A driving style recognition method, system, storage medium and electronic device

Abstract

The application relates to a driving style recognition method, system, storage medium and electronic equipment. Driving data is acquired, the driving data is divided into scenes, a segment set under each driving scene is obtained, and the segment set comprises driving segment data; the segment set under each driving scene is subjected to data processing to obtain a style label; the data processing comprises: dimension reduction is performed on the driving segment data under the driving scene to extract a feature set corresponding to the segment set, the feature set comprising driving features corresponding to the driving segment data; the driving features contained in the feature set are clustered to obtain the number of driving features contained in the feature set belonging to each driving style and the driving style weight; based on the number of each driving style and the driving style weight, the style label under the driving scene corresponding to the feature set is obtained; and based on the style label under each driving scene, the driving style corresponding to the driving data is obtained. The application improves the recognition accuracy and generalization ability.

Description

Technical Field

[0001] This application relates to the fields of computer science and intelligent connected vehicle technology, and in particular to a method, system, storage medium and electronic device for driving style recognition. Background Technology

[0002] With the rapid development of the automotive industry, electrification, connectivity, intelligence, and sharing have become strategic trends in the global automotive sector. Against this backdrop, intelligent connected vehicles are gradually becoming the focus of the industry. Advanced Driver Assistance Systems (ADAS), as a core component of intelligent connected vehicles, not only significantly improve driver convenience and vehicle safety but also effectively enhance risk avoidance capabilities, providing more comprehensive protection for the health and safety of road users. In the further development of ADAS, the accurate identification and application of driving styles plays a crucial role. Driving style recognition not only helps optimize the functions of ADAS, enabling them to adaptively adjust according to driver behavior, but also promotes personalized driving experiences. For example, by identifying a driver's conservative, average, or aggressive driving style, ADAS systems can adjust power output modes, braking response, etc., accordingly to meet the personalized needs of different drivers. Therefore, building an efficient and accurate driving style recognition system has become a key task in promoting the technological advancement of intelligent connected vehicles.

[0003] In the field of driving style recognition, scholars both domestically and internationally have conducted extensive research. Early studies mainly focused on questionnaire surveys or human scoring methods, emphasizing the impact of factors such as culture, age, and gender on differences in driver behavior. These methods often require well-designed questionnaires, and subjective factors have a significant influence on the recognition process, making it difficult to accurately identify driving styles. In recent years, with the development of intelligent connected vehicles, the sensors installed on vehicles have become increasingly diversified, and using historical data collected by sensors or cameras for algorithmic recognition has become the preferred method.

[0004] For example, some related technologies use exploratory data analysis to identify candidate driving style features and their relationship with driving style, and then select key features to build a linear regression model to identify the driver's driving style. However, this method may face problems such as incomplete or biased feature selection during the feature selection stage, which may affect the model's predictive accuracy. Furthermore, driving style identification methods based on a single linear regression model may have limitations in generalization ability because they may not be able to fully capture the complex interactive behaviors and nonlinear relationships in driving data.

[0005] Therefore, in order to improve the accuracy of recognition and the generalization ability of the model, it may be necessary to consider adopting more advanced feature extraction methods and loss optimization methods. Summary of the Invention

[0006] This application provides a method, system, storage medium, and electronic device for driving style recognition to address the problem of low prediction accuracy in related technologies.

[0007] Firstly, a method for driving style recognition is provided, which includes:

[0008] Acquire driving data and divide the driving data into scenarios to obtain a set of segments for each driving scenario, the set of segments including driving segment data;

[0009] A style labeling model is used to process data sets of segments under various driving scenarios to obtain style labels for each driving scenario. The data processing includes: dimensionality reduction of the driving segment data contained in the segment set under the driving scenario to extract the feature set corresponding to the segment set, the feature set including driving features corresponding to the driving segment data contained in the segment set; clustering the driving features contained in the feature set to obtain the number of driving features contained in the feature set belonging to each driving style and the driving style weight; and obtaining the style label for the driving scenario corresponding to the feature set based on the number of driving styles and the driving style weight.

[0010] Based on the style labels for each driving scenario, the driving style corresponding to the driving data is obtained.

[0011] In some embodiments, the driving scenarios include lane-changing scenarios, acceleration scenarios, and deceleration scenarios.

[0012] In some embodiments, the driving data is divided into scenarios to obtain a set of segments for each driving scenario, specifically including:

[0013] Data of a first time length is extracted from the driving data to obtain a set of segments in the lane change scenario. The data of the first time length satisfies the following condition: one of the sampling time points in the data of the first time length is the lane change time point.

[0014] The driving data is extracted as driving segment data for a second time length to obtain a segment set under the acceleration driving scenario. The second time length data satisfies the following condition: in the second time length data, in two adjacent sampling time points, the speed of the later sampling time point is greater than the speed of the earlier sampling time point.

[0015] The driving data is extracted into a third time length as driving segment data to obtain a segment set under the deceleration driving scenario. The third time length data satisfies the following condition: in the third time length data, in two adjacent sampling time points, the speed of the later sampling time point is less than the speed of the earlier sampling time point.

[0016] In some embodiments, a style tagging model is used to process the data sets of segments under each driving scenario to obtain style tags for each driving scenario, specifically including:

[0017] The driving segment data contained in the segment set of each driving scenario is used as input data and input into the style label model corresponding to each driving scenario to output the style label corresponding to each driving scenario.

[0018] In some embodiments, the driving segment data contained in the segment set under the driving scenario is subjected to dimensionality reduction to extract the feature set corresponding to the segment set, specifically including:

[0019] The driving segment data contained in the segment set under the driving scenario is reduced in dimensionality by using the temporal autoencoder corresponding to the driving scenario to extract the driving features corresponding to each driving segment data and obtain the feature set corresponding to the segment set under the driving scenario.

[0020] In some embodiments, the quantity includes quantity percentage or driving time percentage.

