Method and device for generating automatic driving decision based on hybrid expert model

By combining expert models with LSTM and Transformer to process driving data of the vehicle and surrounding vehicles, future driving trajectories are generated. This solves the problem of insufficient scenario expansion capability of existing autonomous driving decision-making and planning models, and achieves more accurate autonomous driving decisions.

CN118928464APending Publication Date: 2024-11-12CHONGQING CHANGAN TECH CO LTD
View PDF 0 Cites 4 Cited by

Patent Information

Application Number
CN202410998641.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

Existing autonomous driving decision-making and planning models have poor scenario expansion capabilities, making it difficult to meet the needs of various complex driving scenarios.

Method used

A hybrid expert model is adopted, in which the driving data of the autonomous vehicle and surrounding vehicles are input into the hybrid expert model. The future driving data is predicted by multiple expert models, driving trajectory is generated and autonomous driving decision is made. The sparsity of the model is optimized by using gating network and backbone network, and the temporal information is processed by combining LSTM and Transformer models.

Benefits of technology

It improves the accuracy of autonomous driving decision-making and planning, adapts to various autonomous driving scenarios, and enhances the model's scenario expansion capabilities and generalization.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118928464A_ABST
    Figure CN118928464A_ABST
Patent Text Reader

Abstract

The invention relates to a method and device for generating an automatic driving decision based on a hybrid expert model, and the method comprises the steps: transmitting the driving data of a vehicle in a preset simulation environment to a hybrid expert model which is obtained through the modeling of a plurality of expert models, the plurality of expert models are used for jointly predicting a plurality of special domain automatic driving tasks; after training, the hybrid expert model is used for taking continuous time sequence driving data of the vehicle within a first preset duration before the current moment as input and outputting continuous time sequence driving data within a second preset duration after the current moment; based on the driving data as input of a hybrid expert model, obtaining an output result of the hybrid expert model; and generating a corresponding driving track based on the driving data, and making an automatic driving decision for the vehicle based on the driving track. According to the invention, the problem that the model scene expansion capability of automatic driving decision planning is poor in the prior art is solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of autonomous driving technology, and in particular to a method and device for generating autonomous driving decisions based on a hybrid expert model. Background Art

[0002] An autonomous vehicle is an intelligent integrated system that integrates environmental perception, decision planning, motion control, voice interaction, and wireless communication technologies. Decision planning technology is an important part of realizing autonomous driving. Since it involves both lateral and longitudinal motion, the vehicle's lane changing behavior is risky and logically complex. Current decision planning methods are mainly divided into rule-based and optimization-based methods. Among them, the rule-based method is limited by artificially designed and rigid rules, the lane changing strategy is too conservative, and it is only applicable to specific scenarios. The optimization-based method regards the lane changing decision planning problem as a constrained nonlinear, non-convex optimization problem. Due to the high difficulty of solving this optimization problem, this type of method is difficult to meet real-time requirements.

[0003] It can be seen that the decision-making and planning methods for autonomous driving in the existing technology have poor scenario expansion capabilities. Summary of the invention

[0004] The present application provides a method and device for generating autonomous driving decisions based on a hybrid expert model to solve the problem of poor model scenario expansion capability for autonomous driving decision planning in the prior art.

[0005] In a first aspect, the present application provides a method for generating autonomous driving decisions based on a hybrid expert model, comprising: sending driving data of a self-vehicle in a preset simulation environment to the hybrid expert model, wherein the driving data includes the driving data of the self-vehicle and the driving data of surrounding vehicles associated with the self-vehicle, and the hybrid expert model is modeled based on multiple expert models, and the multiple expert models are used to jointly predict multiple domain-specific autonomous driving tasks; after training, the hybrid expert model is used to take the continuous time series driving data of the vehicle within a first preset time length before the current moment as input, and output the continuous time series driving data within a second preset time length after the current moment; based on the driving data as the input of the hybrid expert model, the output result of the hybrid expert model is obtained, wherein the output result represents the continuous time series driving data of the self-vehicle within the preset time length after the current moment; based on the driving data, a corresponding driving trajectory is generated, and an autonomous driving decision is made for the self-vehicle based on the driving trajectory.

[0006] In a second aspect, the present application provides a device for generating autonomous driving decisions based on a hybrid expert model, comprising: a sending module, used to send driving data of a self-vehicle in a preset simulation environment to the hybrid expert model, wherein the driving data includes the driving data of the self-vehicle and the driving data of surrounding vehicles associated with the self-vehicle, and the hybrid expert model is modeled based on multiple expert models, and the multiple expert models are used to jointly predict multiple domain-specific autonomous driving tasks; after training, the hybrid expert model is used to take the continuous time series driving data of the vehicle within a first preset time length before the current moment as input, and output the continuous time series driving data within a second preset time length after the current moment; a first processing module, used to obtain the output result of the hybrid expert model based on the driving data as the input of the hybrid expert model, wherein the output result represents the continuous time series driving data of the self-vehicle within the preset time length after the current moment; a second processing module, used to generate a corresponding driving trajectory based on the driving data, and make an autonomous driving decision for the self-vehicle based on the driving trajectory.

[0007] In a third aspect, the present application provides an electronic device comprising: at least one communication interface; at least one bus connected to the at least one communication interface; at least one processor connected to the at least one bus; and at least one memory connected to the at least one bus, wherein the processor is configured to: execute the method for generating autonomous driving decisions based on a hybrid expert model as described in any one of the embodiments of the first aspect above.

[0008] In a fourth aspect, the present application also provides a computer storage medium storing computer executable instructions, wherein the computer executable instructions are used to execute the method for generating an autonomous driving decision based on a hybrid expert model as described in any one of the embodiments of the first aspect above.

[0009] Beneficial effects of the present application: In the embodiment of the present application, after sending the driving data of the vehicle in the preset simulation environment to the hybrid expert model, the driving data of the vehicle in a continuous time series within a preset time after the current moment output by the hybrid expert model can be obtained, and then the corresponding driving trajectory is generated based on the driving data, and the autonomous driving decision of the vehicle is made based on the driving trajectory. Since the hybrid expert model in the embodiment of the present application is modeled by multiple expert models, and the multiple models are used to jointly predict multiple domain-specific autonomous driving tasks, that is, the corresponding expert models can be set according to different demand tasks to adapt to different autonomous driving scenarios, the accuracy of autonomous driving decision planning is improved, thereby solving the technical problem of poor model scenario expansion capability for autonomous driving decision planning in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1A flowchart of a method for generating autonomous driving decisions based on a hybrid expert model provided in the present application;

[0011] Figure 2 One of the optional flow charts of a method for generating autonomous driving decisions based on a hybrid expert model provided in an embodiment of the present application;

[0012] Figure 3 A schematic diagram of the architecture of a hybrid expert model provided in an embodiment of the present application;

