A traffic dangerous driving scene intelligent identification and classification method and system

Abstract

The application provides a traffic dangerous driving scene intelligent identification and classification method and system, the traffic dangerous driving scene intelligent identification and classification method comprises the following steps: extracting each accident characteristic value corresponding to each preset accident characteristic in a traffic accident video; converting each accident characteristic value into a numerical form respectively; performing dimension reduction by using classification principal component analysis; obtaining the best clustering number and clustering result by using the elbow method and the K-means clustering method, and dividing the traffic accident video into multiple different categories according to the clustering result; and determining the main features of each category and the dangerous driving categories corresponding to each category by using statistical analysis. The method can intelligently identify different types of traffic dangerous driving scenes, and better perform driver risk perception training and assist scene identification and decision-making of an automatic driving system.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent transportation systems, in particular to a traffic dangerous driving scene recognition and classification method and system. BACKGROUND

[0002] With the rapid development of autonomous driving technology, traffic safety has become one of the key issues. The existing autonomous driving systems use vision, radar and lidar sensors to perceive the environment around the vehicle, which has improved the safety of driving to a certain extent, but the accurate identification and classification of traffic dangerous scenes is still a challenge, especially in complex and dynamic traffic environments, the existing technology shows certain limitations in dealing with diversified dangerous driving behaviors.

[0003] The existing autonomous driving systems realize basic safety functions through simple obstacle detection, vehicle tracking or pedestrian detection, but when encountering complex traffic situations, especially in dangerous traffic scenes involving pedestrian crossing and vehicle illegal lane changing, the recognition ability of these systems is usually insufficient. The existing methods often rely on single sensor data or overly simple rules, lack comprehensive analysis of multi-factor complex interactions, and cannot effectively deal with potential risks in highway, urban road or complex intersection scenes. SUMMARY

[0004] In order to solve the problems of the prior art, the present application provides a traffic dangerous driving scene intelligent recognition and classification method and system.

[0005] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0006] A traffic dangerous driving scene intelligent recognition and classification method, characterized by: for at least two traffic accident videos, according to steps S1 to S5, the classification of traffic accident videos is realized:

[0007] Step S1, a preset number of traffic accident videos are obtained, and each accident feature value corresponding to each accident feature in each traffic accident video is extracted by using deep learning and computer vision technology, and then step S2 is executed;

[0008] Step S2, for each accident feature value corresponding to each traffic accident video, combine the preset conversion rule, convert each accident feature value into a numerical form, i.e. accident numerical feature, and then perform step S3;

[0009] Step S3, based on the accident numerical features corresponding to each traffic accident video, dimension reduction is performed by using classification principal component analysis to obtain principal component scores, and then step S4 is executed;

[0010] Step S4, based on the principal component scores, the elbow method and the K-means clustering method are used to obtain the optimal clustering number and clustering results, and the total traffic accident videos are divided into different categories according to the clustering results, and then step S5 is executed;

[0011] Step S5, based on the traffic accident videos contained in each category, the main features of each category are determined by using statistical analysis combined with the accident feature values corresponding to each traffic accident video in each category, and then the dangerous driving categories corresponding to each category are determined.

[0012] Further, the step S1 obtains the traffic accident videos from the image capture devices carried by the accident participants or non-accident participants.

[0013] Further, the step S3 uses classification principal component analysis to reduce dimensionality to obtain principal component scores according to the following steps:

[0014] Step S31, based on the accident numerical features corresponding to each traffic accident video, an original data matrix is constructed, and the covariance matrix of the original data matrix is calculated, and then step S32 is executed;

[0015] Step S32, the characteristic values and the characteristic vectors corresponding to the characteristic values are calculated according to the covariance matrix, and the characteristic values are sorted in descending order to obtain λ1, λ2…λ m , and the characteristic vectors a1, a2…a m corresponding to the characteristic values, and the i-th principal component is Y i =Xa i (1≤i≤m), wherein X is the original data matrix, a i is the i-th characteristic vector, and then step S33 is executed;

[0016] Step S33, based on the characteristic values sorted in descending order, the principal component contribution rate and the cumulative contribution rate are calculated according to the following formula, when the cumulative contribution rate reaches the preset threshold value, the first k principal components are selected to form a new feature space, and then step S34 is executed:

[0017]

[0018] wherein λ k is the characteristic value of the k-th principal component, is the sum of all characteristic values, is the principal component contribution rate, is the cumulative contribution rate;

[0019] Step S34, based on the characteristic vectors of the original data matrix, a loading matrix is constructed, and the principal component scores of the first k principal components are obtained by using the following formula:

[0020]

[0021] where PC i represents the principal component score of the i-th principal component, n is the total number of accident numerical features, w ij is the coefficient of the i-th row and j-th column in the feature loading matrix, x j is the j-th accident numerical feature.

