Autonomous driving test scene extraction method based on one-dimensional residual convolutional autoencoder

Abstract

A method for extracting autonomous driving test scenes based on a one-dimensional residual convolutional autoencoder comprises the following steps: 1) preprocessing a natural driving data set; 2) selecting a number of different scene data to form a high-dimensional scene feature data set; 3) using a one-dimensional residual convolutional autoencoder to perform feature dimensionality reduction on each scene data of the high-dimensional scene feature data set to form a low-dimensional feature type data set; 4) considering the weight difference between different features, respectively calculating the weights corresponding to various feature vectors of the low-dimensional feature type data set, and weighting the feature data corresponding to various feature vectors of the low-dimensional feature type data set to obtain the weighted low-dimensional feature type data set; 5) using a K-Means clustering model to cluster the weighted low-dimensional feature type data set to obtain the scene data classification of the low-dimensional feature type data set; 6) naming each scene category according to a number of scene data corresponding to each scene category in the high-dimensional scene feature data set.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving, and in particular to a method for extracting autonomous driving test scenarios based on a one-dimensional residual convolutional autoencoder. Background Art

[0002] Natural driving datasets refer to a data set formed by scene information data of several autonomous vehicles in real driving environments. At present, in the field of autonomous vehicles, natural driving datasets are usually used to simulate and verify autonomous vehicles based on various test scenarios, in order to reduce the testing costs of autonomous vehicles while still ensuring the safety and reliability of autonomous vehicles in real driving environments.

[0003] By analyzing the natural driving data collected by autonomous vehicles during driving, developers can better understand the driving patterns of autonomous vehicles in real driving environments, make full use of the effective information therein, extract the accurate features of various real driving scenarios, and use them for the development and verification of the control algorithms of the autonomous driving system, and study and evaluate the performance and limitations of the autonomous driving system to guide the optimization and improvement of the entire system.

[0004] At present, in the field of test scene extraction for autonomous vehicles, attempts have been made to use machine learning algorithms to automatically learn and discover patterns and features in data and extract real driving scenes from them. For example, the "Classical and Extreme Scene Division and Extraction Method Based on Connected Vehicle Driving Data" with publication number CN112036297A uses the isolation forest algorithm to segment natural driving data, then extracts scene features through an autoencoder network model, and finally obtains different types of scenes through the K-Means algorithm.

[0005] However, although this test scenario extraction method based on machine learning has better flexibility and adaptability, the autoencoder network in this method uses a fully connected layer with a large number of parameters. It is very easy to overfit when the data input dimension is high. In addition, in the process of clustering analysis, the weight differences between different features are ignored, which can easily amplify the influence of minor features and downplay the influence of important features, making the final result inconsistent with reality. Summary of the invention

[0006] The purpose of the present invention is to provide an autonomous driving test scene extraction method based on a one-dimensional residual convolutional autoencoder to address the corresponding deficiencies of the prior art. It not only uses multiple convolutional layers and pooling layers to form a new autoencoder, but also uses parameter-sharing convolution kernels to greatly reduce the number of autoencoder parameters and speed up the training of the autoencoder. At the same time, by introducing the concept of residual modules, problems such as gradient vanishing and gradient explosion are avoided, making deep neural networks easier to train. In addition, the weight differences between different features are taken into account, so that the final classification results are in line with reality.

[0007] The purpose of the present invention is to adopt the following scheme to achieve: a method for extracting autonomous driving test scenarios based on a one-dimensional residual convolutional autoencoder, comprising the following steps:

[0008] 1) Obtain a natural driving data set and preprocess the natural driving data set;

[0009] 2) Select a number of different scene data from the natural driving dataset to form a high-dimensional scene feature dataset;

[0010] 3) Using a one-dimensional residual convolutional autoencoder, the feature dimension reduction of each scene data of the high-dimensional scene feature data set is performed to form a low-dimensional feature type data set;

[0011] 4) Calculate the weights corresponding to each type of feature vector of the low-dimensional feature type data set respectively, and weight the feature data corresponding to each type of feature vector of the low-dimensional feature type data set to obtain a weighted low-dimensional feature type data set;

[0012] 5) Use the K-Means clustering model to cluster the weighted low-dimensional feature type data set to obtain the scene data classification of the low-dimensional feature type data set and number it;

[0013] 6) Setting the naming conditions for each scene category, and naming each scene category according to a number of scene data corresponding to each scene category in the high-dimensional scene feature data set.

