multiple feature values ​​in the first feature vector to obtain the calibrated feature vector. Furthermore, the calibrated feature vector Inputting a fully connected neural network with a fully connected regression layer yields a feature vector as the output. The various feature parameters are then integrated to output the collision avoidance behavior prediction vector. Figure 3 The matrix shown is [y1, y2, ..., y]. m ]).

[0078] In an optional embodiment, in step S2511, the long short-term memory sub-model in the initial collision avoidance behavior prediction model is used to perform time-correlation learning on the simulated motion parameters in the training sample set, generating a first feature vector corresponding to the simulated motion parameters, including:

[0079] Step S25111: In the multiple layer structures of the Long Short-Term Memory (LSTM) sub-model, temporal correlation learning is performed on the simulated motion parameters in the training sample set at multiple consecutive time points to generate a first feature vector; wherein, each layer structure includes at least: a forgetting gate component, an input gate component, and an output gate component; the forgetting gate component is used to: during the temporal correlation learning process, based on the simulated motion parameters output by the previous layer structure at the current time point and the output feature vector of the current layer structure at the previous time point, determine the portion of network state parameters to be retained at the current time point from the output network state parameters of the previous time point in the current layer structure; the input gate component is used to: during the temporal correlation learning process, based on the simulated motion parameters output by the previous layer structure at the current time point... The simulation motion parameters and the output feature vector of the current layer structure at the previous time step are used to determine the second part of the network parameters to be updated at the current time step in the current layer structure, and to create candidate network state parameters at the current time step in the current layer structure. The candidate network state parameters are used to determine the second part of the network state parameters to be updated at the current time step, and the output network state parameters at the current time step are determined based on the first part of the network state parameters and the second part of the network state parameters. The output gating component is used to: determine the output feature vector at the current time step in the current layer structure based on the simulation motion parameters output by the previous layer structure at the current time step and the output feature vector of the current layer structure at the previous time step in the temporal correlation learning process.

[0080] The following combination Figure 4 , Figure 5 The above methods will be further explained.

[0081] Figure 5 This is a schematic diagram of another optional collision avoidance behavior prediction model for an autonomous vehicle according to an embodiment of the present invention, such as... Figure 4 , Figure 5 As shown, LSTM networks can perform time series prediction in a chain-like, iterative manner. Specifically, at time t, the input to a neuron in a certain layer of the LSTM network includes the variable x output by the previous layer of the LSTM network at the current time. t The variable h output by the current LSTM network layer at time t-1t-1 It's also worth noting that LSTM neural networks incorporate multiple gating mechanisms and unit states C, which can autonomously determine the degree to which feature information is retained. Specifically, these gating mechanisms include a forget gate, used to store the calculated variable f... t As a gating mechanism, it determines the cell state C corresponding to time t-1. t-1 The state C of the unit at time t t The discarded part; the input gate, which can selectively "memorize" input variables, and can be used to select the input variable x of the network at time t. t Retained in cell state C t Furthermore, the forget gate can be combined with the output of the input gate and passed to the next state (including but not limited to: as input to the next LSTM network layer, as input to subsequent output gates); the input gate can be used to determine the output variable h at the current time step. t Unit state C t .

[0082] Still as Figure 4 As shown, the information discarded by the forget gate can be represented by the following formula (1):

[0083] f t =σ(W f ·[h t-1 x t ]+b f ) Formula (1)

[0084] In the above formula (1), W f Let [h] be the weight matrix. t-1 x t The symbol ] represents concatenating vector h with vector x, where vector b is the bias term and σ is the sigmoid activation function. It's worth noting that the sigmoid activation function can be used to map input variables to an output between 0 and 1.