[0022] Further, the step S4 comprises the following steps of using elbow method and K-means clustering method to obtain the optimal clustering number and clustering result:

[0023] In step S41, the within-cluster sum of squares under different clustering numbers k is calculated according to the following formula based on the preset different clustering numbers k, and the within-cluster sum of squares curve under different clustering numbers k is drawn, and then step S42 is executed:

[0024]

[0025] where WCSS is the within-cluster sum of squares, k is the clustering number, C i represents the i-th cluster, x j is the j-th data point belonging to the i-th cluster, c i is the centroid of the i-th cluster.

[0026] In step S42, the value at the elbow point of the within-cluster sum of squares curve is selected as the optimal clustering number k based on the within-cluster sum of squares curve under different clustering numbers k, that is, the k value at the bending point of the within-cluster sum of squares curve is selected as the optimal clustering number k.

[0027] In step S43, k centroid vectors are randomly initialized based on the optimal clustering number k, and the Euclidean distance of each data point to the k centroids is calculated according to the following formula, and each data point is assigned to the cluster where the nearest centroid is located, wherein the data point is the principal component score, and then step S44 is executed:

[0028]

[0029] where d(x i , c j ) is the Euclidean distance of each data point to the k centroids, x il is the data point, c jl is the j-th centroid, and p is the data dimension.

[0030] In step S44, the centroid vector of each cluster is recalculated based on the data points of each cluster, and the centroid position is the mean of all data points in the cluster, and then step S45 is executed.

[0031]

[0032] wherein c j is the centroid vector of each cluster, N j is the number of data points in the jth cluster, x i is the data point;

[0033] Step S45, repeating the data point assignment and centroid updating steps until the centroid positions converge or a predetermined maximum number of iterations is reached.

[0034] Further, the step S5 determines the main features of each category respectively by statistical analysis, specifically: the statistical analysis is performed on the accident feature values corresponding to the traffic accident videos included in each category, if an accident feature value accounts for more than 50% in the corresponding category, the accident feature value is the main feature of the corresponding category; if multiple accident feature values account for more than 50% and 30% in the corresponding category, the accident feature values are the main features of the corresponding category; if the proportions of all accident feature values in the corresponding category do not exceed 50%, the two accident feature values with the largest proportions are selected as the main features of the corresponding category.

[0035] A traffic dangerous driving scene intelligent recognition and classification system, characterized in that it comprises:

[0036] A data acquisition module: configured to acquire a certain number of traffic accident videos;

[0037] A feature extraction module: configured to extract each accident feature value corresponding to each preset accident feature in the traffic accident video by using deep learning and computer vision technology;

[0038] A dimension reduction processing module: configured to perform dimension reduction on the extracted features according to classification principal component analysis;

[0039] A clustering analysis module: configured to divide the traffic accident videos into multiple different categories by using elbow method and K-Means clustering method;

[0040] A statistical analysis module: configured to determine the main features of each category and the dangerous driving categories corresponding to each category.

[0041] Compared with the prior art, the beneficial effects of the present application are that: the dimension reduction processing of the accident numerical features of the traffic accident videos is performed by classification principal component analysis, at the same time, the traffic accident videos are divided into multiple categories by using clustering analysis, and the main features of each category are determined by using statistical analysis, so as to determine the dangerous driving categories corresponding to each category; the classification results of the present application can also be used as input for advanced driver assistance system (ADAS) and automatic driving vehicle, to assist the automatic driving system in dangerous driving scene recognition and decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is a flow chart of a traffic dangerous driving scene intelligent recognition and classification method.

[0043] Figure 2 is a schematic diagram of the elbow method WCSS curve. DETAILED DESCRIPTION

[0044] The technical solutions of the present application will be described in detail below with reference to the drawings.