[0013] Figure 4 A schematic diagram of a processing flow based on a hybrid expert model provided in an embodiment of the present application;

[0014] Figure 5 A schematic diagram of the structure of the Transformer-Encoder provided in the embodiment of the present application;

[0015] Figure 6 A schematic diagram of the structure of the LSTM provided in the embodiment of the present application;

[0016] Figure 7 A second optional flowchart of a method for generating autonomous driving decisions based on a hybrid expert model provided in an embodiment of the present application;

[0017] Figure 8 A schematic diagram of extracting driving characteristic parameters provided in an embodiment of the present application;

[0018] Fig. 9 An optional flowchart of a method for generating autonomous driving decisions based on a hybrid expert model provided in an embodiment of the present application;

[0019] Fig.10 An update flow chart of trajectory prediction based on a hybrid expert model provided in an embodiment of the present application;

[0020] Fig.11 A flow chart of a method for decision-making and planning for autonomous driving of a vehicle based on a hybrid expert model provided in an embodiment of the present application;

[0021] Fig.12 A schematic diagram of data division provided in an embodiment of the present application;

[0022] Fig.13 A data preprocessing flow chart provided for an embodiment of the present application;

[0023] Fig.14 A schematic diagram of a device for generating an autonomous driving decision based on a hybrid expert model provided by the present application;

[0024] Fig.15 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0025] The following will describe the implementation methods of the present application with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. The present application can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be understood that the preferred embodiments are only for illustrating the present application, not for limiting the scope of protection of the present application.

[0026] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present application, and thus the drawings only show components related to the present application rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed at will, and the component layout may also be more complicated.

[0027] For ease of description, spatial relative terms may be used in the text to describe the relative position or movement of an element or feature as shown in the figure relative to another element or feature, such as "inside", "outside", "inner side", "outer side", "below", "below", "above", "above", "front", "back", etc. Such spatial relative terms are intended to include different orientations of the device in use or operation in addition to the orientation depicted in the figure. For example, if the device in the figure undergoes a position flip or a change in posture or a change in motion state, then these directional indications will also change accordingly, for example: an element described as "below other elements or features" or "below other elements or features" will subsequently be oriented as "above other elements or features" or "above other elements or features". Therefore, the example term "below..." can include both upper and lower orientations. The device can be oriented otherwise (rotated 90 degrees or in other directions) and the spatial relative descriptors used in the text are interpreted accordingly.

[0028] First, the relevant terms involved in the embodiments of the present application are explained.

[0029] Autonomous driving scenario: It is an organic combination and comprehensive reflection of scene elements such as the driving environment of the autonomous driving car, traffic participants and driving behavior. It can be decomposed into a sequence composed of one or more primitive scenes occurring in sequence, which will last for a period of time and has dynamic change characteristics.

[0030] Expert models are models built based on machine learning algorithms such as decision trees or random forests. These models perform well in complex data analysis scenarios. These models rely on the knowledge and experience of experts to build them. When building an expert model, its indicators, thresholds, and leaf node scores are set based on the knowledge and experience of experts, which enables the model to better adapt to specific business needs and data characteristics.

[0031] Specific functional scenarios: scenarios designed for typical functions of self-driving cars, such as following a vehicle, changing lanes, overtaking, and other functional scenarios.

[0032] Adaptive Moment Estimation (Adam): This algorithm is evolved from gradient descent. It combines the advantages of adaptive gradient and gradient descent algorithms, comprehensively considers the first-order moment estimation and second-order moment estimation of the gradient, calculates the update step size, ensures that the parameter update is not affected by the scaling transformation of the gradient, and can naturally implement the step size annealing process. It is used for problems where the gradient is sparse or the gradient has a lot of noise.

[0033] First-order moment estimation: Generally, the first-order origin moment and the first-order central moment, that is, the expectation and variance, are calculated according to the distribution of the population and the probability function. The function of the population moment is obtained and the function of the sample moment is used to replace the function of the population moment.

[0034] Time series model: A time series is a collection of data arranged in time order, which can be observed and collected at any time interval (for example, every second, every minute, every hour, every day, every month, every year, etc.). In statistics, time series analysis includes a series of methods for analyzing time series data, with the aim of extracting meaningful statistical information from the data and making predictions for the future.

[0035] In order to solve the technical problem that the model scenario expansion capability of autonomous driving decision planning in the prior art is poor, the present application provides a method for generating autonomous driving decisions based on a hybrid expert model, such as Figure 1 As shown, the steps of the method include:

[0036] Step 101, sending the driving data of the vehicle in the preset simulation environment to the hybrid expert model, wherein the driving data includes the driving data of the vehicle and the driving data of surrounding vehicles associated with the vehicle, and the hybrid expert model is modeled based on multiple expert models, and the multiple expert models are used to jointly predict multiple domain-specific autonomous driving tasks; after training, the hybrid expert model is used to take the continuous time series driving data of the vehicle within a first preset time length before the current moment as input, and output the continuous time series driving data within a second preset time length after the current moment;

[0037] In the embodiment of the present application, a corresponding simulation environment can be built based on VTD (Virtual Test Drive) according to the road scene of the HighD dataset, and the traffic flow generated by the VTD's built-in driver model can be used as the background test vehicle. The simulation environment data is read and input into the hybrid expert model through transmission control protocol (TCP) communication. The HighD dataset is a high-precision dataset for autonomous driving of automobiles. The dataset contains data from a variety of different automobile sensors, such as lidar, camera, and IMU, and can also include high-precision map data. The self-vehicle in the embodiment of the present application refers to the vehicle that currently needs to make an autonomous driving decision, and the driving data of the self-vehicle and the surroundings include driving data such as speed, acceleration, and current lane.

[0038] In addition, the expert model in the embodiment of the present application is used to jointly predict the domain-specific autonomous driving tasks. Different domain-specific autonomous driving tasks can correspond to different expert models. For example, the domain-specific autonomous driving tasks are following and changing lanes on the highway, corresponding to an expert model; the domain-specific autonomous driving tasks are straight or turning at the intersection, corresponding to an expert model; or the domain-specific autonomous driving tasks are on and off ramps, which can also correspond to an expert model. It can be seen that in the embodiment of the present application, specific expert models and their numbers can be set according to different application scenarios, that is, as long as the autonomous driving tasks during the vehicle driving process can be achieved through the expert model in the embodiment of the present application. Further, the first preset time length and the second preset time length in the embodiment of the present application can be set accordingly according to actual needs. For example, the first preset time length is 3s and the second preset time length is 5s, that is, the driving trajectory of 5s after the current moment can be predicted by the mitigation expert model through the driving data of the first 3s of the current moment.