[0085] Still as Figure 4 As shown, further, the information stored in the cell state is calculated using the input gate. Specifically, for example, the updated value i is first determined based on the sigmoid layer. t Then, the tanh layer of the input gate creates a new candidate value vector. As shown in formulas (2) and (3) below:

[0086] i t =σ(W i ·[h t-1 x t ]+b i ) Formula (2)

[0087]

[0088] Furthermore, the cell state C corresponding to time t-1 is... t-1 With the output variable f of the forget gate t Multiply, and add the Hadamard product to the result. Obtain the network state parameters C at the current moment. t It can be shown in the following formula (4):

[0089]

[0090] Still as Figure 4 As shown, further, based on the gating mechanism of the output gate, the output variable of the output gate is determined. Specifically, the sigmoid layer is used to select the output portion of the cell state, and then the selection result is calculated with the cell state processed by the tanh layer of the output gate to obtain the output variable h of the output gate. t It can be shown in the following formulas (5) and (6):

[0091] O t =σ(W o ·[h t-1 x t ]+b o ) Formula (5)

[0092]

[0093] In an optional embodiment, in step S2512, the attention mechanism submodule in the initial collision avoidance behavior prediction model is used to adjust the weights of multiple feature values ​​in the first feature vector to obtain the second feature vector, which includes:

[0094] Step S25121: Using the fully connected network parameters and preset activation function in the attention mechanism submodule, perform dependency analysis on multiple feature values ​​in the first feature vector to generate the target weight set corresponding to the first feature vector;

[0095] Step S25122: Update the weights of multiple feature values ​​using the target weight set to obtain the second feature vector.

[0096] In the optional technical solutions provided in steps S25121 to S25122 above, the fully connected network parameters and the preset activation function can be redistributed to each weight parameter in the weight matrix W according to the importance of each feature parameter, so as to guide the collision avoidance behavior prediction model to dynamically focus on the feature parameters that have a greater influence on collision avoidance behavior, thereby enhancing the prediction ability of the collision avoidance behavior prediction model. It should also be noted that the preset activation function may include, but is not limited to, the Rectified Linear Unit (ReLU) activation function and the sigmoid activation function. The ReLU activation function can be used to introduce nonlinear characteristics to increase the complexity of the functional relationship in the neural network, and the sigmoid activation function can be used to map the input variable to an output variable between 0 and 1. The target weight set can be a new set of weight parameters obtained after redistributing the weight parameters of each feature parameter.

[0097] In an optional embodiment, in step S22, based on the target motion parameters, the collision avoidance behavior prediction model is used to predict the collision avoidance behavior of the autonomous vehicle, and the target collision avoidance behavior parameters corresponding to the target motion parameters are obtained as follows:

[0098] Step S221: Using the long short-term memory sub-model in the target collision avoidance behavior prediction model, the target motion parameters are learned in terms of temporal correlation to generate the third feature vector corresponding to the target motion parameters. The weights of multiple feature values ​​in the third feature vector are evenly distributed.

[0099] Step S222: Using the attention mechanism submodule in the target collision avoidance behavior prediction model, the weights of multiple feature values ​​in the third feature vector are adjusted to obtain the fourth feature vector;

[0100] Step S223: Perform fully connected regression on the fourth feature vector to obtain the collision avoidance behavior target vector corresponding to the target motion parameters;

[0101] Step S224: Use the collision avoidance behavior target vector to determine the target collision avoidance behavior parameters corresponding to the target motion parameters.

[0102] Still as Figure 3 , Figure 4 , Figure 5 As shown, the attention mechanism can include an "incentive" unit and a "weighted" unit. The "incentive" unit can be a self-gating mechanism, comprising two activation functions ( Figure 5 The ReLU activation function and sigmoid activation function shown), and two fully connected layers ( Figure 5As shown in the diagram (fully connected layer 1 and fully connected layer 2), it should also be noted that fully connected layer 1 can use compression ratio to perform dimensionality reduction on the data, while fully connected layer 2 can return the dimensionality-reduced data to its original dimension.

[0103] Specifically, the input variable of the attention mechanism is the output variable h of the two-layer LSTM network unit, the fully connected network parameters are denoted as E1 and E2, the set of weights of each feature parameter obtained by reallocation based on the fully connected network parameters is denoted as s, the ReLU activation function is denoted as δ, the sigmoid activation function is denoted as σ, and the processing of the "activation" unit can be shown in the following formula (7):

[0104] s=σ(E2δ(E1h)) Formula (7)

[0105] Furthermore, the "weighting" unit processes the feature vector h output by the two-layer LSTM network unit using the weight set s obtained from the "excitation" unit, as shown in the following formula (8):

[0106]

[0107] Furthermore, the feature vector output by the attention mechanism unit... Each feature parameter in the algorithm is input into a fully connected regression layer. By integrating these feature parameters, the collision avoidance behavior target vector y is output.