[0045] Reference Figure 1 A traffic dangerous driving scene intelligent recognition and classification method, characterized in that: for at least two traffic accident videos, the classification of the traffic accident videos is realized according to steps S1 to S5:

[0046] Step S1, using the driving recorders carried by the accident participants, 1013 traffic accident videos including urban trunk roads, secondary trunk roads, rural roads and highway ramp road sections are obtained, and each accident feature value corresponding to each preset accident feature in each traffic accident video is extracted using deep learning and computer vision technology. Specifically, the accident features are the types of the accident participants, the start and end positions, the driving trajectories, the motion directions and the motion distances, and then step S2 is executed;

[0047] Step S2, for each accident feature value corresponding to each traffic accident video, each accident feature value is converted into a numerical form, i.e. an accident numerical feature, by combining the optimal scaling rule, and the accident numerical feature is standardized, i.e. each accident numerical feature is scaled to a range with a mean value of 0 and a standard deviation of 1, to eliminate the differences between the accident numerical features, and then step S3 is executed;

[0048] Step S3, based on the accident numerical features corresponding to each traffic accident video, dimension reduction is performed using classification principal component analysis to obtain principal component scores, and then step S4 is executed;

[0049] Specifically, the step S3 includes:

[0050] Step S31, based on the accident numerical features corresponding to each traffic accident video, an original data matrix is constructed, and the covariance matrix of the original data matrix is calculated, and then step S32 is executed;

[0051] Step S32, according to the covariance matrix, the eigenvalues and the eigenvectors corresponding to the eigenvalues are calculated, and the eigenvalues are sorted in descending order to obtain λ1, λ2…λ m , and the eigenvectors a1, a2…a m corresponding to the eigenvalues, and the ith principal component is Y i =Xa i (1≤i≤m), wherein X is the original data matrix, ai For the i-th eigenvector, then step S33 is performed;

[0052] Step S33, based on the descendingly arranged eigenvalues, the principal component contribution rate and the cumulative contribution rate are calculated according to the following formula: when the cumulative contribution rate reaches 80%, the first k principal components are selected to form a new feature space, in this embodiment, as shown in Table 1, the first 3 principal components are selected to form a new feature space, then step S34 is performed:

[0053]

[0054] wherein λ k is the eigenvalue of the k-th principal component, is the sum of all eigenvalues, is the principal component contribution rate, is the cumulative contribution rate;

[0055] Table 1

[0056] Principal component Eigenvalue Principal component contribution rate (%) 1 1.787 35.741 2 1.467 29.337 3 0.89 17.803 Total 4.144 82.881

[0057] Step S34, based on the eigenvectors of the original data matrix, a loading matrix is constructed, and the principal component scores of the first k principal components are obtained by using the following formula:

[0058]

[0059] wherein PC i represents the principal component score of the i-th principal component, n is the total number of numerical characteristics of accidents, w ij is the coefficient of the i-th row and the j-th column in the feature loading matrix, x j is the j-th numerical characteristic of accidents;

[0060] In this embodiment, the three principal component scores are respectively:

[0061] RC1: y1 = -0.036x1-0.665x2+0.690x3-0.097x4+0.266x5

[0062] RC2: y2 = 0.629x1+0.214x2-0.050x3-0.470x4+0.579x5

[0063] RC3: y3 = 0.427x1-0.128x2-0.085x3+0.850x4+0.267x5;

[0064] Step S4, based on the principal component scores, the elbow method and the K-means clustering method are used to obtain the optimal clustering number and clustering results, and all traffic accident videos are divided into multiple different categories according to the clustering results, then step S5 is performed;

[0065] Specifically, step S4 includes:

[0066] Step S41: Based on the preset number of clusters k, calculate the sum of squared intra-cluster errors for different numbers of clusters k according to the following formula, and plot the curves of the sum of squared intra-cluster errors for different numbers of clusters k. Then proceed to step S42:

[0067]

[0068] Where WCSS is the sum of squared errors within a cluster, k is the number of clusters, and C i Let x represent the i-th cluster. j For the j-th data point belonging to the i-th cluster, c i Let be the centroid of the i-th cluster;

[0069] Step S42: Based on the intra-cluster error sum of squares curves under different cluster numbers k, select the value at the elbow of the intra-cluster error sum of squares curve as the optimal cluster number k, that is, select the k value at the bend of the intra-cluster error sum of squares curve as the optimal cluster number k, such as... Figure 2 As shown, in this embodiment, the optimal number of clusters is 4.