[0039] Step 102, based on the driving data as the input of the hybrid expert model, obtaining the output result of the hybrid expert model, wherein the output result represents the driving data of the vehicle in a continuous time series within a preset time period after the current moment;

[0040] Step 103, generating a corresponding driving trajectory based on the driving data, and making an automatic driving decision for the vehicle based on the driving trajectory.

[0041] Through the above-mentioned steps 101 to 103 in the embodiment of the present application, after sending the driving data of the vehicle in the preset simulation environment to the hybrid expert model, the driving data of the vehicle in a continuous time series within a preset time after the current moment output by the hybrid expert model can be obtained, and then the corresponding driving trajectory is generated based on the driving data, and the autonomous driving decision of the vehicle is made based on the driving trajectory. Since the hybrid expert model in the embodiment of the present application is modeled by multiple expert models, and the multiple models are used to jointly predict multiple domain-specific autonomous driving tasks, that is, the corresponding expert models can be set according to different demand tasks to adapt to different autonomous driving scenarios, which improves the accuracy of autonomous driving decision planning, thereby solving the technical problem of poor model scenario expansion capability for autonomous driving decision planning in the prior art.

[0042] It should be noted that although multiple expert models can be set in the embodiment of the present application, in actual needs, only some of the expert models can be trained, and then driving data prediction can be performed based on the trained expert models. Other untrained expert models can be used as functions to be expanded.

[0043] Before the driving data is predicted, the hybrid expert model in the embodiment of the present application must first be trained, and before the training, it needs to be modeled. Therefore, the method of training the hybrid expert model involved in the embodiment of the present application is as follows: Figure 2 The steps shown may include:

[0044] Step 201, constructing a driving data set; wherein the driving data set includes a plurality of trajectory data, each trajectory data includes current vehicle data and weekly vehicle data associated with the vehicle;

[0045] The driving data set constructed in the embodiment of the present application can also be obtained from the data set HighD. The constructed posture data set includes multiple trajectory data. Taking the autonomous driving tasks of following a vehicle and changing lanes as an example, the multiple trajectory data are used to determine whether the current vehicle is following the surrounding vehicle, or whether it is changing from following the vehicle to changing lanes based on certain situations.

[0046] Step 202, modeling a hybrid expert model, wherein the modeled hybrid expert model includes a gating network, a backbone network, and multiple expert models;

[0047] In the embodiment of the present application, the distribution of training data among different expert models is achieved through a gating network, so that each expert model can focus on processing the task it is best at, thereby achieving the sparsity of the model; the gating network, as part of the hybrid expert model, receives a single data element as input and then outputs a weight, which represents the contribution of each expert model to processing the input data; the expert model will selectively input the corresponding data through the gating network to achieve its corresponding expert task.

[0048] In this regard, the above method of modeling the hybrid expert model may further include the following steps:

[0049] Step 11, setting a long short-term memory network (Long Short-Term Memory, LSTM) and a Transformer model in the backbone network, wherein the Transformer model and the LSTM adopt a fully connected layer structure, and the output of the Transformer model is the input of the LSTM;

[0050] Step 12: Determine the corresponding expert model according to the task in the autonomous driving scenario, and set the number of outputs of the gating network to be consistent with the number of expert models, where the output of the LSTM is the input of the gating network, and the output of the gating network represents the evaluation probability of the corresponding expert model.

[0051] In this regard, the basic architecture of the hybrid expert model in the specific example is as follows Figure 3 As shown in the figure, GateNetwork is a gated network. The number of outputs of the gated network corresponds to the number of experts in the expert network. The output of the gated network is G(X). g1 to g4 represent the probability evaluation of the gated network on the corresponding experts. The gated output in round t of training is as shown in Formula 1. The input x represents the feature vector of the input model, and the output is the value that the expert model wants to fit or classify. The entire model can be regarded as multiple optimization tasks, namely the classification task of the gated network and the regression or classification tasks of multiple experts.

[0052]

[0053] In the embodiments of the present application, the actual application scenarios of autonomous driving tasks are combined, experts are divided according to the task type (lane changing, following a vehicle), and LSTM and Transformer, which can handle timing information well, are used as the backbone network of the overall architecture. Both the expert network and the gated network adopt a simple fully connected layer structure. This can effectively reduce the size occupied by the model and shorten the running time of inference.

[0054] The hybrid expert model architecture in the embodiment of this application is as follows Figure 4As shown, the gray part is what has been implemented in this application, the white functional blocks are expandable functional blocks, the output of the model is the motion trajectory after 5 seconds, and all the gating networks and backbone networks share the same backbone network.

[0055] The backbone network directly connects the output of the transformer-encoder to the input of the LSTM. The transformer module structure is as follows: Figure 5 As shown in the figure, Embedding represents word encoding before feature input, Positional represents position encoding, Multi-HeadAttention represents multi-head attention module, FeedForward is a feedforward neural network, and N represents multiple layers of this structure; after data is encoded, multi-head attention, and feedforward neural network, the feature relationship between data can be effectively extracted, and intermediate features can be constructed as LSTM input. LSTM can effectively transmit and express information in time series, and this process will not cause long-term useful information to be ignored. At the same time, it can also solve the gradient disappearance and gradient explosion problems in time series neural networks. The LSTM structure is as follows: Figure 6 As shown, X t is the intermediate feature vector of the input, C t-1 The memory unit left over from the previous layer, H t-1 It is the hidden state of the previous layer. The solid-line frame is the forget gate, and the dotted-line frame is the memory gate.

[0056] Step 203 , training the hybrid expert model based on the driving data set, so that the hybrid expert model takes the continuous time series driving data of the vehicle within a first preset time period before the current moment as input, and outputs the continuous time series driving data within a second preset time period after the current moment.

[0057] The method of training the hybrid expert model in step 203 may include the following steps:

[0058] Step 21, determining a loss function according to the task in the autonomous driving scenario, and initializing parameters in the hybrid expert model after modeling;

[0059] It can be seen that after completing the modeling of the hybrid expert model, it is necessary to design a multi-task joint optimization solution to solve the training convergence problem. First, it involves two main tasks: the classification of driving function scenarios (following the car, changing lanes), and the regression of future trajectories. The loss function is calculated as shown in formula (2). The front of the right side of the formula is the mean square error loss function of the expert output part, where is the true label, y i is the predicted value, n is the number of samples, and the right side of the equation is the cross entropy loss function of the output part of the gating network, where y ic The true label, pic is the predicted value and c is the number of categories. In the actual operation process, the actual predicted value was adjusted. The model can directly output the lateral and longitudinal speed and acceleration of the main vehicle, and then calculate the subsequent trajectory based on the speed and acceleration and the current frame position through the kinematic equation and then calculate the loss function. Through this reflection, both the trajectory and the control amount can be output. Finally, the adaptive momentum gradient algorithm is selected as the optimizer so that the global optimum can be found faster during training without falling into the local optimum.