[0021] In some embodiments, the driving features contained in the feature set are clustered to obtain the number of driving features in the feature set belonging to each driving style and the driving style weights, specifically including:

[0022] The driving features contained in the feature set are clustered to obtain multiple sub-feature sets, and the number of driving features divided from the feature set is included in each sub-feature set.

[0023] Based on the driving features included in each of the sub-feature sets, the driving style to which each sub-feature set belongs, and the corresponding driving style weight, are determined.

[0024] In some embodiments, based on the driving features included in each of the sub-feature sets, the driving style to which each sub-feature set belongs, and the corresponding driving style weight, are determined, specifically including:

[0025] Based on the driving segment data corresponding to the driving features included in each of the sub-feature sets, obtain the preset variable values ​​corresponding to each of the sub-feature sets;

[0026] Based on the magnitude of the preset variable values ​​corresponding to each of the sub-feature sets, the driving style to which each sub-feature set belongs, and the corresponding driving style weight, are obtained.

[0027] In some embodiments, the preset variable value is obtained based on velocity or acceleration.

[0028] In some embodiments, the preset variable values ​​are selected from maximum speed, mean speed, standard deviation of speed, maximum acceleration, mean acceleration, standard deviation of acceleration, maximum lateral speed, mean lateral speed, and standard deviation of lateral speed.

[0029] In some embodiments, style labels for driving scenarios corresponding to the feature set are obtained based on the quantity and weight of each driving style, specifically including:

[0030] Multiply the quantity of each driving style by the corresponding driving style weight to obtain the product value of each driving style;

[0031] The product values ​​are summed to obtain the style label for the driving scenario.

[0032] In some embodiments, the driving style corresponding to the driving data is obtained based on the style tags under each driving scenario, specifically including:

[0033] The style tags under each driving scenario are summed to obtain the driving score of the driving data.

[0034] Based on the scoring range of each driving style and the driving score, the driving style corresponding to the driving data is determined.

[0035] In some embodiments, the driving styles include conservative, normal, and aggressive styles.

[0036] In some embodiments, before processing the data sets of segments for each driving scenario using a style tagging model, the method further includes:

[0037] Construct a sample dataset, which includes several driving data sets;

[0038] The driving data is divided into scenarios to obtain a set of segments for each driving scenario;

[0039] Construct initial style label models for various driving scenarios;

[0040] Using a set of clips from various driving scenarios, the initial style label model for each driving scenario is iteratively trained to obtain the style label model for each driving scenario.

[0041] In some embodiments, the style tagging model includes a temporal autoencoder and a deep clustering model;

[0042] The temporal autoencoder is used to: reduce the dimensionality of the driving segment data contained in the segment set under the driving scenario, so as to extract the feature set corresponding to the segment set, wherein the feature set includes driving features corresponding to the driving segment data contained in the segment set;

[0043] The deep clustering model is used to: cluster the driving features contained in the feature set to obtain the number of driving features in the feature set belonging to each driving style and the driving style weight; and obtain the style label of the driving scenario corresponding to the feature set based on the number of each driving style and the driving style weight.

[0044] In some embodiments, the temporal autoencoder includes an encoder for extracting driving features and a decoder for autoregression of driving features;

[0045] The encoder includes a first layer based on a temporal convolutional network (TCN) and the activation function ReLU, a second layer based on MaPool, a third layer based on bidirectional LSTM, and a fourth layer based on bidirectional LSTM.

[0046] The decoder consists of a first layer based on upsampling and a second layer based on deconvolution.

[0047] In some embodiments, the deep clustering model uses the K-means++ algorithm to determine the initial cluster centers and the MiniBatchKmeans algorithm to perform clustering.

[0048] Secondly, a driving style recognition system is provided, comprising:

[0049] An acquisition module is used to acquire driving data and divide the driving data into scenarios to obtain a set of segments under each driving scenario, wherein the set of segments includes driving segment data.

[0050] A style tag model is used to process data sets of segments under various driving scenarios to obtain style tags for each driving scenario. The data processing includes: reducing the dimensionality of the driving segment data contained in the segment set under the driving scenario to extract the feature set corresponding to the segment set, the feature set including driving features corresponding to the driving segment data contained in the segment set; clustering the driving features contained in the feature set to obtain the number of driving features contained in the feature set belonging to each driving style and the driving style weight; and obtaining the style tag for the driving scenario corresponding to the feature set based on the number of driving styles and the driving style weight.

[0051] The driving style recognition module is used to obtain the driving style corresponding to the driving data based on the style labels under each driving scenario.

[0052] Thirdly, a storage medium is provided that stores a computer program, which, when executed by a processor, implements the driving style recognition method as described above.

[0053] Fourthly, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program that runs on the processor, and the processor executes the computer program to implement the driving style recognition method as described above.

[0054] The beneficial effects of the technical solution provided in this application include:

[0055] This application provides a method, system, storage medium, and electronic device for driving style recognition. This application comprehensively considers various driving styles of a driver under different driving scenarios based on driving data over a period of time, taking into account clustering errors or the possibility that the driver may exhibit unstable driving styles. This makes the recognition of driving styles more accurate. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 A flowchart of a driving style recognition method provided in an embodiment of this application;

[0058] Figure 2 A schematic diagram of a timing autoencoder provided in an embodiment of this application;

[0059] Figure 3 This is a schematic diagram of style tag model clustering provided in an embodiment of this application;

[0060] Figure 4 Multiple driver score distribution charts provided in embodiments of this application;

[0061] Figure 5 Radar image of a lane change scenario provided in an embodiment of this application;

[0062] Figure 6 Radar image of an acceleration driving scenario provided in the embodiments of this application;

[0063] Figure 7A radar image of a deceleration driving scenario provided in an embodiment of this application. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0065] See Figure 1 As shown in the figure, this application provides a method for driving style recognition, which includes the following steps:

[0066] 101: Acquire driving data and divide the driving data into scenarios to obtain a set of segments for each driving scenario, the set of segments including driving segment data.