[0014] Preferably, the specific calculation process of the weight corresponding to any type of feature vector in the low-dimensional feature type data set is as follows:

[0015] 1) The feature data x corresponding to the feature vector j of the i-th scene data in the low-dimensional feature type dataset i-j Standardize the process;

[0016] 2) Calculate the proportion of the i-th scene data in the feature vector j:

[0017]

[0018] In the formula, p i-jis the proportion of the i-th scene data in the feature vector j, x′ i-j is the feature data corresponding to the feature vector j of the i-th scene data after normalization, and n is the number of scene data in the low-dimensional feature type data set;

[0019] 3) Calculate the information entropy value E of feature vector j in the low-dimensional feature type data set j :

[0020]

[0021] In the formula, E j is the information entropy value of feature vector j in the low-dimensional feature type dataset, n is the number of scene data in the low-dimensional feature type dataset, and p i-j is the proportion of the i-th scene data in the feature vector j;

[0022] 4) Calculate the weight w of feature vector j according to the following formula j :

[0023]

[0024] In the formula, w j is the weight of the feature vector j of the low-dimensional feature type dataset, s is the number of feature vector types in the low-dimensional feature type dataset, E j is the information entropy value of feature vector j in the low-dimensional feature type data set.

[0025] Preferably, the method of preprocessing the natural driving data set includes data cleaning and / or data filtering.

[0026] Preferably, a time length threshold is set, and the time length of each segment of scene data selected from the natural driving data set is within the time length threshold range.

[0027] Preferably, the one-dimensional residual convolutional autoencoder includes an encoder and a decoder, the encoder includes a first convolutional layer, a second convolutional layer, a third convolutional layer, and a first bottleneck layer, a first maximum pooling layer is provided between the first convolutional layer and the second convolutional layer, and a second maximum pooling layer is provided between the third convolutional layer and the first bottleneck layer;

[0028] The decoder includes a first deconvolution layer, a second deconvolution layer, a third deconvolution layer, and a second bottleneck layer, a first upsampling layer is provided between the second bottleneck layer and the first deconvolution layer, a second jump connection point is provided between the first upsampling layer and the first deconvolution layer, a second upsampling layer is provided between the second deconvolution layer and the third deconvolution layer, and a first jump connection point is provided between the second upsampling layer and the second deconvolution layer;

[0029] The input features of the encoder enter the encoder through the first convolution layer, then pass through the first maximum pooling layer, the second convolution layer, the third convolution layer, the second maximum pooling layer in sequence, and finally are output from the first bottleneck layer to the second bottleneck layer of the decoder, and then pass through the first upsampling layer, the first deconvolution layer, the second deconvolution layer, the second upsampling layer in sequence, and finally are output from the third deconvolution layer;

[0030] The output features of the first maximum pooling layer are added to the output features of the second deconvolution layer through the first skip connection point as input features of the second upsampling layer;

[0031] The output features of the third convolutional layer are added to the output features of the first upsampling layer through the second jump connection point as input features of the first deconvolution layer.

[0032] Preferably, the output features of the first convolutional layer, the second convolutional layer, the third convolutional layer, the first deconvolutional layer, the second deconvolutional layer, and the third deconvolutional layer are all processed in a batch normalization manner and then input into the next layer.

[0033] The advantages of the present invention are the following two points:

[0034] ① The present invention adopts a one-dimensional residual convolutional autoencoder which is different from the prior art. Experimental data show that the one-dimensional residual convolutional autoencoder adopted by the present invention can significantly reduce the number of model parameters and speed up the training of the model by using a parameter-sharing convolution kernel;

[0035] ② When the prior art directly uses the conventional K-Means clustering model to cluster the data set, the K-Means clustering model will directly regard the importance of each feature as the same, but in fact the importance of different features is not the same, so the final result obtained by the prior art does not conform to the actual situation. The present invention uses information entropy to calculate the weights of different features, and before using the conventional K-Means clustering model to cluster the data set, it first distinguishes the importance of various feature vectors in the data sample, so that the final result is consistent with the actual situation, thereby having real guiding significance in the field of test scene extraction for autonomous driving vehicles. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flow chart of the present invention;

[0037] Figure 2 A schematic diagram of scene data of the high-dimensional scene feature data set of the present invention;

[0038] Figure 3 Schematic diagram of the structure of the one-dimensional residual convolutional autoencoder of the present invention;

[0039] Figure 4A schematic diagram showing a performance comparison between the one-dimensional residual convolutional autoencoder of the present invention and other common autoencoders;

[0040] Figure 5 A comparison chart of clustering performance between the present invention and other clustering algorithms;

[0041] Figure 6 It is a schematic diagram of the silhouette coefficient under different K values ​​of the present invention;

[0042] Figure 7 It is a data statistical diagram corresponding to each scene type in the embodiment of the present invention. DETAILED DESCRIPTION

[0043] like Figure 1 As shown, a method for extracting autonomous driving test scenarios based on a one-dimensional residual convolutional autoencoder includes the following steps:

[0044] 1) Obtain a natural driving data set and preprocess the natural driving data set;

[0045] The natural driving dataset is generally a public dataset or road test data collected in real time, and includes a large amount of data corresponding to several feature types.

