Method for identifying driving behavior pattern before collision of unsupervised auxiliary driving vehicle

Abstract

The invention discloses an unsupervised auxiliary driving vehicle pre-collision driving behavior pattern recognition method, which comprises the following steps: extracting original data before collision, and obtaining key driving parameters; the method comprises the following steps: segmenting continuous driving behavior data of an auxiliary driving vehicle before collision through a hierarchical Dirichlet process hidden semi-Markov model, and identifying different driving operation characteristics; and adopting a clustering analysis technology to classify the driving behavior modes, and extracting different driving modes. According to the method, the limitation of a specific driving scene and the complexity that a supervision model depends on annotation can be made up, and the applicability and usability of the method are improved on the premise that the behavior pattern recognition precision is guaranteed; technical support is provided for improving the safety of an auxiliary driving system and accurately recognizing the risk driving state, traffic safety technical innovation and application expansion are promoted, and construction of safer and more efficient traffic ecology is assisted.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent transportation and driver safety assessment, and particularly to an unsupervised method for recognizing pre-collision driving behavior patterns of assisted driving vehicles. Background Art

[0002] At present, there is a certain scale of assisted driving vehicles in China, and their continuously increasing accident rate has also become a pain point in the governance of modern urban traffic. The road transport safety authorities already have a mature technical system for analyzing the operation characteristics in traditional vehicle traffic accidents. However, due to problems such as the operation of the assisted driving system and the interaction between humans and vehicles, the operation characteristics of assisted driving vehicles are significantly different from those of traditional vehicles.

[0003] Various methods for identifying vehicle operation characteristics have been proposed. Most of them adopt supervised methods to develop driver behavior models in specific driving scenarios, such as following a leading vehicle at intersections or on highways. However, due to the complexity of accident scenarios and the diversity of the operation data dimensions of assisted driving vehicles, it is quite difficult to prepare complete driving scenario annotations.

[0004] Therefore, it is necessary to provide a new method for identifying the risky driving state of drivers to make up for the deficiencies of traditional driving behavior pattern recognition methods. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide an unsupervised method for recognizing pre-collision driving behavior patterns of assisted driving vehicles, which can make up for the limitations of specific driving scenarios and the complexity of supervised model dependence on annotations, and improve the applicability and usability of the method on the premise of ensuring the accuracy of behavior pattern recognition; provide technical support for improving the safety and accurate recognition of risky driving states of assisted driving systems, promote the innovation and application expansion of traffic safety technologies, and help build a safer and more efficient traffic ecosystem.

[0006] The present invention adopts the following technical solutions to solve the technical problems:

[0007] An unsupervised method for recognizing pre-collision driving behavior patterns of assisted driving vehicles, comprising the following steps:

[0008] S1, extracting the original data before collision and obtaining key driving parameters;

[0009] S2, segmenting the continuous driving behavior data of the assisted driving vehicle before collision through a hierarchical Dirichlet process hidden semi-Markov model to identify different driving operation characteristics;

[0010] S3, after completing the segmentation of driving behavior, using a Gaussian mixture model - latent Dirichlet allocation model to cluster the segmented data to identify typical pre-collision driving behavior patterns.

[0011] Further, in step S1, the method for extracting the original data before collision and obtaining the key driving parameters is as follows:

[0012] Extract natural driving data from CAN, T - BOX or in - vehicle global positioning system, and process outliers and missing values; based on the assisted - driving vehicle's control behavior and vehicle operation segments, extract key driving parameters.

[0013] Further, the key driving parameters include speed, acceleration, steering wheel angle, brake pedal and throttle pedal parameters.

[0014] Further, in step S2, the method for segmenting the continuous driving behavior data of the assisted - driving vehicle before collision by the hierarchical Dirichlet process hidden semi - Markov model includes:

[0015] 1) Model initialization

[0016] The hierarchical Dirichlet process hidden semi - Markov model uses the Dirichlet process as a non - parametric prior, the observation distribution adopts a Gaussian distribution, the initial mean is set to 0, and the covariance matrix is the identity matrix to ensure the standardized processing of multi - dimensional driving behavior data; the duration of the hidden state is modeled by a Poisson distribution, and the expected value is set to 6, reflecting the typical duration of each stage of the behavior before collision;

[0017] 2) Parameter inference

[0018] Use the Gibbs sampling method to iteratively infer the model parameters; this method updates the hidden state sequence according to the posterior distribution in each iteration, gradually optimizing the state transition probability matrix and observation distribution parameters until the model converges; through iteration, the model can automatically identify the changes in the hidden state in the data and achieve accurate segmentation of driving behavior;

[0019] 3) Hidden state extraction and segmentation

[0020] Based on the inferred hidden state sequence, divide the original driving behavior data into several segments, and each segment corresponds to a specific driving behavior pattern.