[0067] Since the driving data collection may include more data not required by this application, this application selects multiple signal fields to compose the driving data. These include vehicle x-axis velocity, vehicle y-axis velocity, vehicle x-axis acceleration, vehicle y-axis acceleration, etc., where the x-axis and y-axis are the x-axis and y-axis in the whole vehicle coordinate system. The driving data is multi-dimensional time-series data synchronized with time stamps, and the sampling frequency can be set to once per second.

[0068] Because driving data is collected over a long period of time and at a high frequency, it may contain erroneous or missing values. Therefore, driving data can be cleaned.

[0069] The cleaning process includes: correcting outliers, filling in missing values, and removing or reducing noise and interference.

[0070] Specifically, outliers in driving data are identified and processed, and methods such as deletion, interpolation, or fitting are used to correct them, ensuring the integrity and accuracy of the driving data. For missing data, interpolation or other appropriate methods are used to complete it, ensuring the continuity of the driving data. Algorithms such as Kalman filtering or low-pass filtering are applied to remove or reduce noise and interference in the time series, ensuring the smoothness and continuity of the driving data, thereby providing high-quality data support for subsequent processing.

[0071] The outliers in this application include the following categories:

[0072] (1) For abnormal data caused by sensor reading errors or data transmission, this application adopts a combination of statistical analysis methods and visualization technology for identification. Specifically, by calculating statistical characteristics such as standard deviation and mean, an abnormal threshold is set, and values ​​that deviate from the abnormal threshold are marked as abnormal values. In this process, visualization tools such as anomaly detection charts are used to display extreme values ​​or outliers in the data for further processing.

[0073] (2) Another type of outlier is when the data remains constant over a long period of time, which usually indicates that the sensor has failed to accurately reflect the actual driving situation. For this type of signal, this application uses downsampling technology to process it, so as to reduce the interference of redundant data on subsequent analysis and avoid the analytical bias caused by it.

[0074] (3) In response to the abnormal data caused by frequent random fluctuations or periodic disturbances, this application uses a Kalman filter to smooth the multivariate time series data, thereby eliminating uncorrelated fluctuations and ensuring the consistency of the data in space and time.

[0075] Since drivers may exhibit different driving styles in different driving scenarios, for example, some aggressive drivers may find it difficult to increase their speed in congested driving environments and actually perform more conservatively or normally, the driving style in a single driving scenario is insufficient to fully represent the driver's overall style. Therefore, in step 101, this application divides the driving data into multiple segments of different driving scenarios, with the aim of achieving comprehensive identification and evaluation of the driver's style through comprehensive analysis of driving styles in different driving scenarios.

[0076] 102: Use a style tag model to process the fragment sets in each driving scenario to obtain style tags for each driving scenario.

[0077] Specifically, for each segment set in a driving scenario, data processing is performed separately to obtain the style tags corresponding to each driving scenario.

[0078] The data processing includes the following steps:

[0079] 201: Dimensionally reduce the driving segment data contained in the segment set under the driving scenario to extract the feature set corresponding to the segment set, the feature set including driving features corresponding to the driving segment data contained in the segment set.

[0080] In step 201, by reducing the dimensionality of each driving segment data, the driving features corresponding to the driving segment data can be extracted, and these driving features form a feature set.

[0081] 202: Cluster the driving features contained in the feature set to obtain the number of driving features in the feature set belonging to each driving style and the driving style weight.

[0082] In step 202, the driving features contained in the feature set are classified to obtain the number of driving features belonging to different driving styles, and at the same time, the corresponding driving style weights are obtained.

[0083] It is understandable that the number of driving styles mentioned above can be divided according to actual needs. For example, the driving styles include conservative, normal and aggressive styles.

[0084] 203: Based on the quantity and weight of each driving style, obtain the style label for the driving scenario corresponding to the feature set.

[0085] In step 203, the number and weight of each driving style are calculated to obtain the style label for the driving scenario.

[0086] In order to comprehensively evaluate the driver's driving style, this application introduces a driving style weight in step 102. The reason for introducing the driving style weight is that each driving style (such as conservative style, normal style and aggressive style) is actually a relative result of the driver in the same driving scenario. This relative result takes into account the specific driving environment faced by the driver in the specific driving scenario.

[0087] Step 102 clusters the driving features under the same driving scenario, and calculates the style label obtained by the quantity and weight of each driving style under the same driving scenario based on the clustering results. This style label takes into account the specific driving situation faced by the driver under the specific driving scenario, so as to quantify the specific driving environment and provide a more comprehensive and objective evaluation for subsequent driving style recognition.

[0088] The aforementioned style labels are more like soft labels, which are a comprehensive result of the driver's driving data over a period of time under different driving scenarios. They take into account clustering errors or the possibility that the driver may exhibit unstable driving styles, that is, different driving styles may be exhibited in the same driving scenario. This makes the identification of driving styles in this application more accurate.

[0089] 103: Based on the style labels of each driving scenario, obtain the driving style corresponding to the driving data.

[0090] In step 103, the style labels of the driver's driving data over a period of time under different driving scenarios are comprehensively considered to obtain a final result. This final result reflects the driver's driving style during this period, which helps to optimize the function of the advanced driver assistance system, enabling it to adaptively adjust according to the driver's behavior habits and promote the realization of a personalized driving experience.

[0091] In this application, driving scenarios can be divided according to actual needs. As an example, based on the driver's own control of the vehicle, driving scenarios can generally be divided into three categories, namely, driving scenarios include lane changing scenarios, acceleration driving scenarios, and deceleration driving scenarios.

[0092] The data used in this application primarily comes from real-world vehicle driving data, thus covering various driving scenarios. Considering the strong correlation between driving scenarios and driving styles, the same driver may exhibit different driving styles in different scenarios and at different stages. For example, a cautious driver in an acceleration scenario might display an aggressive driving style in a lane-changing scenario. Therefore, identifying driving styles across multiple scenarios is of great significance. One of the main contents of this application is to divide driving data into different scenario events according to the real-world significance of the data: acceleration scenarios, deceleration scenarios, and lane-changing scenarios. Through this division, this application can perform detailed style identification for different driving behavior patterns, providing a systematic method for comprehensively evaluating a driver's driving style.