[0046] In this embodiment, the natural driving data set adopts the American DAS2 natural driving data set, and the sampling frequency is 10 Hz;

[0047] The preprocessing methods for natural driving datasets include data cleaning and / or data filtering, for example:

[0048] ① For missing data, the average value of all data before and after the missing data within one period (1s) is used to fill in the missing data;

[0049] ② Delete the duplicate scene data directly;

[0050] ③ In order to eliminate the noise and outliers in the vehicle speed, acceleration, steering wheel angle and other data caused by vehicle shaking during data collection, robust local weighted linear regression is used to filter the data;

[0051] 2) Select a number of different scene data from the natural driving dataset to form a high-dimensional scene feature dataset;

[0052] It is worth noting that the time length of the scene data is too long, that is, the time span of the scene is large, the data contained corresponds to multiple different types of scenes, and the feature types contained are complex and not representative, and cannot be used as scene data for extracting feature types; and the time length of the scene data is too short, that is, the time span of the scene is short, and the data contained is insufficient to distinguish the feature types represented by the scene.

[0053] Therefore, before selecting scene data from the natural driving data set, it is necessary to set a time length threshold to ensure that the time length of all selected scene data must be within the time length threshold range.

[0054] The time length threshold is an empirical value, which usually includes a minimum value and a maximum value. For example, in this embodiment, the minimum time length is set to 0.5 seconds, and the maximum time length is set to 10 seconds, that is, in the natural driving data set, as long as the natural driving data corresponding to the continuous scene segment with a time length between 0.5 seconds and 10 seconds can be extracted as scene data.

[0055] This embodiment extracts a total of 4736 scenes data from the American DAS2 natural driving dataset. The set formed by these 4736 scenes data is a high-dimensional scene feature dataset.

[0056] 3) Using a one-dimensional residual convolutional autoencoder, the feature dimension reduction of each scene data of the high-dimensional scene feature data set is performed to form a low-dimensional feature type data set;

[0057] like Figure 2 As shown, each segment of scene data in the high-dimensional scene feature data set includes vehicle speed, vehicle acceleration, steering wheel angle, left and right turn signal signals, brake pedal status, yaw rate, pitch angular velocity, roll angular velocity, position coordinates (latitude and longitude) of the vehicle, and relative position coordinates and relative driving speed of the vehicle and other road vehicles corresponding to multiple moments, and multiple feature types need to be reduced through feature dimensionality reduction.

[0058] In this embodiment, a one-dimensional residual convolutional autoencoder is used to reduce the dimension of each scene data of a high-dimensional scene feature data set. Figure 3 As shown, it includes an encoder and a decoder, the encoder includes a first convolutional layer, a second convolutional layer, a third convolutional layer, and a first bottleneck layer, a first maximum pooling layer is arranged between the first convolutional layer and the second convolutional layer, and a second maximum pooling layer is arranged between the third convolutional layer and the first bottleneck layer;

[0059] The decoder includes a first deconvolution layer, a second deconvolution layer, a third deconvolution layer, and a second bottleneck layer, a first upsampling layer is provided between the second bottleneck layer and the first deconvolution layer, a second jump connection point is provided between the first upsampling layer and the first deconvolution layer, a second upsampling layer is provided between the second deconvolution layer and the third deconvolution layer, and a first jump connection point is provided between the second upsampling layer and the second deconvolution layer;

[0060] The input features of the encoder enter the encoder through the first convolution layer, then pass through the first maximum pooling layer, the second convolution layer, the third convolution layer, the second maximum pooling layer in sequence, and finally are output from the first bottleneck layer to the second bottleneck layer of the decoder, and then pass through the first upsampling layer, the first deconvolution layer, the second deconvolution layer, the second upsampling layer in sequence, and finally are output from the third deconvolution layer;

[0061] The output features of the first maximum pooling layer are added to the output features of the second deconvolution layer through the first skip connection point as input features of the second upsampling layer;

[0062] The output features of the third convolutional layer are added to the output features of the first upsampling layer through the second jump connection point as input features of the first deconvolution layer.

[0063] The mathematical expression of the one-dimensional convolution process of the one-dimensional residual convolution autoencoder is as follows:

[0064]

[0065] Where δ(·) is the activation function, and Represent the i-th feature vector of the l-th and l-1-th convolutional layers respectively, is the weight between the convolution kernel and the input data, Represents the bias term, and N is the dimension of the convolutional layer input data.

[0066] The deconvolution process is similar to the convolution process, but it should be noted that the weights in the above formula need to be transposed to obtain the convolution kernel in the one-dimensional deconvolution calculation. In order to ensure that the data size of the deconvolution output is larger than the input data size, the input data of the deconvolution needs to be padded (edge ​​filling) before the convolution calculation.

