Scene heterogeneous-oriented multi-dimensional abnormal driving behavior feature analysis and recognition method

Abstract

The invention relates to a scene heterogeneous-oriented multi-dimensional abnormal driving behavior feature analysis and recognition method. The method comprises the following steps of S1, performing multi-dimensional element analysis and scene classification on a scene; s2, collecting driving information of people, vehicles, roads and environments; s3, identifying and manually calibrating the abnormal driving behavior data; s4, on the basis of original time sequence data, according to driving tasks and abnormal driving behavior modes of different scenes, constructing a refined evaluation system, and according to a large amount of actually measured data, performing machine learning to determine an index threshold value; s5, when the vehicle drives into the scene road section, recognizing the current road section through a vehicle positioning system, and recognizing the scene information of the current road section; s6, according to the acquired scene information, acquiring a driving task and an evaluation system of a corresponding scene; and S7, identifying the abnormal driving behavior of the driver according to the refined evaluation system in the scene. According to the invention, the abnormal driving behavior can be analyzed in detail, and the development of road traffic safety can be promoted.

Description

Technical Field

[0001] The present invention relates to the fields of traffic safety and driving behavior, and more specifically, to a method for analyzing and identifying multi-dimensional abnormal driving behavior characteristics in heterogeneous scenarios. Background Art

[0002] With the continuous development of randomized intelligent transportation systems, research on traffic safety is becoming increasingly mature both domestically and internationally. However, different research methods and indicators can produce significantly different analytical results in different driving scenarios. Therefore, accurately analyzing driving behavior in different driving scenarios is of great significance to traffic safety research. However, existing methods for analyzing driving behavior characteristics in driving behavior simulations have the following limitations:

[0003] 1. Insufficient consideration of multidimensional driving behavior characteristics. Traditional driving behavior models mostly focus on single-dimensional characteristics, such as vehicle trajectory or driver physiological indicators. They fail to comprehensively consider key characteristics such as driver-vehicle-road-environment, and in particular, fail to consider risk factors based on the driver's multidimensional characteristics.

[0004] 2. Inadequate scenario classification. Previous driving model simulation research has often focused on studying driving behavior in a single scenario, failing to consider the different driving tasks and risk characteristics of different scenarios and conducting comparative analysis of driving behavior in different scenarios. 3. Inadequate classification of abnormal driving behavior. Abnormal driving behavior poses a serious threat to traffic safety. Existing driving behavior models typically categorize abnormal driving behavior in a general manner, failing to provide detailed classification based on different scenarios. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a multi-dimensional abnormal driving behavior feature analysis and identification method for scene heterogeneity. It classifies driving behavior scenes through multi-dimensional driving behavior feature element analysis, comprehensively considers multi-dimensional element analysis of traffic scenes, and proposes a detailed evaluation system under different scenes, so as to specifically analyze the influencing factors of driving behavior in the scene, and better identify and analyze abnormal driving behavior.

[0006] The technical solution adopted by the present invention to solve the technical problem is to construct a multi-dimensional abnormal driving behavior feature analysis and identification method for scene heterogeneity, including the following steps:

[0007] S1. Analyze and classify scenes in multiple dimensions;

[0008] S2, collect driving information of people, vehicles, roads and environment;

[0009] S3. Identify and manually calibrate abnormal driving behavior data, and classify abnormal driving behaviors into universal abnormal driving behaviors and specific abnormal driving behaviors based on vehicle trajectory information and driving tasks in different scenarios;

[0010] S4. Based on the original time series data, a detailed evaluation system is constructed according to driving tasks in different scenarios and abnormal driving behavior patterns. Machine learning is performed based on a large amount of measured data to determine the indicator thresholds.

[0011] S5. When the vehicle enters the scene section, the vehicle positioning system identifies the current section and identifies the scene information of the current section;

[0012] S6. Obtaining a driving task and evaluation system corresponding to the scenario based on the acquired scenario information;

[0013] S7. Identify the driver's abnormal driving behavior based on the refined evaluation system in the scenario.

[0014] According to the above solution, in step S1, the scene elements are analyzed from the aspects of road grade, traffic volume, and road section.

[0015] According to the above scheme, in step S2, the collected information about people, vehicles, roads, and environment is used to extract vehicle data, which includes vehicle trajectory, longitudinal acceleration, lateral speed, lateral acceleration, lateral offset, TTC, and ETTC;

[0016] Extracting driver information, the driver information including eye movement frequency and steering wheel grip;

[0017] Extracting road information, wherein the road information includes road structure and road type;

[0018] The surrounding vehicle information is extracted, where the surrounding vehicle information includes track information and speed of the surrounding vehicles.