[0093] The driving data is divided into scenarios to obtain a set of segments for each driving scenario, specifically including:

[0094] Data of a first time length is extracted from the driving data to obtain a set of segments in the lane change scenario. The data of the first time length satisfies the following condition: one of the sampling time points in the data of the first time length is the lane change time point. The first time length can be set according to actual needs. For example, as an example, data of 20 seconds before and after the lane change time point can be extracted as driving segment data. It can be understood that the lane change time point should be ensured to be in the middle of the first time length as much as possible, rather than at the two ends.

[0095] The driving data is extracted as driving segment data for a second time length to obtain a segment set under the acceleration driving scenario. The data for the second time length satisfies the following condition: in the data for the second time length, the speed at the later sampling time point is greater than the speed at the earlier sampling time point in two adjacent sampling time points. The second time length can be set according to actual needs. For example, the second time length is 70s.

[0096] Data of a third time length is extracted from the driving data to obtain a set of segments for deceleration scenarios. The data of the third time length satisfies the following condition: in two adjacent sampling time points, the speed at the later sampling time point is less than the speed at the earlier sampling time point. The third time length can be set according to actual needs; for example, it can be 70 seconds.

[0097] In step 102 above, the style tag model is used to process the data sets of segments under each driving scenario to obtain the style tags for each driving scenario, specifically including:

[0098] The driving segment data contained in the segment set of each driving scenario is used as input data and input into the style label model corresponding to each driving scenario to output the style label corresponding to each driving scenario.

[0099] In this application, a corresponding style label model is trained for each driving scenario. For example, if there are three driving scenarios, three style label models are trained. The driving segment data contained in the segment set corresponding to each driving scenario is input into the respective style label model for data processing, thereby outputting the style label corresponding to each driving scenario.

[0100] The style tag model includes a temporal autoencoder, a deep clustering model, and a loss function calculation module.

[0101] The temporal autoencoder is used to: reduce the dimensionality of the driving segment data contained in the segment set under the driving scenario, so as to extract the feature set corresponding to the segment set, wherein the feature set includes driving features corresponding to the driving segment data contained in the segment set;

[0102] The deep clustering model is used to: cluster the driving features contained in the feature set to obtain the number of driving features in the feature set belonging to each driving style and the driving style weight; and obtain the style label of the driving scenario corresponding to the feature set based on the number of each driving style and the driving style weight.

[0103] The loss function calculation module is used to calculate the loss of the style tag model in order to measure the accuracy of the style tag model.

[0104] The feature extraction method based on temporal autoencoders aims to extract key information from driving data. The temporal autoencoder mainly includes an encoder for extracting driving features and a decoder for autoregression of these features. Unlike traditional neural networks that only perform forward propagation, the temporal autoencoder learns to map the input data to a latent space for representation, and then maps this representation back to the original space, achieving data compression and reconstruction. In this process, the encoder is responsible for converting the input data into an encoded representation in a hidden layer, while the decoder restores this encoded representation to reconstructed data that closely approximates the original input. The training objective of the temporal autoencoder is to minimize the error between the input data and the reconstructed data, typically using mean squared error as the loss function to measure this difference. Furthermore, the activation function introduces nonlinearity into the network, enabling the temporal autoencoder to capture and learn complex patterns and structures in the data. Through this feature extraction method, this application can effectively learn useful features from driving data that contribute to driving style recognition.

[0105] Specifically, time-series autoencoders can take various forms, including but not limited to those constructed from Long Short-Term Memory (LSTM) networks or Gated Recurrent Units (GRUs). While LSTM networks effectively address long-term dependencies, they exhibit long training times and a tendency to overfit when processing time-series data. Therefore, to better extract information from driving style time series data and address the inability of traditional LSTM networks to perform parallel computation, such as... Figure 2 As shown, as a preferred embodiment, this application selects a hybrid network architecture that integrates a Temporal Convolutional Network (TCN) and a Long Short-Term Memory (LSTM) network. This hybrid architecture is designed to combine the efficiency of TCN in processing time-series data with the advantages of LSTM in capturing long-term data dependencies, thereby providing a more powerful and adaptable autoencoder.

[0106] An Auto-Encoder based on TCN and LSTM networks, comprising an encoder and a decoder.

[0107] The encoder includes a first layer based on a temporal convolutional network (TCN) and the activation function ReLU (i.e., Figure 2 TCN+ReLU in the middle), and the second layer based on MaPool (i.e. Figure 2 MaPool in the middle), the third layer based on bidirectional LSTM (i.e. Figure 2 BiLSTM in the middle) and the fourth layer based on bidirectional LSTM (i.e. Figure 2 BiLSTM in (the part).

[0108] Temporal Convolutional Networks (TCNs) are used to process temporal data and can capture dependencies. The ReLU (Rectified Linear Unit) activation function is used to increase the network's non-linearity, helping it learn more complex patterns. During data processing, preliminary feature extraction is performed on the input driving data. Convolutional operations capture local features in the data, and the ReLU activation function introduces non-linearity.

[0109] Max pooling is used to downsample data, reducing its dimensionality while preserving important features. During data processing, downsampling is performed on data that has undergone TCN+ReLU processing to reduce the amount of data and decrease the complexity of subsequent calculations.

[0110] The two-layer bidirectional long short-term memory (BiLSTM) network can simultaneously consider information from the past and future, demonstrating excellent handling capabilities for long-term dependencies in time-series data. During data processing, the first BiLSTM layer receives the output of the MaxPool layer and further processes the data to extract higher-level features. The second BiLSTM layer continues to process the data based on the first layer, mapping the data to a lower-dimensional latent vector space while preserving as much useful information as possible.

[0111] The decoder includes a first layer based on upsampling (i.e. Figure 2 UpSample in the middle) and the second layer based on deconvolution (i.e. Figure 2 (Deconvolution in the middle).