[0067] The pooling layer of the one-dimensional residual convolutional autoencoder follows the one-dimensional convolutional layer, and downsamples the features output by the convolutional layer to reduce model parameters and calculation amount, which also helps to alleviate overfitting.

[0068] The upsampling layer of the one-dimensional residual convolutional autoencoder is used to upsample the features so as to better reconstruct the original input data.

[0069] The one-dimensional residual convolutional autoencoder learns features by reconstructing the original input, and its loss function is:

[0070]

[0071] Where N is the total number of samples, x i is the i-th sample data input, y iis the i-th sample data output, and Loss is the loss function.

[0072] In this embodiment, the network parameters when the loss function Loss obtains the minimum value are the parameters of the trained one-dimensional residual convolutional autoencoder neural network, as shown in Table 1:

[0073]

[0074]

[0075] Table 1

[0076] Compared with ordinary autoencoders, the one-dimensional residual convolutional autoencoder can significantly reduce the number of model parameters and speed up model training by using parameter-sharing convolution kernels.

[0077] The training process of the one-dimensional residual convolutional autoencoder is a parameter learning process. In this process, the learning goal of the one-dimensional residual convolutional autoencoder is to make the network output equal to the input.

[0078] The type of output features of the first bottleneck layer in the one-dimensional residual convolutional autoencoder is determined by the number of input features and the encoder parameters.

[0079] In this embodiment, the activation function used to train the one-dimensional residual convolutional autoencoder is the Relu activation function, as shown in the following formula:

[0080] ReLU(x)=max(0,x)

[0081] Where x is the input real value or vector.

[0082] For real-valued x, the ReLU function turns negative values ​​to zero, while for non-negative values ​​x, the ReLU function remains unchanged. Therefore, for real-valued x, the output of the ReLU function can be expressed as: if x is greater than or equal to zero, the output is x; if x is less than zero, the output is zero.

[0083] In deep learning, the ReLU function is often used to select activation functions to introduce nonlinear properties and nonlinear transformations. It has the characteristics of simple calculation and high efficiency, and can alleviate the problem of gradient disappearance, which is helpful for network training and convergence. In each neuron of a neural network, the ReLU function is usually applied to the output of the neuron to introduce nonlinear transformations and activations.

[0084] For vector x, the ReLU function applies the ReLU operation to each element in the vector to obtain the corresponding output vector. Therefore, x can represent a single real value or a vector composed of multiple real values ​​in the ReLU function.

[0085] like Figure 4 As shown, the ordinary autoencoder (AE), regularized autoencoder (RAE), denoising autoencoder (DAE), convolutional autoencoder (CAE) and the one-dimensional residual convolutional autoencoder (1DRCAE) described in the present invention are compared from three aspects of the mean square error (MSE), mean absolute error (MAE) and root mean square error (RMSE) of network training.

[0086] The results show that the one-dimensional residual convolutional autoencoder (1DRCAE) described in the present invention has a more excellent feature extraction capability. This embodiment uses a one-dimensional residual convolutional autoencoder to perform feature dimensionality reduction on each scene data of the high-dimensional scene feature data set. In fact, each scene data is input into the trained one-dimensional residual convolutional autoencoder, and finally the output of the first bottleneck layer is the reduced dimensionality feature data corresponding to each scene data, as shown in Table 2:

[0087]

[0088] Table 2

[0089] It should be noted that in order to prevent the data distribution from becoming dispersed due to internal variable offset after multiple convolutions, resulting in problems such as slow model training speed and low efficiency, the present invention also connects a batch normalization (BN) layer after each convolution layer, that is, the output features of the first convolution layer, the second convolution layer, the third convolution layer, the first deconvolution layer, the second deconvolution layer, and the third deconvolution layer are all processed in a batch normalization manner and then input into the next layer.

[0090] Compared with traditional normalization technology, the BN layer has the following advantages: First, it accelerates the convergence speed of deep neural networks and shortens the training time; second, it can effectively normalize the feature distribution in the network, reduce the risk of overfitting, and improve the generalization performance of the model; third, it can reduce the scale difference of features between different layers, thereby making the initialized weights more robust.

[0091] The BN layer normalizes the data of each batch so that the mean and variance of each feature channel are close to 0 and 1 in each batch, thereby avoiding the problems of gradient vanishing and explosion in deep networks and improving the stability and generalization performance of the network.

[0092] In this embodiment, the batch input data x is assumed to be B = {x1, x2, ..., x m}, the mathematical expression of the BN layer calculation process is as follows:

[0093]

[0094]

[0095]

[0096]

[0097] In the formula, μ β is the mean of the batch data, is the variance of the batch data, is the normalized data, y i is the output data after batch normalization, γ is the scale factor, β is the translation factor, and ε is a positive number that avoids division by zero, which is infinitely close to 0.

