An Uncertainty Estimation Method for an End-to-End Autonomous Driving Decision Algorithm

Abstract

The present invention discloses a method for estimating the uncertainty of an end-to-end autonomous driving decision algorithm, belonging to the field of autonomous driving. An end-to-end network model constructed by n multi-view-based end-to-end decision algorithms is used as the basic model; a public data set is used as the training data set and randomly divided into a total training set and a validation set according to a ratio; the total training set data is evenly divided into n sub-data sets, and each sub-data set in the n sub-data sets is responsible for training and validating an end-to-end network model; since the parameters of each end-to-end network model are randomly initialized before training, N end-to-end decision networks with the same structure but different parameters can be obtained as the final decision network model. Then, the model is used to perform inference on the same scenario, and the results separately inferred by each network are calculated using set operations, so as to calculate the uncertainty of the network for corner and speed estimation, improving the accuracy of uncertainty estimation.

Description

Technical Field

[0001] The present invention belongs to the field of autonomous driving, and particularly relates to a method for estimating the uncertainty of an end-to-end autonomous driving decision algorithm. Background Art

[0002] The number of motor vehicles in China is increasing continuously. Coupled with the uneven driving skills of drivers, the probability of highway traffic congestion and traffic accidents is getting higher and higher. Autonomous driving relies on computer technology and artificial intelligence technology to achieve driving, which can improve travel efficiency and greatly avoid traffic congestion and traffic accidents caused by driver misoperations.

[0003] The decision-making and planning module is one of the key modules in the autonomous driving technology system. After receiving the perception information, the decision-making module analyzes the current environment, calculates a reasonable control quantity, and conveys the control quantity to the lower-level execution module. The end-to-end decision-making method is an implementation method of the decision-making model, which mainly designs a model by relying on methods such as deep learning or deep reinforcement learning, and uses a large amount of human driving data for training, so that the model learns the relationship from the input image to the control quantities such as the steering wheel angle and speed of the output.

[0004] Currently, people attach great importance to the safety of autonomous driving. However, the "end-to-end decision-making algorithm" is a typical black-box structure, and researchers cannot know whether the decision-making quantity output by the decision-making algorithm is credible, that is, there is a lack of a method for judging the credibility of the output quantity of the decision-making algorithm, which makes the safety concern problem of autonomous driving vehicles using this type of decision-making algorithm prominent. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for estimating the uncertainty of an end-to-end autonomous driving decision algorithm, so as to solve the problems that the existing end-to-end autonomous driving decision algorithm lacks an evaluation of the credibility of the output quantity of the decision-making algorithm, resulting in the prominent safety concern problem of autonomous driving vehicles using this type of decision-making algorithm.

[0006] To achieve the above purpose, the present invention adopts the following technical solutions:

[0007] A method for estimating the uncertainty of an end-to-end autonomous driving decision algorithm includes the following steps:

[0008] S1. Based on the end-to-end decision-making algorithm with a multi-perspective attention mechanism, establish n end-to-end decision-making models to be measured for which the uncertainty needs to be estimated;

[0009] S2. Use the public data set as the training sample data to train the decision-making models established in S1, and obtain n end-to-end decision-making network models with the same structure but different parameters. The detailed process is as follows:

[0010] S2.1. Randomly select a portion of the sample data from the training sample data as the total training set and the validation set respectively in proportion.

[0011] S2.2. Randomly and evenly divide the sample data of the total training set into n sub-datasets, and randomly divide each sub-dataset into a sub-training set and a sub-validation set in proportion.

[0012] S2.3. Use the sub-datasets obtained in S2.2 to train and validate the n models established in S1. Each model selects a sub-dataset as the input for independent training to obtain n end-to-end decision network models with different parameters but the same structure.

[0013] S3. Use the validation set obtained in S2.1 as the input data, and input it into the n end-to-end decision network models with different parameters but the same structure trained in S2.3 respectively to calculate n sets of corner values and speed values.

[0014] S4. Calculate the average value and variance of the n sets of corners and the average value and variance of the n sets of speeds obtained in S3 respectively; use the average value of the n sets of corners as the final control corner output, and the variance of the n sets of corners as the uncertainty of the corner; use the average value of the n sets of speeds as the final control speed output, and the variance of the n sets of speeds as the uncertainty of the speed.

[0015] Further, the detailed process of step S2.3 is as follows:

[0016] S2.3.1. Randomly initialize the parameters of the n end-to-end decision network models.