[0112] Upsampling is used to upsample the data, increasing its dimensionality and restoring its original size. During data processing, the latent vector from the last layer (the second BiLSTM layer) of the Encoder is received and upsampled to prepare for subsequent deconvolution operations.

[0113] Deconvolution is used to reconstruct upsampled data into a form similar to the original input data. During data processing, deconvolution is performed on the upsampled data in an attempt to reconstruct the original input data. By continuously adjusting the network parameters, the reconstructed data is made as close as possible to the original input data.

[0114] By designing the aforementioned temporal autoencoder, the dimensionality of the driving segment data contained in the segment set under the driving scenario can be reduced to extract the feature set corresponding to the segment set. Specifically, the temporal autoencoder corresponding to the driving scenario is used to sequentially reduce the dimensionality of the driving segment data contained in the segment set under the driving scenario to extract the driving features corresponding to each driving segment data. The driving features corresponding to each driving segment data together constitute the feature set corresponding to the segment set under the driving scenario.

[0115] Through the processing of the Encoder and Decoder, the Temporal Autoencoder (TAE) network ultimately abstracts the input driving data into a low-dimensional representation (latent vector). This low-dimensional representation retains the key information and features of the original data while undergoing dimensionality reduction and optimization. It can be utilized by subsequent deep clustering models and loss function calculation modules, providing strong support for accurately identifying driving styles.

[0116] After driving features are extracted by a temporal autoencoder, these features are typically used to perform driving style segmentation, grouping data using algorithms such as K-means and hierarchical clustering. Since clustering algorithms are inherently unsupervised learning methods, they do not rely on predefined label information. The main goal of these algorithms is to explore and discover the internal structure and patterns of the data. However, evaluating the accuracy of clustering results becomes challenging in the absence of true labels. The key issue is how to quickly and accurately extract features and effectively evaluate clustering results in a label-free manner. To address this problem, this application introduces internal evaluation metrics, such as the silhouette coefficient, to quantitatively measure the quality of clustering results, thus providing an effective solution to the label-free clustering problem.

[0117] This application does not impose any restrictions on the clustering algorithm, but as a preferred option, the deep clustering model described in this application uses the K-means++ algorithm to determine the initial cluster centers and the MiniBatchKmeans algorithm to perform clustering.

[0118] Compared to the conventional K-means algorithm, the core advantage of the MiniBatch K-means algorithm lies in its choice of processing a subset (i.e., a mini-batch) of data instead of using the entire dataset in each iteration. This approach significantly improves computational efficiency and reduces computational resource consumption when dealing with large datasets. By processing data in batches, MiniBatch K-means also helps alleviate memory pressure, enabling the algorithm to handle large datasets that exceed memory capacity limits. This is particularly important in practical applications, as the amount of data generated by modern intelligent connected vehicles is increasingly massive, and traditional clustering algorithms may encounter performance bottlenecks when processing this data.

[0119] Specifically, driving features are extracted using a temporal autoencoder to obtain a feature set. These driving features are then divided into k clusters using the K-means++ algorithm. The cluster centers are then updated using the MiniBatchKmeans algorithm, resulting in k initial cluster centers, representing k driving styles. For example, in this application, there are three driving styles, hence three initial cluster centers. During cluster training, to more accurately assign samples to their corresponding clusters, the distance from each sample to the cluster center is calculated. Then, the probability of a sample belonging to a cluster center is calculated using the Student's distribution. The closer a sample is to a cluster center, the higher its probability of belonging to that cluster, similar to the idea of ​​soft labeling.

[0120] Regarding the calculation of the loss function, this application uses joint optimization of two loss functions. The first loss function is the mean-square error (MSE) between the encoder and decoder of the temporal autoencoder. This ensures that the low-dimensional sequence after the encoder can well represent the original data. The second loss function is provided by the KL divergence of the clustering algorithm used in the deep clustering model, such as... Figure 3 As shown, minimizing the clustering loss allows the low-dimensional sequence to be better assigned to the k clusters by the clustering model. Compared with traditional methods that only minimize the autoencoder loss or only minimize the clustering loss, the joint optimization of the autoencoder and clustering loss in this application has better accuracy.

[0121] Understandably, this application requires model training before using the aforementioned style tag model for recognition.

[0122] Specifically, before using the style tagging model to process the fragment sets under each driving scenario, the method further includes the following steps:

[0123] 301: Construct a sample dataset, which includes several driving data sets;

[0124] 302: Divide the driving data into scenarios to obtain a set of segments for each driving scenario;

[0125] 303: Construct initial style tag models for various driving scenarios;

[0126] 304: Using the fragment sets of various driving scenarios, the initial style label model for each driving scenario is iteratively trained to obtain the style label model for each driving scenario.

[0127] Understandably, the methods of processing driving data during training and recognition are basically the same, so I won't go into details here.

[0128] In step 202 above, the driving features contained in the feature set are clustered to obtain the number of driving features in the feature set belonging to each driving style and the driving style weights, specifically including the following:

[0129] 401: Cluster the driving features contained in the feature set to obtain multiple sub-feature sets, and the number of driving features divided from the feature set in each sub-feature set.

[0130] In step 401, the driving features contained in the feature set are clustered using a deep clustering model of the style tag model, thereby dividing the feature set into multiple sub-feature sets. Each sub-feature set corresponds to a driving style, and each sub-feature set includes a certain number of driving features.

[0131] The quantity can refer to the proportion of quantities. For example, driving data is divided into three driving scenarios: lane changing, acceleration, and deceleration. The lane changing scenario segment set includes 100 driving segment data. After extraction, the resulting sub-feature set includes 100 driving features. Among these 100 driving features, 80 are clustered into style A, 15 into style B, and 5 into style C. Therefore, in the lane changing scenario, the proportion of style A is 80%, style B is 15%, and style C is 5%.

[0132] Styles A, B, and C are all different from each other, and each represents one of the following styles: conservative, ordinary, and radical.