[0060]

[0061] It can be seen that after determining the optimization scheme, it is necessary to further determine the initial parameters of the hybrid expert model and the hyperparameters of the training. The input of the hybrid expert model is limited by the following initial parameters: the number of categories (class), the time series length (seq), and the driving feature parameter (feature) length. The hyperparameter is the number of samples (batch) for each iteration. The hybrid expert model in the embodiment of the present application outputs a 4-dimensional feature vector of lateral and longitudinal acceleration and speed. The intermediate parameters of the hybrid expert model include the number of layers of transformer and LSTM and the feature lengths of the intermediate layers. The input driving feature parameter length is 42, and the intermediate feature length is set to increase by 2 times layer by layer starting from 64, and then decrease by 2 times back to 64. The remaining adjustable hyperparameters (learning rate lr, learning rate change rate, total number of training rounds) are adjusted accordingly according to the actual training effect.

[0062] Step 22, after initializing the parameters, the hybrid expert model is trained based on the driving data set until the curve change rate of the loss function is less than a first preset threshold in multiple consecutive rounds of training.

[0063] After setting the initialization parameters, you can start model training; batch and lr use the default values ​​of 64 and 0.01, and can be adjusted according to the change of loss (loss function) during training. If the loss function curve changes slowly, you need to increase lr. If the loss does not converge after increasing it, start to increase batch until the loss curve has a high decline rate in the early stage and a slow decline in the later stage.

[0064] The optimal parameter preservation condition of the hybrid expert model is determined by the following two indicators: indicator 1: the loss change rate is less than 0.001 in 10 consecutive rounds after the loss changes smoothly; indicator 2: the average displacement error of the intermediate test (ADE formula 3) (where Y i is the true value of the i-th point, f(x i ) is the predicted value of the i-th point, T is the number of points in a trajectory), the final displacement error (FDE Formula 4) (where T represents the last point of the trajectory, Y T is the truth value of the last point, f(x T) is the predicted value of the last point), mean square error (MSE formula 5), ​​root mean square error (RMSE formula 6) (where N is the number of sample points, Y i is the true value of the i-th point, f(x i ) is the predicted value of the i-th point) exceeds the best result saved previously.

[0065]

[0066] It can be seen that in the embodiment of the present application, the past state characteristics of traffic participants in specific functional scenarios are utilized, the environmental time series feature information is extracted through the backbone network, and the expert network predicts the future time series output to realize the end-to-end decision-making planning process of autonomous driving.

[0067] Furthermore, before model training, a driving data set needs to be constructed. Therefore, the method of constructing the driving data set involved in the above step 201 is as follows: Figure 7 The steps shown may include:

[0068] Step 701, collecting driving trajectory data, and filtering the driving trajectory data according to preset conditions to obtain multiple trajectory data, wherein filtering based on preset conditions means filtering the driving trajectory data according to preset time series length and preset collection frequency to obtain multiple trajectory data; the trajectory data includes the driving trajectory data of the vehicle and the ID of the surrounding vehicle;

[0069] In this regard, in a specific example of an embodiment of the present application, data collection can be performed with a maximum time series length of 8 seconds and a data set acquisition frequency of 25 Hz. The frame number of the trajectory data required for data collection based on this method must be greater than 200 frames, so when extracting single trajectory data, it is necessary to directly eliminate the trajectories with a maximum frame number less than 200.

[0070] Step 702 : extracting driving characteristic parameters from the trajectory data, and constructing a driving data set based on the driving characteristic parameters.

[0071] The method involved in the above step 702 can be specifically implemented through the following steps:

[0072] Step 31, adding the driving data of the surrounding vehicle based on the ID in each frame of data in each track data based on the position information;

[0073] Step 32, extracting sliding window data from each frame of each trajectory data to obtain driving characteristic parameters, wherein the driving characteristic parameters include first attribute data associated with the vehicle and second attribute data associated with the current driving environment;

[0074] Step 33, normalizing the first attribute data and the second attribute data, and storing the normalized data in a target file, wherein each track data corresponds to a target file, and the track data and the sliding window data are mapped by an index;

[0075] Step 34 , constructing a driving data set using the data in the target file corresponding to each track data.

[0076] In the specific example, the relative speed (V) corresponding to the current frame can be filled in according to the surrounding vehicle IDs of the eight directions (front, back, left, right, left front, right front, left rear, right rear) recorded in the trajectory data. x ,V y ), relative acceleration (a x ,a y ), relative position (x, y). Then, sliding window data extraction is performed on a single trajectory, and the interval of sliding window extraction is 1 frame. The lane ID number of the first frame of the extracted 200 frames of sliding window data is the initial lane ID of the trajectory. If the lane IDs of 10 consecutive frames of data at the end of the sliding window are not equal to the initial lane ID, then the type of the sliding window data is considered to be lane change, otherwise it is car following. Figure 8 This is a schematic diagram for extracting driving characteristic parameters. Figure 8 Middle W i represents the sliding window data of the i-th frame, which includes all parameters related to the vehicle (vehicle width, vehicle length, speed, acceleration) and quantities related to the environment (lane ID, relative speed of surrounding vehicles, relative acceleration of surrounding vehicles, relative position of surrounding vehicles, headway, headway, collision time); W1 represents the trajectory data of the following vehicle type, and W2 represents the trajectory data of the lane changing type. All relative objects of the surrounding vehicles are the main vehicle. Finally, calculate all the features ( V x ,V y ,a x ,a y ,X,Y, etc.), the features with large variance are subjected to maximum-minimum normalization (Formula 7) (where x represents the feature vector, min(x) represents the calculation of the minimum value of the feature vector in all samples, and max(x) represents the calculation of the maximum value of the feature vector in all samples), and other features are subjected to mean-variance normalization (Formula 8) (where mean(x) represents the calculation of the average value of all feature vectors, and std(x) represents the calculation of the standard deviation of all feature vectors). The normalization formula is as follows:

[0077]

[0078] The corresponding trajectory file data and sliding window data are saved separately into one file by index mapping, so that the original file does not need to be modified and the memory occupied when reading the data can be reduced.

[0079] It can be seen that in the embodiment of the present application, the driving parameters include but are not limited to the main vehicle speed, the relative distance of the surrounding vehicles, the relative speed, etc.; that is, the driving parameter information of the self-vehicle and the surrounding vehicles in the real driving environment is learned through the neural network to achieve a more humanized driving style. In addition, by using multiple expert models in the hybrid expert model, it can also ensure that the scene expansion will not affect the generalization of the model in multiple scenes, thereby obtaining a more reliable autonomous driving decision-making planning model.