[0108] In the above optional embodiments, the technical effects that can be achieved are: the attention mechanism can fully capture the dependencies between the various feature parameters of the feature vector output by the two-layer LSTM network unit; the activation function and fully connected layer can limit the complexity of the collision avoidance behavior prediction model, thereby helping the neural network to generalize better and thus improving the applicability of the collision avoidance behavior prediction model; and, by filtering and reweighting the output variables of the two-layer LSTM network unit, the model parameters that have a significant impact on collision avoidance prediction in the collision avoidance behavior prediction model can be improved, thereby improving the prediction ability of the collision avoidance behavior prediction model.

[0109] The following combination Figure 6 The method provided by the present invention will be further described.

[0110] Figure 6 This is a flowchart of another optional process for testing an autonomous vehicle according to an embodiment of the present invention, such as... Figure 6As shown, in the optional technical solution provided by the present invention, a CPNCO scenario is constructed using VTD simulation software to simulate an automatic emergency braking (AEB) driving scenario under actual dangerous road conditions. During the simulation process, the AEB algorithm can be connected to control the autonomous vehicle to avoid collisions, and test data can be collected at a preset sampling frequency (e.g., 50Hz) through the RDB Viewer interface tool inside the VTD simulation software.

[0111] Furthermore, as an optional implementation, the main vehicle is set to travel at a constant speed of 20-60 km / h, and pedestrians cross the main vehicle's path at a constant speed of 4-6 km / h. Additionally, the pedestrians' crossing direction is set to be at different angles to the front direction of the main vehicle. It should be noted that the technical solution provided by this invention can be tested multiple times by adjusting motion parameters such as the main vehicle's speed, the pedestrian's speed, and the relative positions of the main vehicle and pedestrians.

[0112] Furthermore, as an optional implementation, the motion state of the main vehicle and pedestrians can be observed through the Scenario Editor 2D interface and IG 3D interface of the simulation software to determine whether the main vehicle has completed collision avoidance behavior under different test conditions, and save the test data obtained in each simulation test.

[0113] Furthermore, the test data is screened to determine the input parameters of the initial collision avoidance behavior prediction model. These input parameters may include, but are not limited to: the vehicle's speed, the pedestrian's speed, the vehicle's deceleration, the pedestrian's deceleration, and the relative speed between the vehicle and the pedestrian. The output parameters of the initial collision avoidance behavior prediction model may include, but are not limited to: the vehicle's AEB activation state, the time-to-collision (TTC) at the AEB activation moment, and the longitudinal spacing.

[0114] Furthermore, during the experiment, training data corresponding to driving segments that started when the vehicle's speed was greater than zero and ended when the vehicle's position no longer changed after braking to a stop were selected as valid experimental data. Based on this valid experimental data, training, validation, and test sets were obtained. Next, the training and validation sets were input into the intelligent collision avoidance behavior prediction model (i.e., the initial collision avoidance behavior prediction model mentioned above), and the intelligent collision avoidance behavior prediction model was trained.

[0115] Furthermore, during the training process, the value of the loss function is observed, and the convergence of the intelligent collision avoidance behavior prediction model is determined based on the value of the loss function. When the value of the loss function approaches zero and remains stable during a certain training process, the current intelligent collision avoidance behavior prediction model can be determined as the target collision avoidance behavior prediction model. Furthermore, the above test set is input into the current intelligent collision avoidance behavior prediction model to test the model's collision avoidance prediction capability.

[0116] The technical effects achieved by the technical solution provided by this invention are as follows:

[0117] (1) Based on the LSTM neural network model, the vehicle collision avoidance behavior is adaptively predicted, and the test data is intelligently processed and analyzed to update the collision avoidance behavior prediction model, which reduces the complexity of the collision avoidance behavior prediction model and improves its applicability.

[0118] (2) Based on the LSTM network, the potential correlation of vehicle collision avoidance behavior data is determined from the time dimension, which improves the prediction ability of the collision avoidance behavior prediction model.

[0119] (3) Based on the attention mechanism, the feature parameters that have a significant impact on collision avoidance behavior are determined, which improves the learning ability and accuracy of the collision avoidance behavior prediction model.