[0133] The quantity can also be a percentage of driving time. In this case, when extracting driving segment data, the duration of driving segment data under various driving scenarios can be the same or different. For example, if the duration is the same, taking the above-mentioned lane-changing scenario segment set as an example with 100 driving segment data, 80 are clustered into style A, 15 into style B, and 5 into style C. In the calculation, both the numerator and denominator are multiplied by the duration of the driving segment data. Therefore, in the final lane-changing scenario, the driving time percentage of style A is 80%, the driving time percentage of style B is 15%, and the driving time percentage of style C is 5%.

[0134] 402: Based on the driving features included in each of the sub-feature sets, determine the driving style to which each of the sub-feature sets belongs, and the corresponding driving style weight.

[0135] The methods for determining Style A, Style B, and Style C for different driving scenarios are as follows:

[0136] When the feature set is the feature set corresponding to the segment set in the lane change scenario, based on the driving features included in each of the sub-feature sets, the driving style to which each sub-feature set belongs and the corresponding driving style weight are determined, specifically including the following steps:

[0137] 501: Based on the driving segment data corresponding to the driving features included in each of the sub-feature sets, obtain the preset variable values ​​corresponding to each of the sub-feature sets.

[0138] 502: Based on the magnitude of the preset variable values ​​corresponding to each of the sub-feature sets, obtain the driving style to which each of the sub-feature sets belongs, and the corresponding driving style weight.

[0139] For example, taking the above-mentioned lane-changing scenario segment set including 100 driving segment data as an example, the preset variable value of the driving segment data corresponding to the 80 driving features of style A is 'a', the preset variable value of the driving segment data corresponding to the 15 driving features of style B is 'b', and the preset variable value of the driving segment data corresponding to the 15 driving features of style C is 'c'. If the preset variable value is based on speed or acceleration, a > b > c. Since the greater the speed or acceleration, the faster the lane change, the more aggressive it is, style A is an aggressive style, style B is a normal style, and style C is a conservative style.

[0140] As for the driving style weights, for example, the weight of aggressive style is x, the weight of normal style is y, and the weight of conservative style is z, and x, y, and z have been assigned values ​​in advance.

[0141] Similarly, when the feature set is the feature set corresponding to the segment set in the acceleration driving scenario, based on the driving features included in each of the sub-feature sets, the driving style to which each sub-feature set belongs and the corresponding driving style weight are determined, specifically including the following steps:

[0142] 601: Based on the driving segment data corresponding to the driving features included in each of the sub-feature sets, obtain the preset variable values ​​corresponding to each of the sub-feature sets.

[0143] 602: Based on the magnitude of the preset variable values ​​corresponding to each of the sub-feature sets, obtain the driving style to which each of the sub-feature sets belongs, and the corresponding driving style weight.

[0144] Similarly, when the feature set is the feature set corresponding to the segment set in the deceleration driving scenario, based on the driving features included in each of the sub-feature sets, the driving style to which each sub-feature set belongs and the corresponding driving style weight are determined, specifically including the following steps:

[0145] 701: Based on the driving segment data corresponding to the driving features included in each of the sub-feature sets, obtain the preset variable values ​​corresponding to each of the sub-feature sets.

[0146] 702: Based on the magnitude of the preset variable values ​​corresponding to each of the sub-feature sets, obtain the driving style to which each of the sub-feature sets belongs, and the corresponding driving style weight.

[0147] In step 203, based on the quantity and weight of each driving style, the style label for the driving scenario corresponding to the feature set is obtained, specifically including the following steps:

[0148] 801: Multiply the quantity of each driving style by the corresponding driving style weight to obtain the product value of each driving style.

[0149] For example, taking the above lane-changing scenario segment set as an example, which includes 100 driving segment data, in the lane-changing scenario, style A is aggressive style, style B is normal style, and style C is conservative style. Then the product value of aggressive style is 80%x, the product value of normal style is 15%y, and the product value of conservative style is 5%z.

[0150] 802: Sum the product values ​​to obtain the style label for the driving scenario.

[0151] For example, taking the above lane change scenario as an example, which includes 100 driving segment data, the style label for the lane change scenario is 80%x+15%y+5%z.

[0152] Similarly, driving data can be divided into style labels for accelerating driving scenarios and style labels for decelerating driving scenarios.

[0153] It is understandable that the above-mentioned preset variable values ​​can be selected from maximum speed, mean speed, standard deviation of speed, maximum acceleration, mean acceleration, standard deviation of acceleration, maximum lateral speed, mean lateral speed, and standard deviation of lateral speed.

[0154] Referring to Table 2 below, the speed standard deviation can easily distinguish styles, so it is the preferred choice. Although it seems that other indicators, such as the acceleration standard deviation under lane change, cannot be distinguished when other indicators are used as preset variable values ​​in Table 2, it is understandable that Table 2 cannot be compared due to an unreasonable selection of the number of decimal places (three), thus leading to the inability to distinguish. In fact, the distinction can be made by setting an appropriate number of decimal places.

[0155] It is understandable that the maximum speed, maximum acceleration, and maximum lateral speed are all average values ​​calculated from all driving segment data included in the same style. For example, the maximum speed is the average value calculated from the maximum speed of all driving segment data included in the same style. If a certain style contains 80 driving segment data, and each driving segment data has a maximum speed, then the average of the 80 maximum speeds is calculated to obtain the preset variable value.

[0156] In step 103, based on the style tags for each driving scenario, the driving style corresponding to the driving data is obtained, specifically including:

[0157] 901: Sum the style tags under each driving scenario to obtain the driving score of the driving data.

[0158] 902: Based on the scoring range of each driving style and the driving score, determine the driving style corresponding to the driving data.

[0159] By comparing the driving score of the driving data with the score range of each driving style, it can be determined which score range the driving score falls into, thereby determining the driving style corresponding to the identified driving data, and thus determining the driver's driving style.