[0098] In this embodiment, the batch data refers to the number of data sent to the one-dimensional residual convolutional autoencoder in each batch. For example, if the batch is 32, it means that 32 data are sent each time, and the mean of the batch data is the mean of these 32 data.

[0099] 4) Calculate the weights corresponding to each type of feature vector of the low-dimensional feature type data set respectively, and weight the feature data corresponding to each type of feature vector of the low-dimensional feature type data set to obtain the weighted low-dimensional feature type data set.

[0100] Before clustering, information entropy is used to calculate the weights of various feature vectors, which can better measure the degree of data dispersion and uncertainty within the cluster. Smaller information entropy corresponds to higher weights, indicating that the feature vector is more able to reflect the differences between different samples compared with other feature vectors;

[0101] A larger information entropy corresponds to a lower weight, indicating that the feature vector is less able to reflect the differences between different samples compared with other feature vectors;

[0102] For example, the steps to calculate the weight corresponding to the feature vector A in the low-dimensional feature type data set are as follows:

[0103] 1) Standardize the feature data xi-A corresponding to the feature vector A of the i-th scene data in the low-dimensional feature type data set (for example, the range standardization method, or the Z-score standardization method, etc.);

[0104] 2) Calculate the proportion of the i-th scene data in the feature vector A:

[0105]

[0106] In the formula, p i-A is the proportion of the i-th scene data in the feature vector A, x′ i-Ais the feature data corresponding to the feature vector A of the i-th scene data after normalization, and n is the number of scene data in the low-dimensional feature type data set;

[0107] 3) Calculate the information entropy value E1 of the feature vector A in the low-dimensional feature type data set:

[0108]

[0109] Where E1 is the information entropy value of feature vector A in the low-dimensional feature type dataset, n is the number of scene data in the low-dimensional feature type dataset, and p i-A is the proportion of the i-th scene data in the feature vector A;

[0110] 4) Calculate the weight w of feature vector A according to the following formula: A :

[0111]

[0112] In the formula, w A is the weight of the feature vector j of the low-dimensional feature type data set, s is the number of different feature vectors in the low-dimensional feature type data set, and E1 is the information entropy value of the feature vector A in the low-dimensional feature type data set.

[0113] By analogy, the weights of feature vectors A, B, and C in the low-dimensional feature type data set are shown in Table 3:

[0114]

[0115] Table 3

[0116] In this embodiment, the feature data corresponding to each type of feature vector of the low-dimensional feature type data set is weighted, that is, all feature data corresponding to each type of feature vector of the low-dimensional feature type data set are multiplied by the weight corresponding to the feature vector of this type, and the feature data corresponding to each scene data in the weighted low-dimensional feature type data set are shown in Table 4 (according to Table 3, the weight A in Table 4 is 0.3251, the weight B is 0.3655, and the weight C is 0.3094):

[0117]

[0118] Table 4

[0119] 5) Use the K-Means clustering model to cluster the weighted low-dimensional feature type data set to obtain the scene data classification of the low-dimensional feature type data set, and number different scene categories;

[0120] Usually, the method of clustering a data set using the K-Means clustering model is as follows:

[0121] Step 1: Determine the number of categories K by the silhouette coefficient, and randomly select K points as the initial center points within the distribution range of the training data;

[0122] Step 2: According to the minimum distance principle, all data points are divided into the class where the nearest center point is located;

[0123] Step 3: There are several observation data points in each class. Calculate the mean of all data points in K classes as the center point of the next iteration;

[0124] Step 4: Repeat step 2 and step 3 until convergence (the category or center point of each data point no longer changes), and the clustering process ends.

[0125] The K-Means clustering model used in this embodiment is also a conventional K-Means clustering model, so the clustering process is as follows:

[0126] 1) Determine the K value of the K-Means clustering model by the silhouette coefficient method, and randomly initialize the cluster center (that is, randomly select K points as the initial center points within the distribution range of the training data);

[0127] The silhouette coefficient calculation method can be expressed by the following formula:

[0128]

[0129] In the formula, a(i) represents cohesion, b(i) represents separation, and the silhouette coefficient ranges from [-1, 1]. The larger the silhouette coefficient, the better the clustering effect.