[0021] Further, to ensure the effectiveness of the segmentation result, eliminate short time periods with a duration less than 0.2 seconds to exclude meaningless noise.

[0022] Further, in step S3, the method for clustering the segmented data using the Gaussian mixture model - latent Dirichlet allocation model to identify typical pre - collision driving behavior patterns includes:

[0023] 1) Data discretization

[0024] The continuous behavior vectors are discretized into "driving units" using the Gaussian mixture model; the GMM assumes that the data comes from a mixture of multiple Gaussian distributions, and the parameters are optimized through the expectation-maximization algorithm; in the E-step, the probability of each data point belonging to each Gaussian component is calculated, and in the M-step, the model parameters are updated. Finally, each behavior segment is mapped to a discrete class label;

[0025] 2) Topic modeling

[0026] The discretized data is input into the LDA model for topic modeling. Each segmented part is regarded as a "document", and the discrete "driving units" are used as "words". The LDA model is used to mine the potential driving behavior "topics"; the LDA model assumes that each behavior segment may contain multiple behavior patterns, and Gibbs sampling or variational inference is used to deduce the model parameters to determine the probability distribution of the behavior patterns in each segmented part;

[0027] 3) Behavior pattern recognition

[0028] Through clustering, various typical pre-collision driving behavior patterns are identified.

[0029] Furthermore, the pre-collision driving behavior patterns include steady cruising, accelerating to avoid danger, emergency braking, and direction adjustment.

[0030] An unsupervised method for recognizing pre-collision driving behavior patterns of assisted driving vehicles provided by the present invention has the following beneficial effects:

[0031] (1) Automatic segmentation and recognition of driving behavior based on non-parametric Bayesian models:

[0032] The present invention uses the hierarchical Dirichlet process hidden semi-Markov model (HDP-HSMM) to achieve automatic segmentation and dynamic modeling of pre-collision driving behavior of assisted driving vehicles. Different from traditional models that rely on a fixed number of states, the present invention uses non-parametric Bayesian methods and does not require presetting the number of states. It can automatically infer the optimal number of hidden states according to data characteristics and fully adapt to complex and changeable driving behavior scenarios. The model also introduces a semi-Markov process to accurately model the state duration and capture the temporal characteristics of driving behavior. While maintaining high computational performance, this method can more finely depict the changes in pre-collision driving behavior, effectively identify key behavior segments such as steady driving, rapid acceleration, and emergency braking, and provide a reliable technical means for driving behavior analysis and risk prediction. Whether it is the fine-grained analysis of a single driving behavior or the pattern extraction of large-scale behavior data, the present invention shows strong adaptability and stability.

[0033] (2) Extraction of typical risk driving patterns based on behavior topic modeling:

[0034] The present invention realizes the topic clustering and pattern recognition of pre - collision driving behaviors based on segmented behavior data through the Gaussian Mixture Model - Latent Dirichlet Allocation model (GMM - LDA). Compared with traditional methods based on simple clustering or rule - based classification, the present invention can deeply mine the implicit structure in driving behaviors and accurately capture the behavioral differences in different driving scenarios. The behavioral patterns extracted by this method not only improve the recognition accuracy of risky driving states, but also provide behavioral prediction and decision - making support for traffic management and assisted driving systems, helping to improve the intelligent level of traffic safety management. This technology has wide application value in real - time risk warning, driving behavior research, and the development of personalized assisted driving systems, significantly enhancing the scientific nature and effectiveness of driving safety management. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a flowchart of the method of the present invention;

[0036] Figure 2 is an example diagram of time - series segmentation of the present invention;

[0037] Figure 3 is an example diagram of the GMM - LDA clustering result of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0039] The present invention proposes an unsupervised method to extract and identify pre - collision behavior patterns, which combines multivariate time - series segmentation and clustering to explore the combined behavior patterns in the pre - collision process. This method has high robustness, adaptability, and ease of use to make up for the limitations of specific driving scenarios and the complexity of supervised models relying on annotations. The present invention can be applied to accident investigation and ADAS function optimization, providing strong support for traffic safety guarantee and meeting the actual needs of the traffic field for accurately identifying driving behavior patterns and improving traffic safety.