[0160] like Figure 4 As shown in the figure, this illustration presents examples of multiple driver score distributions, further demonstrating the effectiveness of this application in practical applications.

[0161] To verify the accuracy of the method, this application first statistically analyzed the cluster distribution of drivers in various driving scenarios. The category distribution for acceleration, deceleration, and lane changing scenarios is shown in Table 1 below.

[0162] Table 1

[0163]

[0164] The distribution of driver driving style classification results shows a significant difference in the number of drivers in different categories. However, since the data is unlabeled, the clustering results cannot represent their true meaning. Therefore, to further illustrate the different driving styles of drivers in different driving scenarios, this application uses time-varying driving data, spectral entropy, and multiple perspectives such as speed and acceleration to measure the style of each category, and uses this as the basis for calculating the comprehensive category of driving style in multiple scenarios. First, this application calculated the maximum and average values ​​of acceleration, speed, and other signals in the three categories, and the calculation results are shown in Table 2 below.

[0165] Table 2

[0166]

[0167] To more clearly compare the results, the data in the table above are compared using radar charts for different scenarios. The radar charts for each preset variable value under the lane changing, acceleration, and deceleration scenarios are shown below. Figure 5 , Figure 6 and Figure 7 As shown in the radar chart for lane-changing scenarios, the SpeedStd for Category 2 is significantly higher than the other two categories, indicating that Category 2 lane-changing scenarios represent aggressive drivers. Meanwhile, the SpeedStd for Category 0 is the smallest among all categories, indicating that Category 0 represents conservative drivers. The analysis methods for the other two scenarios are similar.

[0168] The above analysis shows that in lane-changing scenarios, categories 0, 1, and 2 represent conservative, normal, and aggressive driving styles, respectively; in acceleration scenarios, categories 0, 1, and 2 represent normal, conservative, and aggressive driving styles, respectively; and in deceleration scenarios, categories 0, 1, and 2 represent conservative, normal, and aggressive driving styles, respectively. This implies that the perception and thresholds for aggressive, normal, and conservative driving differ across different environmental categories. In other words, the same driver may exhibit different driving styles in different driving scenarios, which aligns with current understanding of driving styles. Therefore, it is necessary to segment the data into different driving scenarios and identify driving styles.

[0169] Since clustering tasks themselves lack true labels, it's impossible to compare results using real labels. To compare the proposed method's performance in driving style clustering tasks compared to traditional machine learning methods, internal evaluation metrics such as the Sihouette Coefficient (SC), Davies-Bouldin (DB) score, and Calinski-Harabasz (CH) score were used to compare the algorithm's clustering effectiveness. Table 3 shows a comparison of the proposed method and the traditional K-means method on these three evaluation metrics. Therefore, the proposed method outperforms traditional machine learning clustering methods, represented by K-means, on the SC, DB, and CH evaluation metrics.

[0170] Table 3

[0171] K-means 0.542 0.557 27240.315 PCA+K-means 0.544 0.556 27232.866 This application method 0.554 0.549 27987.227

[0172] Corresponding to the above-described method for recognizing driving styles, this application also provides a system for recognizing driving styles, which includes an acquisition module, a style tag model, and a driving style recognition module.

[0173] The acquisition module is used to acquire driving data and divide the driving data into scenarios to obtain a set of segments under each driving scenario, the set of segments including driving segment data.

[0174] The style tag model is used to process data sets of segments under various driving scenarios to obtain style tags for each driving scenario. The data processing includes: reducing the dimensionality of the driving segment data contained in the segment set under the driving scenario to extract the feature set corresponding to the segment set, the feature set including driving features corresponding to the driving segment data contained in the segment set; clustering the driving features contained in the feature set to obtain the number of driving features contained in the feature set belonging to each driving style and the driving style weight; and obtaining the style tag for the driving scenario corresponding to the feature set based on the number of driving styles and the driving style weight.

[0175] The driving style recognition module is used to obtain the driving style corresponding to the driving data based on the style labels under each driving scenario.

[0176] Corresponding to the above-described method for recognizing driving styles, this application also provides a storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the above embodiments. It should be noted that the storage medium in this application embodiment can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0177] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0178] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0179] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as "C" or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0180] Corresponding to the above-described method for recognizing driving styles, this application also provides an electronic device, including a memory and a processor. The memory stores a computer program that runs on the processor, and the processor executes the computer program to implement the steps of the above embodiments.

[0181] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for driving style recognition, characterized in that, It includes: Acquire driving data and divide the driving data into scenarios to obtain a set of segments for each driving scenario, the set of segments including driving segment data; The style tag model is used to process the fragment sets of each driving scenario to obtain the style tags for each driving scenario. The data processing includes: dimensionality reduction of the driving segment data contained in the segment set under the driving scenario to extract the feature set corresponding to the segment set, the feature set including driving features corresponding to the driving segment data contained in the segment set; clustering the driving features contained in the feature set to obtain the number of driving features contained in the feature set belonging to each driving style and the driving style weight; and obtaining the style label under the driving scenario corresponding to the feature set based on the number of each driving style and the driving style weight. Based on the style labels for each driving scenario, the driving style corresponding to the driving data is obtained.

2. The driving style recognition method as described in claim 1, characterized in that: The driving scenarios include lane changing scenarios, acceleration scenarios, and deceleration scenarios.

3. The driving style recognition method as described in claim 2, characterized in that: The driving data is divided into scenarios to obtain a set of segments for each driving scenario, specifically including: Data of a first time length is extracted from the driving data to obtain a set of segments in the lane change scenario. The data of the first time length satisfies the following condition: one of the sampling time points in the data of the first time length is the lane change time point. The driving data is extracted as driving segment data for a second time length to obtain a segment set under the acceleration driving scenario. The second time length data satisfies the following condition: in the second time length data, in two adjacent sampling time points, the speed of the later sampling time point is greater than the speed of the earlier sampling time point. The driving data is extracted into a third time length as driving segment data to obtain a segment set under the deceleration driving scenario. The third time length data satisfies the following condition: in the third time length data, in two adjacent sampling time points, the speed of the later sampling time point is less than the speed of the earlier sampling time point.