[0017] S2.3.2. Each end-to-end decision network model is trained according to the training set in the sub-dataset it inputs; each end-to-end decision network model is verified according to the sub-validation set in the sub-dataset it inputs until the training loss no longer decreases.

[0018] Further, in the above uncertainty estimation method of the end-to-end autonomous driving decision algorithm, due to the different sample set data collected and the different decision networks selected, the magnitudes and ranges of the corner uncertainty and speed uncertainty are not the same; by setting thresholds corresponding to different training sets and decision algorithms, the uncertainty is divided into four levels: low, medium, high, and extremely high. When the uncertainty value is greater than medium, it indicates that the probability of autonomous driving failure in the current environment is relatively high.

[0019] Further, the value range of n is 5 ≤ n ≤ 8.

[0020] The present invention uses n end-to-end network models constructed based on multi-perspective end-to-end decision algorithms as the basic model; uses a public data set as a training data set, and randomly divides it into a total training set and a verification set according to a ratio; divides the total training set data into n sub-data sets, and each of the n sub-data sets is responsible for training and verifying an end-to-end network model; because the parameters of each end-to-end network model are randomly initialized before training, N end-to-end decision networks with different parameters and the same structure can be obtained as the final decision network model. Using this model, the same scene is inferred through each network separately during reasoning, and then the set calculation method is used to calculate the uncertainty of the network for the angle and speed estimation, thereby improving the accuracy of uncertainty estimation. In addition, the added uncertainty division determines the probability of autonomous driving failure, which is equivalent to adding the credibility judgment of the output of the decision algorithm in the process of calculating uncertainty, which effectively improves the safety of autonomous driving.

[0021] Compared with the prior art, the method of the present invention uses existing public data as sample data, has a large amount of data information, and does not require the collection of implementation data. It is more universal and can estimate network uncertainty without modifying the decision network structure, thereby improving the safety of the entire vehicle using such a network. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is the principle diagram of the method of the present invention;

[0023] Figure 2 This is a flow chart of training N isomorphic and heterogeneous end-to-end decision network models in the embodiment;

[0024] Figure 3 The embodiment uses N trained homogeneous and heterogeneous end-to-end decision network models to calculate uncertainty flow chart. Specific embodiments

[0025] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0026] like Figure 1 As shown, this embodiment provides an uncertainty estimation method for an end-to-end autonomous driving decision algorithm, comprising the following steps:

[0027] S1. An end-to-end decision algorithm based on a multi-view attention mechanism is used to establish five end-to-end decision models that require uncertainty estimation.

[0028] S2, using the Drive 360 dataset as training sample data, trains the model established in S1 to obtain five end-to-end decision network models with different parameters and the same structure. Figure 2 As shown:

[0029] S2.1. Randomly select 90% of the sample data from the training sample data as the total training set, and the remaining 10% of the sample data as the validation set.

[0030] S2.2. Randomly and evenly divide the sample data of the total training set into 5 sub-datasets, and then randomly divide each sub-dataset into a sub-training set and a sub-validation set according to a ratio; among them, the sub-training set accounts for 80% of each sub-dataset, and the sub-validation set accounts for 20% of each sub-dataset.

[0031] S2.3. Use the sub-datasets obtained in S2.2 to train and validate the 5 models established in S1. Each model selects a sub-dataset as the input for independent training to obtain 5 end-to-end decision network models with different parameters but the same structure. Specifically:

[0032] S2.3.1. Randomly initialize the parameters of the 5 end-to-end decision network models;

[0033] S2.3.2. Each end-to-end decision network model is trained according to the training set in the sub-dataset it inputs; each end-to-end decision network model is validated according to the sub-validation set in the sub-dataset it inputs until the training loss no longer decreases.

[0034] S3. Calculate the corner uncertainty and speed uncertainty during the autonomous driving process, as Figure 3 shown:

[0035] S3.1. Use the validation set obtained in S2.1 as the input data and input it into the 5 end-to-end decision network models with different parameters but the same structure trained in S2.3 respectively to calculate 5 sets of corners and speeds;

[0036] S3.2. Calculate the average value and variance of the 5 sets of corners and the average value and variance of the 5 sets of speeds obtained in S4 respectively; and use the average value of the 5 sets of corners as the final control corner output, the variance of the 5 sets of corners as the corner uncertainty; the average value of the 5 sets of speeds as the final control speed output, and the variance of the 5 sets of speeds as the speed uncertainty.