[0070] Step S43: Based on the optimal number of clusters k, randomly initialize k centroid vectors, and calculate the Euclidean distance from each data point to the k centroids according to the following formula. Assign each data point to the cluster containing the nearest centroid, where the data point is the principal component score. Then proceed to step S44:

[0071]

[0072] Where d(x) i ,c j ) is the Euclidean distance from each data point to the k centroids, x il For data points, c jl Let p be the j-th centroid, and p be the data dimension;

[0073] Step S44: Based on the data points of each cluster, recalculate the centroid vector of each cluster, with the centroid position being the mean of all data points within that cluster. Then, proceed to step S45:

[0074]

[0075] Among them, c j For each cluster's centroid vector, N j Let x be the number of data points in the j-th cluster. i For data points;

[0076] Step S45, repeating the data point assignment and centroid update steps until the centroid position converges or reaches a predetermined maximum number of iterations;

[0077] Step S5, based on the traffic accident videos contained in each category, combining the accident feature values corresponding to each traffic accident video in each category, using statistical analysis to determine the main features of each category respectively, and further determining the dangerous driving category corresponding to each category, that is, the accident feature values corresponding to the traffic accident videos contained in each category are counted respectively, if a certain accident feature value accounts for more than 50% in the corresponding category, the accident feature value is the main feature of the corresponding category; if multiple accident feature values account for more than 50% and 30% in the corresponding category, the accident feature value is the main feature of the corresponding category; if the proportion of all accident feature values in the corresponding category does not exceed 50%, the two accident feature values with the largest proportion are selected as the main features of the corresponding category; and the determined main features can be input to the scene recognition module of the automatic driving system.

[0078] A traffic dangerous driving scene intelligent recognition and classification system, characterized in that it comprises:

[0079] A data acquisition module for acquiring a certain number of traffic accident videos;

[0080] A feature extraction module for extracting each accident feature value corresponding to each preset accident feature in the traffic accident video using deep learning and computer vision technology;

[0081] A dimension reduction processing module for reducing the dimension of the extracted features according to the classification principal component analysis;

[0082] A clustering analysis module for dividing the traffic accident videos into multiple different categories using the elbow method and K-Means clustering method;

[0083] A statistical analysis module for determining the main features of each category and the dangerous driving category corresponding to each category.

[0084] The above is only a preferred embodiment of the present application, and does not limit the present application in any way. Any simple modification, change and equivalent structural change made according to the technical essence of the present application to the above embodiment are still within the protection scope of the technical solution of the present application.

Claims

1. A method for intelligent identification and classification of dangerous driving scenarios in traffic, characterized in that: For at least two traffic accident videos, classify the traffic accident videos according to steps S1 to S5: Step S1: Obtain a preset number of traffic accident videos, use deep learning and computer vision technology to extract the accident feature values ​​corresponding to the preset accident features in each traffic accident video, and then execute step S2. Step S2: For each accident feature value corresponding to each traffic accident video, and in accordance with the preset conversion rules, each accident feature value is converted into a numerical form, namely the accident numerical feature. The accident numerical feature is then standardized, and then step S3 is executed. Step S3: Based on the accident numerical features corresponding to each traffic accident video, use classification principal component analysis to reduce dimensionality and obtain principal component scores, then proceed to step S4. Step S4: Based on the principal component scores, the elbow method and K-means clustering method are used to obtain the optimal number of clusters and clustering results. Based on the clustering results, all traffic accident videos are divided into multiple different categories, and then step S5 is executed. Step S5: Based on the traffic accident videos included in each category, and combined with the accident feature values ​​corresponding to each traffic accident video in each category, statistical analysis is used to determine the main features of each category, and then the dangerous driving category corresponding to each category is determined. In step S1, traffic accident video is acquired from an image capture device carried by either the party involved in the accident or a non-party involved in the accident. Step S5 utilizes statistical analysis to determine the main features of each category. Specifically, it involves: statistically analyzing the accident feature values ​​corresponding to the traffic accident videos included in each category; if a certain accident feature value accounts for more than 50% of the corresponding category, then the accident feature value is the main feature of the corresponding category; if multiple accident feature values ​​account for more than 50% and 30% of the corresponding category, then the accident feature value is the main feature of the corresponding category; if the proportion of all accident feature values ​​in the corresponding category does not exceed 50%, then the two accident feature values ​​with the largest proportions are selected as the main features of the corresponding category.