4. The driving style recognition method as described in claim 1, characterized in that: The style tagging model is used to process the fragment sets under various driving scenarios to obtain style tags for each driving scenario, specifically including: The driving segment data contained in the segment set of each driving scenario is used as input data and input into the style label model corresponding to each driving scenario to output the style label corresponding to each driving scenario.

5. The driving style recognition method as described in claim 1, characterized in that: The dimensionality of the driving segment data contained in the segment set under the driving scenario is reduced to extract the feature set corresponding to the segment set, specifically including: The driving segment data contained in the segment set under the driving scenario is reduced in dimensionality by using the temporal autoencoder corresponding to the driving scenario to extract the driving features corresponding to each driving segment data and obtain the feature set corresponding to the segment set under the driving scenario.

6. The driving style recognition method as described in claim 1, characterized in that: The quantity information includes either the percentage of quantity or the percentage of driving time.

7. The driving style recognition method as described in claim 1, characterized in that: Clustering the driving features contained in the feature set to obtain the number of driving features in the feature set belonging to each driving style and the driving style weights, specifically including: The driving features contained in the feature set are clustered to obtain multiple sub-feature sets, and the number of driving features divided from the feature set is included in each sub-feature set. Based on the driving features included in each of the sub-feature sets, the driving style to which each sub-feature set belongs, and the corresponding driving style weight, are determined.

8. The driving style recognition method as described in claim 7, characterized in that: Based on the driving features included in each of the sub-feature sets, the driving style to which each sub-feature set belongs, and the corresponding driving style weight, are determined, specifically including: Based on the driving segment data corresponding to the driving features included in each of the sub-feature sets, obtain the preset variable values ​​corresponding to each of the sub-feature sets; Based on the magnitude of the preset variable values ​​corresponding to each of the sub-feature sets, the driving style to which each sub-feature set belongs, and the corresponding driving style weight, are obtained.

9. The driving style recognition method as described in claim 8, characterized in that: The preset variable values ​​are obtained based on velocity or acceleration.

10. The driving style recognition method as described in claim 9, characterized in that: The preset variable values ​​are selected from maximum speed, mean speed, standard deviation of speed, maximum acceleration, mean acceleration, standard deviation of acceleration, maximum lateral speed, mean lateral speed, and standard deviation of lateral speed.

11. The driving style recognition method as described in claim 1, characterized in that: Based on the quantity and weight of each driving style, style labels for the driving scenarios corresponding to the feature set are obtained, specifically including: Multiply the quantity of each driving style by the corresponding driving style weight to obtain the product value of each driving style; The product values ​​are summed to obtain the style label for the driving scenario.

12. The driving style recognition method as described in claim 1, characterized in that: Based on the style tags for each driving scenario, the driving style corresponding to the driving data is obtained, specifically including: The style tags under each driving scenario are summed to obtain the driving score of the driving data. Based on the scoring range of each driving style and the driving score, the driving style corresponding to the driving data is determined.

13. The driving style recognition method as described in claim 1, characterized in that: The driving styles include conservative, normal, and aggressive.

14. The driving style recognition method as described in claim 1, characterized in that: Before using the style tag model to process the fragment sets under various driving scenarios, the method further includes: Construct a sample dataset, which includes several driving data sets; The driving data is divided into scenarios to obtain a set of segments for each driving scenario; Construct initial style label models for various driving scenarios; Using a set of clips from various driving scenarios, the initial style label model for each driving scenario is iteratively trained to obtain the style label model for each driving scenario.

15. The driving style recognition method as described in claim 14, characterized in that: The style tag model includes a temporal autoencoder and a deep clustering model; The temporal autoencoder is used to: reduce the dimensionality of the driving segment data contained in the segment set under the driving scenario, so as to extract the feature set corresponding to the segment set, wherein the feature set includes driving features corresponding to the driving segment data contained in the segment set; The deep clustering model is used to: cluster the driving features contained in the feature set to obtain the number of driving features in the feature set belonging to each driving style and the driving style weight; and obtain the style label of the driving scenario corresponding to the feature set based on the number of each driving style and the driving style weight.

16. The driving style recognition method as described in claim 15, characterized in that: The temporal autoencoder includes an encoder for extracting driving features and a decoder for autoregression of driving features; The encoder includes a first layer based on a temporal convolutional network (TCN) and the activation function ReLU, a second layer based on MaPool, a third layer based on bidirectional LSTM, and a fourth layer based on bidirectional LSTM. The decoder consists of a first layer based on upsampling and a second layer based on deconvolution.

17. The driving style recognition method as described in claim 15, characterized in that: The deep clustering model uses the K-means++ algorithm to determine the initial cluster centers and the MiniBatchKmeans algorithm to perform clustering.

18. A driving style recognition system, characterized in that, It includes: An acquisition module is used to acquire driving data and divide the driving data into scenarios to obtain a set of segments under each driving scenario, wherein the set of segments includes driving segment data. A style tag model is used to process data sets of segments in various driving scenarios to obtain style tags for each driving scenario. The data processing includes: dimensionality reduction of the driving segment data contained in the segment set under the driving scenario to extract the feature set corresponding to the segment set, the feature set including driving features corresponding to the driving segment data contained in the segment set; clustering the driving features contained in the feature set to obtain the number of driving features contained in the feature set belonging to each driving style and the driving style weight; and obtaining the style label under the driving scenario corresponding to the feature set based on the number of each driving style and the driving style weight. The driving style recognition module is used to obtain the driving style corresponding to the driving data based on the style labels under each driving scenario.

19. A storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the driving style recognition method according to any one of claims 1 to 17.

20. An electronic device comprising a memory and a processor, wherein the memory stores a computer program that runs on the processor, characterized in that: When the processor executes the computer program, it implements the driving style recognition method according to any one of claims 1 to 17.

Images