[0040] REFERENCE Figure 1, the present invention provides an unsupervised method for recognizing driving behavior patterns before collision of an assisted driving vehicle, and conducts relevant research through the following three stages: First, extract the original data before collision and obtain key driving parameters. Second, use the time series segmentation method to divide the driving behavior into stages and identify different driving operation characteristics. Finally, use clustering analysis technology to classify the driving behavior patterns and extract different driving patterns. The method includes the following steps:

[0041] S1, extract the original data before collision and obtain key driving parameters;

[0042] The method for extracting the original data before collision and obtaining key driving parameters is as follows: Extract natural driving data from CAN, T - BOX or in - vehicle global positioning system, and process outliers and missing values; Based on the assisted driving vehicle control behavior and vehicle operation segments, extract key driving parameters. The key driving parameters include speed, acceleration, steering wheel angle, brake pedal and throttle pedal parameters, etc.

[0043] S2, segment the continuous driving behavior data of the assisted driving vehicle before collision through a hierarchical Dirichlet process hidden semi - Markov model, and identify different driving operation characteristics;

[0044] The method for segmenting the continuous driving behavior data of the assisted driving vehicle before collision through a hierarchical Dirichlet process hidden semi - Markov model includes:

[0045] 1) Model initialization

[0046] The hierarchical Dirichlet process hidden semi - Markov model uses the Dirichlet process as a non - parametric prior, avoiding the preset limit on the number of hidden states in the traditional HMM. The observation distribution uses a Gaussian distribution, with the initial mean set to 0 and the covariance matrix as the identity matrix to ensure the standardized processing of multi - dimensional driving behavior data; The duration of the hidden state is modeled by a Poisson distribution, with the expected value set to 6, reflecting the typical duration of each stage of the behavior before collision;

[0047] 2) Parameter inference

[0048] Use the Gibbs sampling method to iteratively infer the model parameters; This method updates the hidden state sequence according to the posterior distribution in each iteration, gradually optimizing the state transition probability matrix and observation distribution parameters until the model converges; Through iteration, the model can automatically identify the changes in the hidden states in the data and achieve accurate segmentation of the driving behavior;

[0049] 3) Hidden state extraction and segmentation

[0050] The original driving behavior data is divided into several segments according to the inferred hidden state sequence, and each segment corresponds to a specific driving behavior pattern, such as steady cruising, hard acceleration, or emergency braking. To ensure the effectiveness of the segmentation results, short segments with a duration less than 0.2 seconds are removed to exclude meaningless noise. Compared with traditional segmentation algorithms, HDP-HSMM has higher flexibility and accuracy in processing complex driving behavior data and can capture subtle differences in behavior changes.

[0051] S3. After completing the driving behavior segmentation, the Gaussian mixture model - Latent Dirichlet Allocation model (GMM-LDA) is used to cluster the segmented data to identify typical pre-collision driving behavior patterns. This method combines the processing ability of the Gaussian mixture model for continuous data with the ability of the LDA model to mine latent topics and is suitable for multi-dimensional and complex driving behavior analysis.

[0052] The method of using the Gaussian mixture model - Latent Dirichlet Allocation model to cluster the segmented data to identify typical pre-collision driving behavior patterns includes:

[0053] 1) Data discretization (GMM)

[0054] Since the LDA model is suitable for discrete data and the segmented driving behavior data is continuous, the Gaussian mixture model (GMM) is used to discretize the continuous behavior vectors into "driving units"; GMM assumes that the data comes from a mixture of multiple Gaussian distributions and optimizes the parameters through the Expectation-Maximization (EM) algorithm; in the E-step, the probability of each data point belonging to each Gaussian component is calculated, and in the M-step, the model parameters are updated. Finally, each behavior segment is mapped to a discrete class label; the number of mixture components is set to 50 to balance the computational efficiency and the expression of behavior characteristics.