[0019] According to the above solution, in step S3, the abnormal driving behavior includes universal abnormal driving behavior and specific abnormal driving behavior;

[0020] The universal abnormal driving behavior described applies to all scenarios, including lane crossing, repeated lane changes, and sudden braking while maintaining lane.

[0021] The specific abnormal driving behavior is applicable to specific scenarios and is determined based on the characteristics of the specific scenarios and driving tasks. When two vehicles interact and the driver needs to slow down in time, if the driver starts slowing down too late, resulting in a traffic conflict between the two vehicles, it can be determined that the driver started slowing down later than a certain point, the deceleration behavior was too gradual to meet the safety distance between the two vehicles, or the deceleration behavior was too drastic, which is an abnormal driving behavior.

[0022] According to the above solution, in step S3, after the abnormal driving behavior is identified, the abnormal driving behavior is manually labeled as universal abnormal driving behavior or specific abnormal driving behavior based on the vehicle trajectory and the scene driving task.

[0023] According to the above scheme, in step S4, the original time series data and scene classification labels are input, key indicators are screened using known machine learning methods, and indicator thresholds are determined based on a clustering algorithm.

[0024] According to the above scheme, the method of screening key indicators by using known machine learning methods and determining indicator thresholds based on clustering algorithms includes the following steps:

[0025] S401, collect the original time series data X=[x1,x2,...,x n ] is preprocessed, including normalization and padding operations; then the feature vector of each time step is taken as an element in the input sequence to construct the input matrix X∈R^{n×d}, where n is the length of the time series and d is the feature dimension of each time step;

[0026] S402. Use the Transformer encoder to process the input sequence X to obtain the encoded output sequence. First, perform a linear transformation on the input sequence X∈R^{n×d}:

[0027] A=XW A ,B=XW B ,C=XW C

[0028] in, Both are learnable weight matrices;

[0029] S403: Perform self-attention calculation on the transformed A, B, and C, and then calculate multiple attention heads in parallel to complete multi-head attention calculation:

[0030]

[0031] MultiHead(X)=Concat(head1,...,head h )W O

[0032] Where each header is calculated as:

[0033]

[0034] Combined with the added positional encoding, the multi-head attention output is fully connected twice as shown below, and the final output feature matrix Z = [z1, z2, ..., z m ]:

[0035] FFN(x)=max(0,xW1+b1)W2+b2

[0036] Based on the output of the Transformer encoder, clustering algorithms or statistical-based methods are used to screen key indicators and determine thresholds.

[0037] According to the above solution, in step S5, when entering the driving scene, the vehicle obtains geographic location information in real time through the positioning system, and completes the scene recognition in combination with the vehicle's perception of surrounding vehicles and environmental information.

[0038] The implementation of the present invention's multi-dimensional abnormal driving behavior feature analysis and identification method for scene heterogeneity has the following beneficial effects:

[0039] The present invention provides a multi-dimensional abnormal driving behavior feature analysis and identification method for scene heterogeneity. Taking into account the different driving tasks of drivers in different scenes, it analyzes the elements of the traffic system, classifies different driving scenes, and makes a more detailed analysis of abnormal driving behavior, which helps to promote the development of highway traffic safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0041] Figure 1 This is a flow chart of the method for analyzing and identifying multi-dimensional abnormal driving behavior characteristics facing scene heterogeneity of the present invention;

[0042] Figure 2 It is a free flow scene diagram of the method for analyzing and identifying multi-dimensional abnormal driving behavior characteristics with respect to scene heterogeneity of the present invention;

[0043] Figure 3 It is a congestion flow scene diagram of the multi-dimensional abnormal driving behavior feature analysis and identification method for scene heterogeneity of the present invention;

[0044] Figure 4 It is a scene diagram of an electrical construction area for the present invention's method for analyzing and identifying characteristics of multi-dimensional abnormal driving behavior with respect to scene heterogeneity. DETAILED DESCRIPTION

[0045] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0046] like Figure 1 As shown, the method for analyzing and identifying abnormal driving behavior characteristics in a multi-dimensional manner for scene heterogeneity of the present invention includes the following steps:

[0047] S1. Analyze the multi-dimensional elements and classify the scenes.

[0048] Analyzing scene elements can be done from multiple perspectives: road grade, traffic volume, and road section. Road grades vary in speed limits, number of lanes, and access points. Traffic volume is categorized as heavy, medium, or light based on road grade. Road sections include general sections, tunnels, intersections, and cross-river bridges. Table 1 provides an example of multi-dimensional element analysis.

[0049] Table 1 Multidimensional factor classification table

[0050]

[0051] S2. Collect driving-related information such as people, vehicles, roads, and environment.