[0080] Based on this, in the embodiment of the present application, for the method of making an automatic driving decision for the vehicle based on the driving trajectory involved in the above step 103, such as Fig. 9 The steps shown may further include:

[0081] Step 901, determining the similarity between the current frame data in the current driving trajectory and the current frame data in the real-time generated driving trajectory;

[0082] Step 902: When the similarity is less than the second preset threshold, continue to make an automatic driving decision for the vehicle based on the current driving trajectory;

[0083] Step 903, when the similarity is greater than or equal to the second preset threshold, the current driving trajectory is replaced by the driving trajectory generated in real time, and the self-driving vehicle makes an automatic driving decision based on the replaced driving trajectory.

[0084] Since the driving trajectory for a period of time in the future is predicted based on the past period of time at the current moment in the embodiment of the present application, the driving trajectory for a period of time in the future is predicted based on the past period of time at each current moment, that is, the driving trajectory is updated in real time. If the driving trajectory predicted at the current moment has a low similarity to the driving trajectory predicted at the previous moment, it means that the current driving trajectory of the vehicle has changed, for example, from following a vehicle to changing lanes, and the current driving trajectory needs to be updated. If the driving trajectory predicted at the current moment has a high similarity to the driving trajectory predicted at the previous moment, it means that the current driving trajectory of the vehicle has not changed and the previous driving trajectory can still be used.

[0085] Therefore, for the above steps 901 to 903 in a specific example, Fig.10As shown in the figure, the control of the motion trajectory is implemented using a double-ended queue. The trajectory stored in the stack is the future position point predicted by the current frame. Whenever a new frame of data arrives, the queue head will pop up the completed frame, and then pop up the data from the queue head and compare it with the new data one by one. If the deviation is less than 0.5, it will be pushed back to the end of the queue. If the deviation is greater than 0.5, all subsequent points will be popped out and the subsequent new trajectory will be directly pushed into the queue. The trajectory frame-by-frame update process is as follows Fig.10 As shown, Q1 represents the control queue, Q2 represents the situation where the current prediction sequence is similar to the remaining unexecuted trajectories, Q3 represents the situation where the current prediction sequence is not similar to the remaining unexecuted trajectories, and Fcur represents the number of frames currently executed.

[0086] The present application is explained below in conjunction with the specific implementation of the embodiment of the present application. The specific implementation provides a method for decision-making and planning for automobile autonomous driving based on a hybrid expert model. In the specific implementation, the overall task is divided into multiple subtasks (highways, on-ramp, off-ramp, intersection, roundabout) through the method of (autonomous driving scene task division), and at the same time, functional experts (following the car, changing lanes) and functional division mode experts (aggressive, conservative) are divided under the task. The idea of ​​the hybrid expert model is used to model the overall task, and the aggregate information is extracted through a unified backbone network. Multiple expert networks jointly predict the domain-specific tasks to solve the generalization problems caused by multi-scenario expansion and scenario expansion. At the same time, the structure of the high-aggregation network also eliminates the problem of decreased accuracy caused by intermediate prediction results. Further, the method steps of this specific implementation are as follows: Fig.11 Shown include:

[0087] Step 1101, driving data set construction:

[0088] The method of constructing the driving dataset specifically includes the following steps:

[0089] Step 41, driving trajectory data preprocessing:

[0090] First, the driving trajectory data is filtered based on the public dataset HighD. The filtering conditions are set according to the maximum time series length and data set acquisition frequency required by the prediction task. The maximum time series length selected in this application is 8 seconds, and the data set acquisition frequency is 25 Hz. The frame number of the required trajectory data must be greater than 200 frames, so when extracting a single trajectory data, it is necessary to directly remove the trajectory with a maximum frame number less than 200. Fig.12 The following is a schematic diagram of data partitioning: Fig.12 Where Ego′ represents the driving parameters of the main vehicle 3 seconds before the current moment, Ego″ represents the driving parameters of the main vehicle 5 seconds after the current moment, t1 represents all driving data of the main vehicle in the first 3 seconds, t2 represents all driving data of the main vehicle in the last 5 seconds, and t3 represents the complete trajectory data used for training.

[0091] The coordinate system of the HighD dataset uses the absolute world coordinate system. In this application, all data is relative to the vehicle coordinate system. Therefore, it is necessary to find the trajectory that travels in the opposite direction relative to the world coordinate system during the traversal of the trajectory data and perform reverse processing on its speed and acceleration. Fig.13 Shown include:

[0092] Step 1301, for the Nth data: determine whether N is greater than the total number of data, if not, then end;

[0093] Step 1302, calculate the total number of frames, and determine whether the total number of frames is greater than 200. If the condition is not met, execute step 1306;

[0094] Step 1303, determine whether the trajectory movement direction is opposite to the world coordinates, if the condition is not met, go to step 1305;

[0095] Step 1304, reverse processing of the speed and acceleration of the host vehicle;

[0096] Step 1305, saving trajectory data;

[0097] Step 1306, execute step 1301 after N+1.

[0098] It should be noted that a complete trajectory data contains not only the vehicle data but also the surrounding vehicle ID of the corresponding frame number, which needs to be further processed for subsequent parameter extraction.

[0099] Step 42, extracting driving characteristic parameters:

[0100] The steps of extracting driving characteristic parameters are as follows: First, fill in the relative speed (V) corresponding to the current frame according to the surrounding vehicle IDs of the eight directions (front, back, left, right, left front, right front, left rear, right rear) recorded in the trajectory data. x ,V y ), relative acceleration (a x ,a y ), relative position (x, y). Then, sliding window data extraction is performed on a single trajectory. The interval of sliding window extraction is 1 frame. The lane ID number of the first frame of the extracted 200 frames of sliding window data is the initial lane ID of the trajectory. If the lane IDs of 10 consecutive frames of data at the end of the sliding window are not equal to the initial lane ID, then the type of the sliding window data is considered to be lane change, otherwise it is car following. Figure 4 This is a schematic diagram for extracting driving characteristic parameters. Figure 4 Middle W irepresents the sliding window data of the i-th frame, which includes all parameters related to the vehicle (vehicle width, vehicle length, speed, acceleration) and quantities related to the environment (lane ID, relative speed of surrounding vehicles, relative acceleration of surrounding vehicles, relative position of surrounding vehicles, headway, headway, collision time); W1 represents the trajectory data of the following type, and W2 represents the trajectory data of the lane changing type. All relative objects of the surrounding vehicles are the main vehicle. Finally, all features ( V x ,V y ,a x ,a y ,X,Y, etc.), perform maximum-minimum normalization on features with large variance, and use mean-variance normalization on other features. Finally, save the corresponding trajectory file data and sliding window data into a separate file using index mapping. This has the advantage of not modifying the original file and reducing the memory occupied when reading data.

[0101] Step 1102: training parameters of the hybrid expert model.