2. The intelligent identification and classification method for dangerous driving scenarios according to claim 1, characterized in that, Step S3 involves dimensionality reduction using categorical principal component analysis to obtain principal component scores, as follows: Step S31: Based on the accident numerical features corresponding to each traffic accident video, construct the original data matrix and calculate the covariance matrix of the original data matrix, and then execute step S32. Step S32: Calculate the eigenvalues ​​and corresponding eigenvectors based on the covariance matrix, and sort the eigenvalues ​​in descending order to obtain λ1, λ2…λ m And the eigenvectors a1, a2…a corresponding to the eigenvalues. m And the i-th principal component is Y i =Xa i (1≤i≤m), where X is the original data matrix, a i Given the i-th feature vector, proceed to step S33; Step S33: Based on the descendingly ordered feature values, calculate the principal component contribution rate and cumulative contribution rate according to the following formula. When the cumulative contribution rate reaches a preset threshold, select the first k principal components to form a new feature space, and then proceed to step S34: Where, λ k Let be the eigenvalue of the k-th principal component. The sum of all eigenvalues. Contribution rate of main components Cumulative contribution rate; Step S34: Based on the eigenvectors of the original data matrix, construct the loading matrix, and obtain the principal component scores of the first k principal components using the following formula: Among them, PC i w represents the principal component score of the i-th principal component, n is the total number of numerical features of the accident, and w ij x is the coefficient in the i-th row and j-th column of the characteristic load matrix. j Let be the numerical feature of the j-th accident.

3. The intelligent identification and classification method for dangerous driving scenarios according to claim 1, characterized in that, Step S4, based on the following steps, uses the elbow method and K-means clustering to obtain the optimal number of clusters and clustering results: Step S41: Based on the preset number of clusters k, calculate the sum of squared intra-cluster errors for different numbers of clusters k according to the following formula, and plot the curves of the sum of squared intra-cluster errors for different numbers of clusters k. Then proceed to step S42: Where WCSS is the sum of squared errors within a cluster, k is the number of clusters, and C i Let x represent the i-th cluster. j For the j-th data point belonging to the i-th cluster, c i Let be the centroid of the i-th cluster; Step S42: Based on the intra-cluster error sum of square curves under different cluster numbers k, select the value at the elbow of the intra-cluster error sum of square curve as the optimal cluster number k, that is, select the k value at the bend of the intra-cluster error sum of square curve as the optimal cluster number k. Step S43: Based on the optimal number of clusters k, randomly initialize k centroid vectors, and calculate the Euclidean distance from each data point to the k centroids according to the following formula. Assign each data point to the cluster containing the nearest centroid, where the data point is the principal component score. Then proceed to step S44: Where d(x) i ,c j ) is the Euclidean distance from each data point to the k centroids, x i For data points, c h Let p be the j-th centroid, and p be the data dimension; Step S44: Based on the data points of each cluster, recalculate the centroid vector of each cluster, with the centroid position being the mean of all data points within that cluster. Then, proceed to step S45: Among them, c j For each cluster's centroid vector, N h Let x be the number of data points in the j-th cluster. i For data points; Step S45: Repeat the data point allocation and centroid update steps until the centroid position converges or the predetermined maximum number of iterations is reached.

4. A system for executing the intelligent recognition and classification method for dangerous driving scenarios according to any one of claims 1 to 3, characterized in that, include: Data acquisition module: used to acquire a preset number of traffic accident videos; Feature extraction module: Used to extract the feature values ​​of each accident in traffic accident video that correspond to the preset accident features using deep learning and computer vision technology; Dimensionality reduction module: used to reduce the dimensionality of extracted features based on categorical principal component analysis; Clustering analysis module: Used to classify traffic accident videos into multiple different categories using the elbow method and K-Means clustering methods; Statistical Analysis Module: Used to determine the main characteristics of each category and the corresponding dangerous driving category.

Images