[0052] The collected information about people, vehicles, roads, and the environment includes vehicle data such as vehicle trajectory, longitudinal acceleration, lateral speed, lateral acceleration, lateral offset, TTC, ETTC, etc.; driver information such as eye movement frequency, steering wheel grip, etc.; road information including road structure, road type, etc.; surrounding vehicle information including trajectory information, speed, and other information of surrounding vehicles.

[0053] S3. Identify and manually calibrate abnormal driving behavior data. Based on vehicle trajectory information and different driving tasks, abnormal driving behaviors are divided into universal abnormal driving behaviors and specific abnormal driving behaviors.

[0054] Abnormal driving behaviors are categorized as universal and specific, as follows: Universal abnormal driving behaviors apply to all scenarios, such as driving in lane while maintaining a lane, repeatedly changing lanes, and sudden braking. Specific abnormal driving behaviors apply to specific scenarios and are determined based on the characteristics of the specific scenario and the driving task. For example, when two vehicles are interacting and the driver needs to slow down, if the driver slows down too late, causing a traffic conflict between the two vehicles, then abnormal driving behavior can be considered as starting to slow down later than a certain point, slowing down too gradually to maintain a safe distance between the two vehicles, or slowing down too drastically.

[0055] After identifying abnormal driving behavior, the abnormal driving behavior is manually labeled as universal abnormal driving behavior or specific abnormal driving behavior based on the vehicle trajectory and scene driving task.

[0056] As shown in Table 2, the specific abnormal driving behaviors and universal abnormal driving behaviors on the main line section of the highway.

[0057] Table 2 Abnormal driving behavior classification and evaluation system under main line scenario classification

[0058]

[0059] S4. Based on the original time series data, a more detailed evaluation system is constructed according to driving tasks in different scenarios and abnormal driving behavior patterns, and machine learning is performed based on a large amount of measured data to determine the indicator thresholds.

[0060] Input the original time series data and scene classification labels, use known machine learning methods (such as random forest and LSTM) to screen key indicators, and determine the indicator threshold based on the clustering algorithm (such as K-means).

[0061] Taking the transformer model as an example, the collected original time series data X=[x1,x2,...,x n ] is preprocessed, including normalization, padding and other operations; then the feature vector of each time step is used as an element in the input sequence to construct the input matrix X∈R^{n×d}, where n is the length of the time series and d is the feature dimension of each time step.

[0062] Then use the Transformer encoder to process the input sequence X to obtain the encoded output sequence. First, perform a linear transformation on the input sequence X∈R^{n×d}:

[0063] A=XW A ,B=XW B ,C=XW C

[0064] in, are all learnable weight matrices.

[0065] Next, self-attention calculation is performed on the transformed A, B, and C, and then multiple attention heads are calculated in parallel to complete the multi-head attention calculation.

[0066]

[0067] MultiHead(X)=Concat(head1,...,head h )W O

[0068] where each header is calculated as:

[0069]

[0070] Combined with the added positional encoding, the multi-head attention output is fully connected twice as shown below, and the final output feature matrix Z = [z1, z2, ..., z m ].

[0071] FFN(x)=max(0,xW1+b1)W2+b2

[0072] Based on the output of the Transformer encoder, clustering algorithms or statistical methods are used to screen key indicators and determine thresholds. The following calculation takes K-meas as an example. The feature representation Z output by the Transformer encoder is Z = [z1, z2, ..., z m The K-means algorithm is used to cluster the normal driving behavior data set, and the normal distribution range of each indicator is determined by the 3σ principle. By analyzing the feature distribution within different clusters, key indicators are screened out and the threshold range of each key indicator is determined.

[0073] When new driving behavior data arrives, the Transformer model first generates its feature representation. The distance between this feature representation and the cluster centers obtained through K-means clustering is then calculated to determine the cluster to which it belongs. Next, the key indicators in the feature representation are checked to see if they exceed a preset threshold. If so, the data is considered abnormal driving behavior.

[0074] S5. When the vehicle enters a scene section, the current section can be identified through the vehicle positioning system, and information about the current section scene can be identified.

[0075] When entering a driving scene, the vehicle can obtain geographic location information in real time through the positioning system, and complete the scene recognition by combining the vehicle's perception of surrounding vehicles and environmental information.

[0076] S6. Obtain a driving task and evaluation system corresponding to the scenario based on the acquired scenario information.

[0077] S7. Identify the driver's abnormal driving behavior based on the detailed evaluation system for the scenario.

[0078] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are protected by the present invention.