[0102] The method of training the parameters of the hybrid expert model specifically includes the following steps:

[0103] Step 51, hybrid expert model building:

[0104] After the data set is prepared, the corresponding results need to be fitted through the neural network model. This application uses a hybrid expert model to model it. The past 3 seconds of continuous time series data is input into the hybrid expert model to predict the trajectory information for the next 5 seconds. In an embodiment of the present application, the distribution of training data among different expert models is achieved through a gating network, allowing each model to focus on processing the task it is best at, thereby achieving the sparsity of the model; the sparse gating network is part of the hybrid expert model, which receives a single data element as input and then outputs a weight. These weights represent the contribution of each expert model to processing the input data; the expert model will selectively input the corresponding data through the gating network to achieve its corresponding expert task.

[0105] The basic architecture of the hybrid expert model is as follows Figure 3 As shown in the figure, GateNetwork is a gated network. The number of outputs of the gated network corresponds to the number of experts in the expert network. The output of the gated network is G(X). g1 to g4 represent the probability evaluation of the gated network on the corresponding experts. The gated output in round t of training is shown in Formula 3. The input x represents the feature vector of the input model, and the output is the value that the expert model wants to fit or classify. The entire model can be regarded as multiple optimization tasks, namely the classification task of the gated network and the regression or classification tasks of multiple experts.

[0106] This application combines the actual application scenarios of autonomous driving tasks, divides experts according to task types (lane changing, following vehicles), uses LSTM and Transformer, which can handle time series information well, as the backbone network of the overall architecture, and both the expert network and the gated network adopt a simple fully connected layer structure, which can effectively reduce the size of the model and shorten the running time of inference.

[0107] The model architecture in the embodiment of this application is as follows Figure 4 As shown, the gray part is what has been implemented in this application, the white functional blocks are expandable functional blocks, the output of the model is the motion trajectory after 5 seconds, and all the gating networks and backbone networks share the same backbone network.

[0108] The backbone network directly connects the output of the transformer-encoder to the input of the LSTM. The transformer module structure is as follows: Figure 5 As shown in the figure, Embedding represents word encoding before feature input, Positional represents position encoding, Multi-HeadAttention represents multi-head attention module, FeedForward is a feedforward neural network, and N represents multiple layers of this structure; after data is encoded, multi-head attention, and feedforward neural network, the feature relationship between data can be effectively extracted, and intermediate features can be constructed as LSTM input. LSTM can effectively transmit and express information in time series, and this process will not cause long-term useful information to be ignored. At the same time, it can also solve the gradient disappearance and gradient explosion problems in time series neural networks. The LSTM structure is as follows: Figure 6 As shown, X t is the intermediate feature vector of the input, C t-1 The memory unit left over from the previous layer, H t-1 It is the hidden state of the previous layer. The solid-line frame is the forget gate, and the dotted-line frame is the memory gate.

[0109] Step 52, multi-task joint optimization scheme design:

[0110] After completing the modeling of the hybrid expert model, it is necessary to design a multi-task joint optimization solution to solve the training convergence problem. The optimization process of the hybrid expert model is basically the same as that of other neural networks. First, it involves two main tasks, the classification of driving function scenarios (following the car, changing lanes), and the regression of future trajectories. The loss function calculation is as shown in formula (2). The front part is the mean square error loss function of the expert output part, and the back part is the cross entropy loss function of the gated network output part. This application has made adjustments to the actual prediction value during the actual operation process. The model directly outputs the lateral and longitudinal speed and acceleration of the main vehicle, and then calculates the subsequent trajectory based on the speed and acceleration and the current frame position through the kinematic equation and then calculates the loss function. The advantage of this is that it can output both the trajectory and the control quantity. Finally, the adaptive momentum gradient algorithm is selected as the optimizer so that the global optimum can be found faster during training without falling into the local optimum.

[0111] Step 53, initialization of model training parameters:

[0112] After determining the optimization plan, it is necessary to further determine the initial parameters of the hybrid expert model and the training hyperparameters; the input of the hybrid expert model is limited by the number of categories (class), time series length (seq), and driving feature parameter (feature) length, and the only adjustable hyperparameter is the number of samples in each iteration (batch); the model output is a 4-dimensional feature vector of lateral and longitudinal acceleration and speed; the intermediate parameters of the model include the number of layers of transformer and LSTM and the feature length of each intermediate layer. The input driving feature parameter length is 42, and the intermediate feature length is set to increase by 2 times from 64 layer by layer, and then decrease by 2 times back to 64. The remaining adjustable hyperparameters (learning rate lr, learning rate change rate, total number of training rounds) are adjusted accordingly according to the actual training results.

[0113] Step 54, model training optimization:

[0114] After specifying the initialization parameters, you can start training the hybrid expert model; batch and lr use the default values ​​of 64 and 0.01, and can be adjusted according to the change of loss during training. If the loss function curve changes slowly, you need to increase lr. If the loss does not converge after increasing it, start to increase batch until the loss curve has a high decline rate in the early stage and a slow decline in the later stage.

[0115] The optimal parameter preservation conditions of the hybrid expert model mainly depend on two indicators: indicator 1: the loss changes smoothly and the change rate is less than 0.001 for 10 consecutive rounds; indicator 2: the average displacement error (ADE formula 3), final displacement error (FDE formula 4), mean square error (MSE formula 5), ​​and root mean square error (RMSE formula 6) of the intermediate test exceed the previously saved optimal result.

[0116] Step 1103, decision-making planning driving trajectory generation.

[0117] The decision-making planning driving trajectory generation method specifically includes the following steps:

[0118] Step 61, building a driving simulation environment: After training the hybrid expert model, you need to build a simulation test environment to verify the trajectory generation effect. In VTD (Virtual Test Drive), build a corresponding simulation environment according to the road scene of the HighD data set, and use the traffic flow generated by the VTD driver model as the background test vehicle. Read the simulation environment data through TCP communication and input it into the hybrid expert model, and then control the main vehicle movement in real time through the output of the hybrid expert model.

[0119] Step 62, generate predicted trajectory: After completing the environment construction and parameter setting of the simulation environment, directly perform simulation. When there are surrounding vehicles in the host vehicle environment, the VTD end will send the environment and vehicle data to the model. After the model end completes the data analysis, it will build the model input and finally output the lateral and longitudinal speeds and accelerations of continuous frames. The predicted amount is returned to the VTD end to calculate the generated trajectory through the kinematic equation.

[0120] Step 63, real-time control of the main vehicle movement: The control quantities received from the hybrid expert model in the simulation include speed, acceleration, and motion trajectory. After each frame of control data arrives, the acceleration and speed of the main vehicle will be directly changed; the control of the motion trajectory is achieved using a double-ended queue, such as Fig.10 As shown in the figure, the trajectory saved in the stack is the future position point predicted by the current frame. Whenever a new frame of data arrives, the queue head will pop out the completed frame, and then pop out the data from the queue head and compare it with the new data one by one. If the deviation is less than 0.5, it will be pushed back to the end of the queue. If the deviation is greater than 0.5, all subsequent points will be popped out and the subsequent new trajectory will be directly pushed into the queue.