[0130] 2) Calculate the Euclidean distance between each feature and the cluster center. That is, when clustering, the Euclidean distance is used as the similarity measurement indicator. The Euclidean distance calculation method is as follows:

[0131]

[0132] In the formula, x is the input data, C i is the i-th cluster center, n is the input data dimension, x j , C ij are x and C respectively i The j-th attribute value of ;

[0133] 3) According to the minimum distance principle, the data is assigned to the cluster with the closest cluster center according to the Euclidean distance of the cluster center;

[0134] 4) Calculate the mean of all objects in each cluster and the sum of squared errors within each cluster;

[0135] 5) Repeat steps 2) to 4), iteratively update the cluster center until the sum of squared errors within each cluster converges, and output the clustering result. That is, based on the Euclidean distance, K-Means clustering needs to minimize the sum of squared errors within the cluster (SSE). With the continuous iteration of the algorithm, the SSE value gradually decreases until it becomes stable. When the SSE value converges, the clustering is completed. The SSE value calculation method is:

[0136]

[0137] In the formula, x is the input data, C i is the i-th cluster center, and SSE is the sum of squared errors within the cluster.

[0138] Since the Davidson-Botting index can be expressed as:

[0139]

[0140] In the formula, N means there are N clusters, is the average Euclidean distance from the i-th class sample to its class center, is the average Euclidean distance from the j-th class sample to its class center, ||w i -w j ||2 is the Euclidean distance between the class centers of the i-th and j-th classes.

[0141] like Figure 5 As shown, the smaller the DBI value is, the better the clustering performance is, that is, the K-Means clustering model of the present invention has the best clustering performance compared with other clustering models.

[0142] like Figure 6 As shown, the clustering result of this embodiment is 7, that is, the clustering K value determined according to the silhouette coefficient is 7, and the scene data of the low-dimensional feature type data set can be divided into 7 categories in total, which are the category numbers of the 7 different types of scene data (the high-dimensional scene feature data set and the low-dimensional feature type data set have the same data classification, and the scene data ID and the scene category ID correspond one to one), for example, scene category A is numbered 1, scene category B is numbered 2, scene category C is numbered 3, and so on, as shown in Table 5.

[0143]

[0144] Table 5

[0145] Reference Figure 7By counting the number of seven types of scene data in the high-dimensional scene feature data set, it can be seen that the scene category A includes 1441 scene data, the scene category B includes 2032 scene data, the scene category C includes 185 scene data, the scene category D includes 718 scene data, the scene category E includes 123 scene data, the scene category F includes 103 scene data, and the scene category G includes 134 scene data.

[0146] 6) Setting the naming conditions for each scene category, and naming each scene category according to a number of scene data corresponding to each scene category in the high-dimensional scene feature data set.

[0147] This embodiment uses detailed data corresponding to feature types such as vehicle speed, vehicle acceleration, steering wheel angle, turn signal signal, etc. in various scene data in the high-dimensional scene feature data set to set the naming conditions, which are as follows:

[0148] ① Determine whether the vehicle is in a uniform speed state or an accelerating state based on the absolute value of the average acceleration and the standard deviation of the acceleration:

[0149] The larger the absolute value of the average acceleration, the greater the difference in the vehicle's driving speed at each moment in the scene data, and the less likely it is that the vehicle is driving at a constant speed;

[0150] The larger the standard deviation of acceleration, the greater the acceleration fluctuation of the vehicle in this scene data, and the more unstable the vehicle's driving condition.

[0151] Ⅰ. If the scene data corresponding to the current scene category meets both of the following conditions, the vehicles in this scene category are in a uniform speed state:

[0152] ⑴The absolute value of the average acceleration is less than 0.1;

[0153] (2) The standard deviation of acceleration is less than 0.2;

[0154] For example, the scene data corresponding to scene category A satisfies the condition that the vehicle is in a uniform speed state, that is, the vehicle in scene category A is in a uniform speed state;

[0155] Ⅱ. If the scene data corresponding to the current scene category satisfies both of the following conditions, the vehicle in this scene category is in an accelerating state:

[0156] ⑴The absolute value of the average acceleration is ≥ 0.1;

[0157] ⑵The standard deviation of acceleration is ≥ 0.2;

[0158] For example, the scene data corresponding to scene category B satisfies the condition that the vehicle is in an accelerating state, that is, the vehicle in scene category B is in an accelerating state;

[0159] ② According to the vehicle speed with the highest proportion in the same scene category, determine the speed level of the vehicle in the scene category and the corresponding vehicle speed status:

[0160] ⑴Low speed level (low speed state): vehicle speed <11m / s;

[0161] ⑵Medium speed level (medium speed state): 11m / s≤Vehicle speed≤22m / s

[0162] ⑶ High-speed level (high-speed state): vehicle speed>22m / s;

[0163] For example, in the scene data corresponding to scene categories A and B, the vehicle speed greater than 22m / s accounts for the highest proportion, that is, the vehicle speed in scene categories A and B is at a high-speed level, that is, the vehicles in scene categories A and B are in a high-speed state;

[0164] For example, in the scene data corresponding to scene categories C and D, the vehicle speed of less than 11m / s accounts for the highest proportion, that is, the vehicle speed in scene categories C and D is at a low speed level, that is, the vehicles in scene categories C and D are in a low speed state;