Claims

1. A multi-dimensional abnormal driving behavior feature analysis and identification method for scene heterogeneity, characterized by: The following steps are involved: S1. Analyze and classify scenes in multiple dimensions; S2, collect driving information of people, vehicles, roads and environment; S3. Identify and manually calibrate abnormal driving behavior data, and classify abnormal driving behaviors into universal abnormal driving behaviors and specific abnormal driving behaviors based on vehicle trajectory information and driving tasks in different scenarios; S4. Based on the original time series data, a detailed evaluation system is constructed according to driving tasks in different scenarios and abnormal driving behavior patterns, and machine learning is performed based on measured data to determine indicator thresholds; S5. When the vehicle enters the scene section, the vehicle positioning system identifies the current section and identifies the scene information of the current section; S6. Obtaining a driving task and evaluation system corresponding to the scenario based on the acquired scenario information; S7. Identify the driver's abnormal driving behavior based on the refined evaluation system in the scenario.

2. The method for analyzing and identifying multi-dimensional abnormal driving behavior characteristics in heterogeneous scenarios according to claim 1 is characterized in that: In step S1, the scene elements are analyzed from the aspects of road grade, traffic volume, and road section.

3. The method for analyzing and identifying multi-dimensional abnormal driving behavior characteristics in heterogeneous scenarios according to claim 1 is characterized in that: In step S2, the collected information about people, vehicles, roads, and environment is used to extract vehicle data, which includes vehicle trajectory, longitudinal acceleration, lateral speed, lateral acceleration, lateral offset, TTC, and ETTC; Extracting driver information, the driver information including eye movement frequency and steering wheel grip; Extracting road information, wherein the road information includes road structure and road type; The surrounding vehicle information is extracted, where the surrounding vehicle information includes track information and speed of the surrounding vehicles.

4. The method for analyzing and identifying multi-dimensional abnormal driving behavior characteristics in heterogeneous scenarios according to claim 1 is characterized in that: In step S3, the abnormal driving behavior includes universal abnormal driving behavior and specific abnormal driving behavior; The universal abnormal driving behavior described applies to all scenarios, including lane crossing, repeated lane changes, and sudden braking while maintaining lane. The specific abnormal driving behavior is applicable to specific scenarios and is determined based on the characteristics of the specific scenarios and driving tasks. When two vehicles interact and the driver needs to slow down in time, if the driver starts slowing down too late, resulting in a traffic conflict between the two vehicles, it is determined that the driver starts slowing down later than a certain point, the deceleration behavior is too gradual to meet the safety distance between the two vehicles, or the deceleration behavior is too drastic, which is an abnormal driving behavior.

5. The method for analyzing and identifying multi-dimensional abnormal driving behavior characteristics in heterogeneous scenarios according to claim 4 is characterized in that: In step S3, after the abnormal driving behavior is identified, the abnormal driving behavior is manually labeled as universal abnormal driving behavior or specific abnormal driving behavior based on the vehicle trajectory and the scene driving task.

6. The method for analyzing and identifying multi-dimensional abnormal driving behavior characteristics in heterogeneous scenarios according to claim 1 is characterized in that: In step S4, the original time series data and scene classification labels are input, key indicators are screened using known machine learning methods, and indicator thresholds are determined based on a clustering algorithm.

7. The method for analyzing and identifying multi-dimensional abnormal driving behavior characteristics in heterogeneous scenarios according to claim 6 is characterized in that: The method of screening key indicators by using known machine learning methods and determining indicator thresholds based on clustering algorithms includes the following steps: S401, collect the original time series data X=[x1,x2,...,x n ] is preprocessed, including normalization and padding operations; then the feature vector of each time step is taken as an element in the input sequence to construct the input matrix X∈R^{n×d}, where n is the length of the time series and d is the feature dimension of each time step; S402. Use the Transformer encoder to process the input sequence X to obtain the encoded output sequence. First, perform a linear transformation on the input sequence X∈R^{n×d}: A=XW A ,B=XW B ,C=XW C in, Both are learnable weight matrices; S403: Perform self-attention calculation on the transformed A, B, and C, and then calculate multiple attention heads in parallel to complete multi-head attention calculation: MultiHead(X)=Concathead1,...,head h )W O Where each header is calculated as: Combined with the added positional encoding, the multi-head attention output is fully connected twice as shown below, and the final output feature matrix Z = [z1, z2, ..., z m ]: FFN(x)=max(0,xW1+b1)W2+b2 Based on the output of the Transformer encoder, clustering algorithms or statistical-based methods are used to screen key indicators and determine thresholds.

8. The method for analyzing and identifying multi-dimensional abnormal driving behavior characteristics in heterogeneous scenarios according to claim 1 is characterized in that: In step S5, when entering the driving scene, the vehicle obtains geographic location information in real time through the positioning system, and completes the scene recognition in combination with the vehicle's perception of surrounding vehicles and environmental information.

Images