[0120] In this embodiment, an apparatus for testing autonomous vehicles is also provided. This apparatus is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, a "module" is a combination of software and / or hardware that can perform a predetermined function. Although the apparatus described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0121] Figure 7 This is a structural block diagram of an optional apparatus for testing autonomous vehicles according to an embodiment of the present invention, such as... Figure 7 As shown, the device includes:

[0122] The acquisition module 701 is used to acquire target motion parameters corresponding to the autonomous vehicle under multiple collision avoidance test conditions; the prediction module 702 is used to predict the collision avoidance behavior of the autonomous vehicle based on the target motion parameters and using a target collision avoidance behavior prediction model to obtain the target collision avoidance behavior parameters corresponding to the target motion parameters. The target collision avoidance behavior prediction model is obtained through machine learning using a training sample set, and the target collision avoidance behavior parameters are used to characterize the collision avoidance behavior of the autonomous vehicle under multiple collision avoidance test conditions; the generation module 703 is used to generate test results of the autonomous vehicle using the target motion parameters and the target collision avoidance behavior parameters. The test results are used to characterize the emergency collision avoidance performance of the autonomous vehicle.

[0123] Optionally, Figure 8 This is a structural block diagram of another optional apparatus for testing autonomous vehicles according to an embodiment of the present invention, such as... Figure 8 As shown, the device includes Figure 7 In addition to all the modules shown, it also includes: a simulation module 704, used to conduct simulation tests on autonomous vehicles using simulation test drive, and to build a training sample set, wherein each training sample in the training sample set includes: simulation motion parameters and simulation collision avoidance behavior parameters corresponding to the simulation motion parameters; and to train the initial collision avoidance behavior prediction model using the training sample set to obtain the target collision avoidance behavior prediction model.

[0124] Optionally, the simulation module 704 is further configured to: conduct simulation tests on autonomous vehicles using simulation test drive, and construct a training sample set, including: conducting simulation tests on autonomous vehicles using simulation test drive to obtain test logs; selecting target test data from the test logs based on multiple simulated driving segments during the simulation test; and constructing a training sample set using the target test data.

[0125] Optionally, the simulation module 704 is further configured to: train the initial collision avoidance behavior prediction model using the training sample set to obtain the target collision avoidance behavior prediction model, including: predicting collision avoidance behavior using the initial collision avoidance behavior prediction model based on the simulated motion parameters in the training sample set, and obtaining the collision avoidance behavior prediction vector corresponding to the simulated motion parameters; calculating the training loss based on the simulated collision avoidance behavior parameters and the collision avoidance behavior prediction vector corresponding to the simulated motion parameters; and adjusting the network parameters of the initial collision avoidance behavior prediction model based on the training loss to obtain the target collision avoidance behavior prediction model.

[0126] Optionally, the simulation module 704 is further configured to: The target collision avoidance behavior prediction model includes a long short-term memory sub-model and an attention mechanism sub-module. The initial collision avoidance behavior prediction model has the same network structure as the target collision avoidance behavior prediction model. Based on the simulated motion parameters in the training sample set, the initial collision avoidance behavior prediction model is used to predict collision avoidance behavior, obtaining the collision avoidance behavior prediction vector corresponding to the simulated motion parameters. This includes: using the long short-term memory sub-model in the initial collision avoidance behavior prediction model to perform temporal correlation learning on the simulated motion parameters in the training sample set, generating a first feature vector corresponding to the simulated motion parameters, wherein the weights of multiple feature values ​​in the first feature vector are evenly distributed; using the attention mechanism sub-module in the initial collision avoidance behavior prediction model to adjust the weights of multiple feature values ​​in the first feature vector, obtaining a second feature vector; and performing fully connected regression processing on the second feature vector to obtain the collision avoidance behavior prediction vector corresponding to the simulated motion parameters.