[0121] Corresponding to the above-mentioned method embodiments of the present application, the following is an apparatus embodiment of the present application, which can be used to execute the method embodiments of the present application.

[0122] Fig.14 This is a schematic diagram of the structure of a device for generating autonomous driving decisions based on a hybrid expert model provided in an embodiment of the present application. The device can be implemented as part or all of an electronic device through software, hardware, or a combination of both. Fig.14 As shown, the device comprises:

[0123] The sending module 1402 is used to send the driving data of the ego vehicle in the preset simulation environment to the hybrid expert model, wherein the driving data includes the driving data of the ego vehicle and the driving data of surrounding vehicles associated with the ego vehicle, and the hybrid expert model is obtained based on multiple expert models, and the multiple expert models are used to jointly predict multiple domain-specific autonomous driving tasks; after training, the hybrid expert model is used to take the continuous time series driving data of the vehicle within a first preset time length before the current moment as input, and output the continuous time series driving data within a second preset time length after the current moment;

[0124] The first processing module 1404 is used to obtain an output result of the hybrid expert model based on the driving data as an input of the hybrid expert model, wherein the output result represents the driving data of the vehicle in a continuous time series within a preset time period after the current moment;

[0125] The second processing module 1406 is used to generate a corresponding driving trajectory based on the driving data, and make an automatic driving decision for the vehicle based on the driving trajectory.

[0126] In an optional implementation manner of the embodiment of the present application, the device in the embodiment of the present application may also include: a construction module, used to construct a driving data set; wherein the driving data set includes multiple trajectory data, each trajectory data includes current vehicle data and surrounding vehicle data associated with the vehicle; a third processing module, used to model the hybrid expert model, wherein the modeled hybrid expert model includes a gated network, a backbone network and multiple expert models; a training module, used to train the hybrid expert model based on the driving data set, so that the hybrid expert model is based on the continuous time series driving data of the vehicle within a first preset time length before the current moment as input, and outputs the continuous time series driving data within a second preset time length after the current moment.

[0127] In an optional implementation manner of the embodiment of the present application, the construction module in the embodiment of the present application may further include: a first processing unit, used to collect driving trajectory data, and filter the driving trajectory data according to preset conditions to obtain multiple trajectory data, wherein filtering based on preset conditions means filtering the driving trajectory data according to a preset time series length and a preset acquisition frequency to obtain multiple trajectory data; the trajectory data includes the driving trajectory data of the vehicle and the ID of the surrounding vehicle; a second processing unit, used to extract driving feature parameters from the trajectory data, and construct a driving data set based on the driving feature parameters.

[0128] In an optional implementation manner of an embodiment of the present application, the second processing unit in the embodiment of the present application may further include: a first processing subunit, used to add the driving data of the surrounding vehicle based on the ID in each frame of data in each trajectory data based on the orientation information; an extraction subunit, used to extract the sliding window data for each frame of data in each trajectory data to obtain driving characteristic parameters, wherein the driving characteristic parameters include first attribute data associated with the vehicle and second attribute data related to the current driving environment; a second processing subunit, used to normalize the first attribute data and the second attribute data, and store the data obtained after the normalization processing in a target file, wherein each trajectory data corresponds to a target file, and the trajectory data and the sliding window data are mapped by an index; a construction subunit, used to construct a driving data set through the data in the target file corresponding to each trajectory data.

[0129] In an optional implementation manner of the embodiment of the present application, the third processing module in the embodiment of the present application may further include: a setting unit, used to set a long short-term memory network LSTM and a Transformer model in the backbone network, wherein the Transformer model and the LSTM adopt a fully connected layer structure, and the output of the Transformer model is the input of the LSTM; a third processing unit, used to determine the corresponding expert model according to the task in the autonomous driving scenario, and set the output number of the gated network to be consistent with the number of expert models, wherein the output of the LSTM is the input of the gated network, and the output of the gated network represents the evaluation probability of the corresponding expert model.

[0130] In an optional implementation manner of an embodiment of the present application, the training module in the embodiment of the present application may further include: a first determination unit, used to determine the loss function according to the task in the autonomous driving scenario, and the parameters in the hybrid expert model after initialization of the modeling; a training unit, used to train the hybrid expert model based on the driving data set after initializing the parameters, until the curve change rate of the loss function is less than a first preset threshold in multiple consecutive rounds of training.

[0131] In an optional implementation manner of the embodiment of the present application, the second processing module in the embodiment of the present application may further include: a second determination unit, used to determine the similarity between the current frame data in the current driving trajectory and the current frame data in the driving trajectory generated in real time; a fourth processing unit, used to continue to make automatic driving decisions for the vehicle based on the current driving trajectory when the similarity is less than a second preset threshold; a fifth processing unit, used to replace the current driving trajectory with the driving trajectory generated in real time when the similarity is greater than or equal to the second preset threshold, and make automatic driving decisions for the vehicle based on the replaced driving trajectory.

[0132] like Fig.15As shown, an embodiment of the present application provides an electronic device, including a processor 1501, a communication interface 1502, a memory 1503 and a communication bus 1504, wherein the processor 1501, the communication interface 1502, and the memory 1503 communicate with each other through the communication bus 1504.

[0133] Memory 1503, used for storing computer programs;

[0134] In one embodiment of the present application, the processor 1501 is used to execute the program stored in the memory 1503 to implement the steps of generating an autonomous driving decision-making method based on a hybrid expert model provided by any of the aforementioned method embodiments. The role it plays is similar and will not be repeated here.

[0135] An embodiment of the present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of generating an autonomous driving decision-making method based on a hybrid expert model as provided in any of the aforementioned method embodiments are implemented.

[0136] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0137] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a general hardware platform, and of course, by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the relevant technology can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0138] It should be understood that the terms used herein are only for the purpose of describing specific example embodiments and are not intended to be limiting. Unless the context clearly indicates otherwise, the singular forms "one", "an" and "said" as used herein may also be meant to include plural forms. The terms "include", "comprise", "contain", and "have" are inclusive, and therefore specify the existence of stated features, steps, operations, elements and / or parts, but do not exclude the existence or addition of one or more other features, steps, operations, elements, parts, and / or combinations thereof. The method steps, processes, and operations described herein are not interpreted as necessarily requiring them to be performed in the specific order described or illustrated, unless the execution order is clearly indicated. It should also be understood that additional or alternative steps may be used.

[0139] The above description is only a specific implementation of the present application, so that those skilled in the art can understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest range consistent with the principles and novel features applied for herein.