[0165] ③ According to the maximum absolute value of the steering wheel angle of the vehicles in the same scene category and the turn signal, determine whether the vehicle in the scene category is in a straight state or a non-straight state:

[0166] ⑴ If the left and right turn signal data in this scene category are both zero, and 0°≤the maximum absolute value of the vehicle's steering wheel angle≤10°, then the vehicle in this scene category is in a straight-ahead state;

[0167] (2) If the amount of left turn signal data or the amount of right turn signal data in the scene category is not zero, and the vehicle in the scene category is in a non-straight state (non-straight state includes lane changing, turning, and turning at an intersection), then:

[0168] ① When the left turn signal data volume in this scenario category accounts for the highest proportion, the turning direction of the vehicle in this scenario category is left. According to the numerical value of the maximum absolute value of the steering wheel turning angle, it is judged that the vehicle in this scenario category is changing lanes to the left, turning left, or turning left at the intersection:

[0169] If 0°≤the maximum absolute value of the vehicle's steering wheel angle≤10°, the vehicle in this scenario category is in a left lane change state;

[0170] If 10°≤the maximum absolute value of the vehicle's steering wheel angle≤50°, the vehicle in this scenario is in a left turning state;

[0171] If 50°≤the maximum absolute value of the steering wheel angle of the vehicle≤180°, the vehicle in this scenario category is in a state of turning left at the intersection;

[0172] ② When the right turn signal data volume in this scenario category accounts for the highest proportion, the turning direction of the vehicle in this scenario category is right. According to the numerical value of the maximum absolute value of the steering wheel turning angle, it is judged that the vehicle in this scenario category is changing lanes to the right, turning right, or turning right at the intersection:

[0173] If 0°≤the maximum absolute value of the vehicle's steering wheel angle≤10°, the vehicle in this scenario is in a right lane change state;

[0174] If 10°≤the maximum absolute value of the vehicle's steering wheel angle≤50°, the vehicle in this scenario is turning right;

[0175] If 50°≤the maximum absolute value of the vehicle's steering wheel angle≤180°, the vehicle in this scenario category is turning right at the intersection;

[0176] It is worth noting that if the amount of left and right turn signal data in the scene category is much smaller than the number of scene data in the scene category (usually a threshold is set for judgment), the proportion of the left and right turn signal data in the scene category cannot be directly used to judge the turning direction of the vehicle in the scene category. The turning direction and status of the vehicle in the scene category can only be judged by the positive and negative signs of the steering wheel turning angle of the vehicle in the scene category combined with the maximum absolute value of the steering wheel turning angle of the vehicle.

[0177] For example, although the amount of left turn signal data in scene category D is not zero, the number of scene data in scene category D is 718, and the amount of left turn signal data in scene category D is 2, that is, the amount of left turn signal data in scene category D is much smaller than the number of scene data in scene category D. At this time, it cannot be directly used to determine the turning direction of the vehicle in this scene category.

[0178] The minimum steering wheel angle in scenario category D is -15.02, that is, the minimum steering wheel angle in scenario category D is a negative value, and its absolute value is 15.02;

[0179] The maximum steering wheel angle in scenario category D is 18.9, that is, the maximum steering wheel angle in scenario category D is a positive value, and its absolute value is 18.9;

[0180] The sign of the steering wheel angle corresponding to the maximum absolute value of the steering wheel turning angle of the vehicle in scene category D represents the turning direction of the vehicle in this scene category, that is, |+18.9|>|-15.02|, and the turning direction of the vehicle in scene category D is right (negative sign means left, positive sign means right).

[0181] After determining the turning direction of the vehicle in scenario category D, the maximum absolute value of the steering wheel turning angle is used to determine whether the vehicle in the scenario category is changing lanes, turning, or turning at an intersection:

[0182] If 0°≤the maximum absolute value of the vehicle's steering wheel angle≤10°, the vehicle in this scenario is in a lane-changing state; if 10°≤the maximum absolute value of the vehicle's steering wheel angle≤50°, the vehicle in this scenario is in a cornering state; if 50°≤the maximum absolute value of the vehicle's steering wheel angle≤180°, the vehicle in this scenario is in an intersection turning state;

[0183] Therefore, the vehicle in scene category D is in a state of turning right.

[0184] To sum up, in this embodiment, according to the above naming rules, characteristic parameters such as vehicle speed, acceleration, steering angle, etc. are selected to perform statistical analysis on the data corresponding to scene categories A, B, C, D, E, F, and G, as shown in Table 6. Seven types of driving scenes can be extracted from the natural driving data set, as shown in Table 6, that is, scene category A is a high-speed uniform straight-driving scene, scene category B is a high-speed accelerated straight-driving scene, scene category C is a scene of decelerating and turning left at an intersection, scene category D is a scene of slowing down and turning right at a low speed, scene category E is a high-speed left lane change scene, scene category F is a high-speed right lane change scene, and scene category G is a scene of decelerating and turning right at an intersection.