[0055] 2) Topic modeling (LDA)

[0056] The discretized data is input into the LDA model for topic modeling. Each segmented data is regarded as a "document", and the discrete "driving units" are used as "words" to mine the potential driving behavior "topics" through the LDA model; the LDA model assumes that each behavior segment may contain multiple behavior patterns and uses Gibbs sampling or variational inference to deduce the model parameters to determine the probability distribution of behavior patterns in each segment.

[0057] 3) Behavior pattern recognition

[0058] Through clustering, multiple typical pre - collision driving behavior patterns are identified, including steady cruising, accelerating to avoid danger, emergency braking, and direction adjustment, etc. These pre - collision driving behavior patterns have obvious behavioral characteristics in different pre - collision scenarios, providing data support for further risk prediction and optimization of assisted driving strategies. The clustering results not only reveal common driving behavior patterns but also reflect the behavioral differences of different drivers in similar scenarios, contributing to the development of personalized driving behavior analysis and assistance systems.

[0059] The present invention can achieve automatic segmentation and recognition of driving behaviors based on non - parametric Bayesian models: The present invention adopts the Hierarchical Dirichlet Process Hidden Semi - Markov Model (HDP - HSMM) to realize the automatic segmentation and dynamic modeling of pre - collision driving behaviors of assisted driving vehicles. Different from traditional models that rely on a fixed number of states, the present invention uses non - parametric Bayesian methods and does not require presetting the number of states. It can automatically infer the optimal number of hidden states according to data characteristics, fully adapting to complex and variable driving behavior scenarios. The model also introduces a semi - Markov process to accurately model the state duration and capture the temporal characteristics of driving behaviors. While maintaining high computational performance, this method can more precisely depict the changes in pre - collision driving behaviors, effectively identify key behavior segments such as steady driving, rapid acceleration, and emergency braking, providing a reliable technical means for driving behavior analysis and risk prediction. Whether it is the fine - grained analysis of a single driving behavior or the pattern extraction of large - scale behavior data, the present invention shows strong adaptability and stability.

[0060] The present invention can achieve the extraction of typical risk driving patterns based on behavior theme modeling: The present invention uses the Gaussian Mixture Model - Latent Dirichlet Allocation Model (GMM - LDA) to realize the theme clustering and pattern recognition of pre - collision driving behaviors based on segmented behavior data. Compared with traditional methods based on simple clustering or rule - based classification, the present invention can deeply mine the implicit structure in driving behaviors and accurately capture the behavioral differences in different driving scenarios. The behavior patterns extracted by this method not only improve the recognition accuracy of risk driving states but also provide behavior prediction and decision - making support for traffic management and assisted driving systems, contributing to improving the intelligent level of traffic safety management. This technology has wide application value in real - time risk warning, driving behavior research, and the development of personalized assisted driving systems, significantly enhancing the scientificity and effectiveness of driving safety management.

[0061] Embodiment

[0062] This embodiment provides an unsupervised method for recognizing pre - collision driving behavior patterns of assisted driving vehicles. Taking the natural driving data of an assisted driving vehicle accident as an example, the specific steps are as follows:

[0063] (1) Extract pre - collision driving behavior data, as shown in Table 1:

[0064] Parameter item Data source Data frequency Unit Driver's brake pedal force value CAN 20hz N Braking force at the ADAS end CAN 20hz bar Acceleration T-box 20hz <![CDATA[m / s 2 > Speed T-box 20hz km / h

[0065] Table 1

[0066] (2) Figure 2 The figure shows an example of segmenting the multivariate variable time series of the accident process before a single collision using HDP-HSMM. The dashed line indicates the segmentation position. It can be seen that the model has a good effect on segmenting the time series, and the state transition of each behavior can be detected.

[0067] (3) Figure 3 The figure shows the clustering effect of GMM-LDA on driving patterns when defining 4 behavior patterns. Different colors are used to represent different behavior patterns. Red represents the driving pattern of stable driving, yellow represents the light braking mode of the machine side, blue represents the mode of the driver stepping on the brake heavily, and purple represents the mode of joint braking by the driver and the machine side 4. Observe Figure 3 It can be seen that LDA can effectively extract different pattern distributions in the time series according to the clustering results of GMM.