Claims

1. A method for generating autonomous driving decisions based on a hybrid expert model, characterized in that: include: Sending the driving data of the self-vehicle in a preset simulation environment to the hybrid expert model, wherein the driving data includes the driving data of the self-vehicle and the driving data of surrounding vehicles associated with the self-vehicle, and the hybrid expert model is modeled based on multiple expert models, and the multiple expert models are used to jointly predict multiple domain-specific autonomous driving tasks; after training, the hybrid expert model is used to take the continuous time series driving data of the vehicle within a first preset time length before the current moment as input, and output the continuous time series driving data within a second preset time length after the current moment; Based on the driving data as the input of the hybrid expert model, an output result of the hybrid expert model is obtained, wherein the output result represents the driving data of the vehicle in a continuous time series within a preset time period after the current moment; A corresponding driving trajectory is generated based on the driving data, and an automatic driving decision is made for the vehicle based on the driving trajectory.

2. The method according to claim 1, characterized in that: Training the hybrid expert model includes: Constructing a driving data set; wherein the driving data set includes a plurality of trajectory data, each of which includes current vehicle data and surrounding vehicle data associated with the vehicle; Modeling the hybrid expert model, wherein the hybrid expert model after modeling includes a gating network, a backbone network and multiple expert models; The hybrid expert model is trained based on the driving data set so that the hybrid expert model takes the continuous time series driving data of the vehicle within a first preset time period before the current moment as input and outputs the continuous time series driving data within a second preset time period after the current moment.

3. The method according to claim 2, characterized in that The constructing of the driving data set includes: Collecting driving trajectory data, and filtering the driving trajectory data according to preset conditions to obtain multiple trajectory data, wherein filtering based on the preset conditions means filtering the driving trajectory data according to a preset time sequence length and a preset collection frequency to obtain multiple trajectory data; the trajectory data includes the driving trajectory data of the vehicle and the ID of the surrounding vehicle; Driving characteristic parameters are extracted from the trajectory data, and the driving data set is constructed based on the driving characteristic parameters.

4. The method according to claim 3, characterized in that Extracting driving characteristic parameters from the trajectory data and constructing the driving data set based on the driving characteristic parameters includes: Adding the driving data of the surrounding vehicle based on the ID in each frame of data in each piece of the trajectory data based on the position information; Claims Extracting sliding window data from each frame of the trajectory data to obtain the driving characteristic parameters, wherein the driving characteristic parameters include first attribute data associated with the vehicle and second attribute data associated with the current driving environment; Normalizing the first attribute data and the second attribute data, and storing the data obtained after the normalization process in a target file, wherein each of the trajectory data corresponds to a target file, and the trajectory data and the sliding window data are mapped by an index; The driving data set is constructed by using the data in the target file corresponding to each piece of the trajectory data.

5. The method according to claim 2, characterized in that: Modeling the hybrid expert model includes: A long short-term memory network LSTM and a Transformer model are set in the backbone network, wherein the Transformer model and the LSTM adopt a fully connected layer structure, and the output of the Transformer model is the input of the LSTM; The corresponding expert model is determined according to the task in the autonomous driving scenario, and the output number of the gating network is set to be consistent with the number of the expert models, wherein the output of the LSTM is the input of the gating network, and the output of the gating network represents the evaluation probability of the corresponding expert model.

6. The method according to claim 5, characterized in that Training the hybrid expert model based on the driving data set includes: Determining a loss function according to the tasks in the autonomous driving scenario and initializing parameters in the modeled hybrid expert model; After initializing the parameters, the hybrid expert model is trained based on the driving data set until the curve change rate of the loss function is less than a first preset threshold in multiple consecutive rounds of training.

7. The method according to claim 1, characterized in that Making an automatic driving decision for the vehicle based on the driving trajectory includes: Determine the similarity between the current frame data in the current driving trajectory and the current frame data in the real-time generated driving trajectory; When the similarity is less than a second preset threshold, continue to make an automatic driving decision for the vehicle based on the current driving trajectory; When the similarity is greater than or equal to the second preset threshold, the real-time generated driving trajectory replaces the current driving trajectory, and the vehicle makes an automatic driving decision based on the replaced driving trajectory.

8. A device for generating autonomous driving decisions based on a hybrid expert model, characterized in that: include: Claims A sending module, used to send the driving data of the ego vehicle in a preset simulation environment to the hybrid expert model, wherein the driving data includes the driving data of the ego vehicle and the driving data of surrounding vehicles associated with the ego vehicle, and the hybrid expert model is modeled based on multiple expert models, and the multiple expert models are used to jointly predict multiple domain-specific autonomous driving tasks; after training, the hybrid expert model is used to take the continuous time series driving data of the vehicle within a first preset time length before the current moment as input, and output the continuous time series driving data within a second preset time length after the current moment; A first processing module, configured to obtain an output result of the hybrid expert model based on the driving data as an input of the hybrid expert model, wherein the output result represents the driving data of the vehicle in a continuous time series within a preset time period after a current moment; The second processing module is used to generate a corresponding driving trajectory based on the driving data, and make an automatic driving decision for the vehicle based on the driving trajectory.

9. The device according to claim 8, characterized in that The device also includes: A construction module, used to construct a driving data set; wherein the driving data set includes a plurality of trajectory data, each of which includes current vehicle data and surrounding vehicle data associated with the vehicle; A third processing module is used to model the hybrid expert model, wherein the hybrid expert model after modeling includes a gating network, a backbone network and multiple expert models; A training module is used to train the hybrid expert model based on the driving data set so that the hybrid expert model takes the continuous time-series driving data of the vehicle within a first preset time period before the current moment as input and outputs the continuous time-series driving data within a second preset time period after the current moment.

10. The device according to claim 9, characterized in that The building blocks include: The first processing unit is used to collect driving trajectory data, and filter the driving trajectory data according to preset conditions to obtain multiple trajectory data, wherein filtering based on the preset conditions means filtering the driving trajectory data according to a preset time sequence length and a preset collection frequency to obtain multiple trajectory data; the trajectory data includes the driving trajectory data of the vehicle and the ID of the surrounding vehicle; The second processing unit is used to extract driving characteristic parameters from the trajectory data and construct the driving data set based on the driving characteristic parameters.

11. An electronic device, comprising: at least one communication interface; at least one bus connected to the at least one communication interface; at least one processor coupled to the at least one bus; At least one memory connected to the at least one bus, wherein the processor is configured to execute the method for generating autonomous driving decisions based on a hybrid expert model as described in any one of claims 1 to 7.

12. A computer storage medium storing computer executable instructions, wherein the computer executable instructions are used to execute the method for generating an autonomous driving decision based on a hybrid expert model as described in any one of claims 1 to 7.

Citation Information

Cited By

  • Intelligent load balancing method based on zero sample hybrid expert model

    CN119225983A

  • An intelligent load balancing method based on zero-sample hybrid expert model

    CN119225983B

  • Hybrid expert strategy fusion automatic driving control method based on world model

    CN121553176A

  • World model based hybrid expert policy fusion autonomous driving control method

    CN121553176B