[0037] The uncertainty estimation method of the end-to-end autonomous driving decision algorithm provided in this embodiment reasons about the same scenario by establishing 5 end-to-end decision network models with different parameters but the same structure, and then uses set calculation for the results respectively inferred by each network to calculate the uncertainty of the network for corner and speed estimation. During the calculation process, the Drive360 dataset is used as the training sample data, and there is no need to collect the data of autonomous driving in real time, ensuring the comprehensiveness of the sample data.

[0038] During use, due to the differences in the collected sample set data and the selected decision network, the magnitudes and ranges of the angular uncertainty and speed uncertainty are not the same. By setting thresholds corresponding to different training sets and decision algorithms, the uncertainty is divided into four levels: low, medium, high, and extremely high. When the uncertainty value is greater than medium, it indicates that the probability of an autonomous driving failure in the current environment is relatively high, and further processing is required. In this embodiment, the uncertainty threshold for speed is as follows: low is <0.05, medium is 0.05 - 0.1, high is 0.1 - 0.2, and extremely high is >0.2; the uncertainty threshold for the angle is as follows: low is <0.1, medium is 0.1 - 0.4, high is 0.4 - 1.0, and extremely high is >1.0.

[0039] In summary, the uncertainty estimation method for the end-to-end autonomous driving decision algorithm provided in this embodiment effectively solves the problem that the existing end-to-end autonomous driving decision algorithms lack the evaluation of the credibility of the output of the decision algorithm, resulting in insufficient safety performance of autonomous driving vehicles using such decision algorithms. Since the obtained sample data set already contains various azimuth data during the driving process, it can be directly used for real-time estimation in the in-vehicle real vehicle scenario without the need for true values.

[0040] The above is only used to illustrate the technical solution of the present invention and not to limit it. Any other modifications or equivalent replacements made by those of ordinary skill in the art to the technical solution of the present invention shall be covered by the scope of the claims of the present invention as long as they do not depart from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for estimating the uncertainty of an end-to-end autonomous driving decision-making algorithm, characterized in that: It includes the following steps: S1. Based on the end-to-end decision algorithm with a multi-perspective attention mechanism, establish n end-to-end decision models to be measured that need to estimate uncertainty; S2. Use the public dataset as the training sample data to train the decision models established in S1, and obtain n end-to-end decision network models with the same structure but different parameters. The detailed process is as follows: S2.

1. Randomly extract a part of the sample data from the training sample data as the total training set and the validation set respectively; S2.

2. Randomly divide the sample data of the total training set into n sub-datasets on average, and randomly divide each sub-dataset into a sub-training set and a sub-validation set according to a certain proportion; S2.

3. Use the sub-datasets obtained in S2.2 to train and validate the n models established in S1. Each model selects a sub-dataset as the input for independent training to obtain n end-to-end decision network models with the same structure but different parameters. Specifically: S2.3.

1. Randomly initialize the parameters of the n end-to-end decision network models; S2.3.

2. Each end-to-end decision network model is trained according to the training set in the sub-dataset it inputs; each end-to-end decision network model is validated according to the sub-validation set in the sub-dataset it inputs until the training loss no longer decreases; S3. Take the validation set obtained in S2.1 as the input data, and input it into the n end-to-end decision network models with the same structure but different parameters trained in S2.3 respectively, and calculate to obtain n groups of corner values and speed values; S4. Calculate the average and variance of the n groups of corners and the average and variance of the n groups of speeds obtained in S4 respectively; take the average of the n groups of corner values as the final control corner output, and the variance of the n groups of corner values as the uncertainty of the corner; take the average of the n groups of speed values as the final control speed output, and the variance of the n groups of speed values as the uncertainty of the speed.

2. The uncertainty estimation method of an end-to-end autonomous driving decision-making algorithm according to claim 1, characterized in that: The magnitudes of the corner uncertainty and speed uncertainty are determined by the sample set data collected and the N decision network models with the same structure but different parameters constructed. By setting the thresholds corresponding to different training sets and N decision network models, the uncertainty is divided into four levels: low, medium, high, and ultra-high. When the uncertainty value is greater than medium, it indicates that the probability of an autonomous driving failure in the current environment is relatively high.

3. A method for estimating the uncertainty of an end-to-end autonomous driving decision-making algorithm according to any one of claims 1 or 2, characterized in that: The value range of n is 5 ≤ n ≤ 8.

Images