[0127] Optionally, the simulation module 704 is further configured to: employ the Long Short-Term Memory (LSTM) sub-model in the initial collision avoidance behavior prediction model to perform temporal correlation learning on the simulated motion parameters in the training sample set, generating a first feature vector corresponding to the simulated motion parameters. This includes: performing temporal correlation learning on the simulated motion parameters in the training sample set at multiple consecutive time points in the multiple layer structures of the LSM sub-model to generate a first feature vector; wherein each layer structure includes at least: a forgetting gate component, an input gate component, and an output gate component; the forgetting gate component is configured to: during the temporal correlation learning process, based on the simulated motion parameters output by the previous layer structure at the current time point and the output feature vector of the current layer structure at the previous time point, determine the portion of network state parameters to be retained at the current time point from the output network state parameters at the previous time point in the current layer structure. The input gating component is used to: determine the second part of the network parameters to be updated at the current time based on the simulated motion parameters output by the previous layer and the output feature vector of the current layer at the previous time during the temporal correlation learning process; create candidate network state parameters for the current time in the current layer; determine the second part of the network state parameters to be updated at the current time using the candidate network state parameters; and determine the output network state parameters at the current time based on the first part of the network state parameters and the second part of the network state parameters. The output gating component is used to: determine the output feature vector at the current time using the output network state parameters at the current time during the temporal correlation learning process, based on the simulated motion parameters output by the previous layer and the output feature vector of the current layer at the previous time.

[0128] Optionally, the simulation module 704 is further configured to: adjust the weights of multiple feature values ​​in the first feature vector using the attention mechanism submodule in the initial collision avoidance behavior prediction model to obtain the second feature vector, including: performing dependency analysis on multiple feature values ​​in the first feature vector using the fully connected network parameters and preset activation function in the attention mechanism submodule to generate a target weight set corresponding to the first feature vector; and updating the weights of multiple feature values ​​using the target weight set to obtain the second feature vector.

[0129] Optionally, the prediction module 702 is further configured to: predict the collision avoidance behavior of the autonomous vehicle based on the target motion parameters using a target collision avoidance behavior prediction model, and obtain the target collision avoidance behavior parameters corresponding to the target motion parameters, including: using the long short-term memory sub-model in the target collision avoidance behavior prediction model to perform temporal correlation learning on the target motion parameters, generating a third feature vector corresponding to the target motion parameters, wherein the weights of multiple feature values ​​in the third feature vector are evenly distributed; using the attention mechanism sub-module in the target collision avoidance behavior prediction model to adjust the weights of multiple feature values ​​in the third feature vector to obtain a fourth feature vector; performing fully connected regression processing on the fourth feature vector to obtain the collision avoidance behavior target vector corresponding to the target motion parameters; and using the collision avoidance behavior target vector to determine the target collision avoidance behavior parameters corresponding to the target motion parameters.

[0130] It should be noted that the above modules can be implemented by software or hardware. For the latter, they can be implemented in the following ways, but are not limited to: all the above modules are located in the same processor; or, the above modules are located in different processors in any combination.

[0131] According to another aspect of the present invention, an electronic device is also provided, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method of testing an autonomous vehicle as described in any of the preceding embodiments.

[0132] Optionally, in this embodiment, the memory may be configured to store a computer program for performing the following steps:

[0133] Step S1: Obtain the target motion parameters corresponding to the autonomous vehicle under multiple collision avoidance test conditions;

[0134] Step S2: Based on the target motion parameters, the collision avoidance behavior prediction model is used to predict the collision avoidance behavior of the autonomous vehicle, and the target collision avoidance behavior parameters corresponding to the target motion parameters are obtained. The target collision avoidance behavior prediction model is obtained by machine learning using a training sample set, and the target collision avoidance behavior parameters are used to characterize the collision avoidance behavior of the autonomous vehicle under multiple collision avoidance test conditions.

[0135] Step S3: Using the target motion parameters and target collision avoidance behavior parameters, generate test results for the autonomous vehicle, wherein the test results are used to characterize the emergency collision avoidance performance of the autonomous vehicle.

[0136] Optionally, in this embodiment, the aforementioned memory may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0137] Optionally, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0138] Step S1: Obtain the target motion parameters corresponding to the autonomous vehicle under multiple collision avoidance test conditions;

[0139] Step S2: Based on the target motion parameters, the collision avoidance behavior prediction model is used to predict the collision avoidance behavior of the autonomous vehicle, and the target collision avoidance behavior parameters corresponding to the target motion parameters are obtained. The target collision avoidance behavior prediction model is obtained by machine learning using a training sample set, and the target collision avoidance behavior parameters are used to characterize the collision avoidance behavior of the autonomous vehicle under multiple collision avoidance test conditions.