[0185]

[0186] Table 6

[0187] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modification made to the present invention by those skilled in the art without departing from the spirit of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for extracting autonomous driving test scenarios based on a one-dimensional residual convolutional autoencoder, characterized in that: The following steps are involved: 1) Obtain a natural driving data set and preprocess the natural driving data set; 2) Select a number of different scene data from the natural driving dataset to form a high-dimensional scene feature dataset; 3) Using a one-dimensional residual convolutional autoencoder, the feature dimension reduction of each scene data of the high-dimensional scene feature data set is performed to form a low-dimensional feature type data set; The one-dimensional residual convolutional autoencoder includes an encoder and a decoder, the encoder includes a first convolutional layer, a second convolutional layer, a third convolutional layer, and a first bottleneck layer, a first maximum pooling layer is provided between the first convolutional layer and the second convolutional layer, and a second maximum pooling layer is provided between the third convolutional layer and the first bottleneck layer; The decoder includes a first deconvolution layer, a second deconvolution layer, a third deconvolution layer, and a second bottleneck layer, a first upsampling layer is provided between the second bottleneck layer and the first deconvolution layer, a second jump connection point is provided between the first upsampling layer and the first deconvolution layer, a second upsampling layer is provided between the second deconvolution layer and the third deconvolution layer, and a first jump connection point is provided between the second upsampling layer and the second deconvolution layer; The input features of the encoder enter the encoder through the first convolution layer, then pass through the first maximum pooling layer, the second convolution layer, the third convolution layer, the second maximum pooling layer in sequence, and finally are output from the first bottleneck layer to the second bottleneck layer of the decoder, and then pass through the first upsampling layer, the first deconvolution layer, the second deconvolution layer, the second upsampling layer in sequence, and finally are output from the third deconvolution layer; The output features of the first maximum pooling layer are added to the output features of the second deconvolution layer through the first skip connection point as input features of the second upsampling layer; The output features of the third convolutional layer are added to the output features of the first upsampling layer through the second jump connection point as input features of the first deconvolutional layer; The output features of the first convolutional layer, the second convolutional layer, the third convolutional layer, the first deconvolutional layer, the second deconvolutional layer, and the third deconvolutional layer are all processed in a batch normalization manner and then input into the next layer; 4) Calculate the weights corresponding to each type of feature vector of the low-dimensional feature type data set respectively, and weight the feature data corresponding to each type of feature vector of the low-dimensional feature type data set to obtain a weighted low-dimensional feature type data set; 5) Use the K-Means clustering model to cluster the weighted low-dimensional feature type data set to obtain the scene data classification of the low-dimensional feature type data set and number it; 6) Setting the naming conditions for each scene category, and naming each scene category according to a number of scene data corresponding to each scene category in the high-dimensional scene feature data set.

2. The method for extracting autonomous driving test scenarios based on a one-dimensional residual convolutional autoencoder according to claim 1, characterized in that: The specific calculation process of the weight corresponding to any type of feature vector in the low-dimensional feature type data set is as follows: 1) The feature data x corresponding to the feature vector j of the i-th scene data in the low-dimensional feature type dataset i-j Standardize the process; 2) Calculate the proportion of the i-th scene data in the feature vector j: In the formula, p i-j is the proportion of the i-th scene data in the feature vector j, x i ' -j is the feature data corresponding to the feature vector j of the i-th scene data after normalization, and n is the number of scene data in the low-dimensional feature type data set; 3) Calculate the information entropy value E of feature vector j in the low-dimensional feature type data set j : In the formula, E j is the information entropy value of feature vector j in the low-dimensional feature type dataset, n is the number of scene data in the low-dimensional feature type dataset, and p i-j is the proportion of the i-th scene data in the feature vector j; 4) Calculate the weight w of feature vector j according to the following formula j : In the formula, w j is the weight of the feature vector j of the low-dimensional feature type dataset, s is the number of feature vectors in the low-dimensional feature type dataset, E j is the information entropy value of feature vector j in the low-dimensional feature type data set.

3. The method for extracting autonomous driving test scenarios based on a one-dimensional residual convolutional autoencoder according to claim 1, characterized in that: The preprocessing methods for natural driving datasets include data cleaning and / or data filtering.

4. The method for extracting autonomous driving test scenarios based on a one-dimensional residual convolutional autoencoder according to claim 1, characterized in that: A time length threshold is set, and the time length of each scene data segment selected from the natural driving data set is within the time length threshold range.

Images