[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An unsupervised assisted driving vehicle pre-collision driving behavior pattern recognition method, characterized in that: The steps include: S1, extracting raw data before the collision and obtaining key driving parameters; S2, segmenting the continuous driving behavior data of the assisted driving vehicle before the collision through the Hierarchical Dirichlet Process Hidden Semi-Markov Model to identify different driving operation characteristics; S3, after completing the driving behavior segmentation, the segmented data is clustered using the Gaussian mixture model-latent Dirichlet allocation model to identify typical pre-collision driving behavior patterns.

2. The unsupervised pre-collision driving behavior pattern recognition method for assisted driving according to claim 1, characterized in that: In step S1, the method of extracting the original data before the collision and obtaining the key driving parameters is as follows: Extract natural driving data from CAN, T-BOX or vehicle-mounted global positioning system, and process outliers and missing values; extract key driving parameters based on assisted driving vehicle control behavior and vehicle operation fragments.

3. The unsupervised pre-collision driving behavior pattern recognition method for assisted driving according to claim 2 is characterized in that: Key driving parameters include speed, acceleration, steering wheel angle, brake pedal and accelerator pedal parameters.

4. The unsupervised pre-collision driving behavior pattern recognition method for assisted driving according to claim 3 is characterized in that: In step S2, the method of segmenting the continuous driving behavior data of the assisted driving vehicle before the collision by using the Hierarchical Dirichlet Process Hidden Semi-Markov Model includes: 1) Model initialization The hierarchical Dirichlet process hidden semi-Markov model uses the Dirichlet process as a non-parametric prior, the observation distribution adopts a Gaussian distribution, the initial mean is set to 0, and the covariance matrix is ​​the unit matrix to ensure the standardized processing of multi-dimensional driving behavior data; the duration of the hidden state is modeled by a Poisson distribution, and the expected value is set to 6, reflecting the typical duration of each stage of pre-collision behavior; 2) Parameter inference The Gibbs sampling method is used to iteratively infer the model parameters. In each iteration, the method updates the hidden state sequence according to the posterior distribution, and gradually optimizes the state transition probability matrix and the observation distribution parameters until the model converges. Through iteration, the model can automatically identify the hidden state changes in the data and achieve accurate segmentation of driving behavior. 3) Hidden state extraction and segmentation By inferring the hidden state sequence, the original driving behavior data is divided into several segments, each of which corresponds to a specific driving behavior pattern.

5. The unsupervised pre-collision driving behavior pattern recognition method for assisted driving according to claim 4 is characterized in that: To ensure the validity of the segmentation results, short periods of time with a duration of less than 0.2 seconds were removed to exclude meaningless noise.

6. The unsupervised pre-collision driving behavior pattern recognition method for assisted driving according to claim 5, characterized in that: In step S3, the method of clustering the segmented data using the Gaussian mixture model-latent Dirichlet allocation model to identify typical pre-collision driving behavior patterns includes: 1) Data discretization Use Gaussian mixture model to discretize continuous behavior vector into "driving units"; GMM assumes that data comes from a mixture of multiple Gaussian distributions and optimizes parameters through expectation-maximization algorithm; E step calculates the probability that each data point belongs to each Gaussian component, and M step updates model parameters, finally mapping each behavior segment to a discrete category label; 2) Topic Modeling The discretized data is input into the LDA model for topic modeling. Each segment is regarded as a "document" and the discrete "driving unit" is regarded as a "word". The LDA model is used to mine potential driving behavior "topics". The LDA model assumes that each behavior segment may contain multiple behavior patterns. Gibbs sampling or variational inference is used to derive model parameters to determine the probability distribution of behavior patterns in each segment. 3) Behavioral pattern recognition Through clustering, several typical pre-collision driving behavior patterns were identified.

7. The unsupervised pre-collision driving behavior pattern recognition method for assisted driving according to claim 6, characterized in that: The pre-collision driving behavior patterns include smooth cruising, accelerating to avoid danger, emergency braking, and steering adjustment.

Images