[0140] Step S3: Using the target motion parameters and target collision avoidance behavior parameters, generate test results for the autonomous vehicle, wherein the test results are used to characterize the emergency collision avoidance performance of the autonomous vehicle.

[0141] Optionally, specific examples in this embodiment can refer to the examples described in the above embodiments and their optional implementations, which will not be repeated here.

[0142] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0143] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0144] In the several embodiments provided by this invention, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection can be through some interfaces; the indirect coupling or communication connection of units or modules can be electrical or other forms.

[0145] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0146] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0147] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0148] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for testing autonomous vehicles, characterized in that, include: Simulation tests were conducted on autonomous vehicles using a simulation-driven approach, and test logs were obtained. Based on multiple simulated driving segments during the simulation test, target test data is obtained by filtering from the test log. Each of the multiple simulated driving segments starts at the moment when the speed of the autonomous vehicle is greater than zero and ends at the moment after the autonomous vehicle brakes and stops for a preset time. A training sample set is constructed using the target test data, wherein each training sample in the training sample set includes: simulated motion parameters and simulated collision avoidance behavior parameters corresponding to the simulated motion parameters; The initial collision avoidance behavior prediction model is trained using the training sample set to obtain the target collision avoidance behavior prediction model; Obtain the target motion parameters of the autonomous vehicle under multiple collision avoidance test conditions; Based on the target motion parameters, the collision avoidance behavior prediction model is used to predict the collision avoidance behavior of the autonomous vehicle, thereby obtaining the target collision avoidance behavior parameters corresponding to the target motion parameters. The target collision avoidance behavior prediction model is obtained through machine learning using a training sample set, and the target collision avoidance behavior parameters are used to characterize the collision avoidance behavior of the autonomous vehicle under multiple collision avoidance test conditions. Using the target motion parameters and the target collision avoidance behavior parameters, test results for the autonomous vehicle are generated, wherein the test results are used to characterize the emergency collision avoidance performance of the autonomous vehicle.

2. The method according to claim 1, characterized in that, The initial collision avoidance behavior prediction model is trained using the training sample set to obtain the target collision avoidance behavior prediction model, which includes: Based on the simulated motion parameters in the training sample set, the initial collision avoidance behavior prediction model is used to predict the collision avoidance behavior, and the collision avoidance behavior prediction vector corresponding to the simulated motion parameters is obtained. The training loss is calculated based on the simulated collision avoidance behavior parameters corresponding to the simulated motion parameters and the collision avoidance behavior prediction vector. The network parameters of the initial collision avoidance behavior prediction model are adjusted based on the training loss to obtain the target collision avoidance behavior prediction model.

3. The method according to claim 2, characterized in that, The target collision avoidance behavior prediction model includes a long short-term memory sub-model and an attention mechanism sub-module. The initial collision avoidance behavior prediction model has the same network structure as the target collision avoidance behavior prediction model. Based on the simulated motion parameters in the training sample set, the initial collision avoidance behavior prediction model is used to predict collision avoidance behavior, resulting in a collision avoidance behavior prediction vector corresponding to the simulated motion parameters, including: Using the long short-term memory sub-model in the initial collision avoidance behavior prediction model, the simulation motion parameters in the training sample set are subjected to time correlation learning to generate a first feature vector corresponding to the simulation motion parameters, wherein the weights of multiple feature values ​​in the first feature vector are evenly distributed. Using the attention mechanism submodule in the initial collision avoidance behavior prediction model, the weights of the multiple feature values ​​in the first feature vector are adjusted to obtain the second feature vector; The second feature vector is subjected to fully connected regression processing to obtain the collision avoidance behavior prediction vector corresponding to the simulated motion parameters.

4. The method according to claim 3, characterized in that, Using the Long Short-Term Memory (LSTM) sub-model in the initial collision avoidance behavior prediction model, temporal correlation learning is performed on the simulated motion parameters in the training sample set to generate a first feature vector corresponding to the simulated motion parameters, including: In the multi-layer structure of the long short-term memory sub-model, the simulation motion parameters in the training sample set are learned temporally at multiple consecutive time points to generate the first feature vector; Each of the plurality of layer structures includes at least: a forgetting gating component, an input gating component, and an output gating component; The forgetting gating component is used to: during the temporal correlation learning process, based on the simulated motion parameters output by the previous layer structure at the current time and the output feature vector of the current layer structure at the previous time, determine the first part of the network state parameters to be retained at the current time from the output network state parameters at the previous time in the current layer structure; The input gating component is used to: during the temporal correlation learning process, based on the simulated motion parameters output by the previous layer structure at the current time and the output feature vector of the current layer structure at the previous time, create candidate network state parameters at the current time in the current layer structure, use the candidate network state parameters to determine the second part of network state parameters to be updated at the current time, and determine the output network state parameters at the current time based on the first part of network state parameters and the second part of network state parameters; The output gating component is used to: during the temporal correlation learning process, based on the simulated motion parameters output by the previous layer structure at the current time and the output feature vector of the current layer structure at the previous time, determine the output feature vector at the current time using the output network state parameters at the current time in the current layer structure.

5. The method according to claim 3, characterized in that, Using the attention mechanism submodule in the initial collision avoidance behavior prediction model, the weights of the multiple feature values ​​in the first feature vector are adjusted to obtain the second feature vector, which includes: Using the fully connected network parameters and preset activation function in the attention mechanism submodule, dependency analysis is performed on the multiple feature values ​​in the first feature vector to generate the target weight set corresponding to the first feature vector. The weights of the multiple feature values ​​are updated using the target weight set to obtain the second feature vector.

6. The method according to claim 3, characterized in that, Based on the target motion parameters, the collision avoidance behavior prediction model is used to predict the collision avoidance behavior of the autonomous vehicle, and the target collision avoidance behavior parameters corresponding to the target motion parameters are obtained as follows: The target collision avoidance behavior prediction model employs the long short-term memory sub-model to perform time-related learning on the target motion parameters, generating a third feature vector corresponding to the target motion parameters, wherein the weights of multiple feature values ​​in the third feature vector are evenly distributed. Using the attention mechanism submodule in the target collision avoidance behavior prediction model, the weights of the multiple feature values ​​in the third feature vector are adjusted to obtain the fourth feature vector; Perform fully connected regression processing on the fourth feature vector to obtain the collision avoidance behavior target vector corresponding to the target motion parameters; The target collision avoidance behavior parameters corresponding to the target motion parameters are determined using the target collision avoidance behavior target vector.

7. An apparatus for testing autonomous vehicles, characterized in that, include: The simulation module is used to conduct simulation tests on autonomous vehicles using simulation test drive, obtain test logs, select target test data from the test logs based on multiple simulated driving segments during the simulation test, construct a training sample set using the target test data, and train an initial collision avoidance behavior prediction model using the training sample set to obtain a target collision avoidance behavior prediction model. Each of the multiple simulated driving segments starts at the moment when the autonomous vehicle's speed is greater than zero and ends at the moment after the autonomous vehicle brakes to a stop for a preset time. Each training sample in the training sample set includes: simulated motion parameters and simulated collision avoidance behavior parameters corresponding to the simulated motion parameters. The acquisition module is used to acquire the target motion parameters corresponding to the autonomous vehicle under multiple collision avoidance test conditions. The prediction module is used to predict the collision avoidance behavior of the autonomous vehicle based on the target motion parameters and the target collision avoidance behavior prediction model, so as to obtain the target collision avoidance behavior parameters corresponding to the target motion parameters. The target collision avoidance behavior prediction model is obtained by machine learning using a training sample set, and the target collision avoidance behavior parameters are used to characterize the collision avoidance behavior of the autonomous vehicle under multiple collision avoidance test conditions. A generation module is used to generate test results for the autonomous vehicle using the target motion parameters and the target collision avoidance behavior parameters, wherein the test results are used to characterize the emergency collision avoidance performance of the autonomous vehicle.

8. An electronic device, characterized in that, The device includes a memory and a processor, characterized in that the memory stores a computer program, and the processor is configured to run the computer program to perform the method for testing an autonomous vehicle according to any one of claims 1 to 6.

Images