Urban road traffic risk analysis method considering time-space distribution of traffic safety events

Abstract

The invention discloses an urban road traffic risk analysis method considering space-time distribution of traffic safety events. The method comprises the following steps: collecting vehicle natural driving data in a certain period of time in a target area through a natural driving experiment; preprocessing the natural driving data of the vehicle; classifying images of a vehicle first view angle video in the preprocessed vehicle natural driving data through a pre-constructed urban road type division model, and dividing the vehicle natural driving data into a plurality of different urban road type data sets according to a classification result; performing traffic safety event extraction on the different urban road type data sets through a pre-constructed traffic safety event extraction model to obtain traffic safety event data under different urban road types; and performing space-time analysis on the traffic safety event data under different road types, and counting risk peak periods and high-risk areas under different road types. According to the method, high-driving-risk areas and high-driving-risk time periods of different urban road types can be efficiently found out.

Description

Technical Field

[0001] The present invention relates to the field of road transport driving risk research, and in particular to an urban road traffic risk analysis method and system that considers the temporal and spatial distribution of driving safety events. Background Art

[0002] With the acceleration of urbanization and the sharp increase in the number of motor vehicles, frequent traffic accidents have become an important issue affecting public safety and urban sustainable development. The complexity of urban transportation systems and the diversity of road types pose huge challenges to driving safety management. Traditional traffic risk assessment methods often rely on post-accident data analysis and lack the ability to monitor and predict driving risks on different road types in real time.

[0003] With the rapid development of intelligent transportation and autonomous driving technologies, various sensors on vehicles, such as GPS, accelerometers, and on-board cameras, can collect a large amount of driving data in real time. These data provide a rich resource for traffic risk assessment. However, due to the huge amount of data and complex structure, how to effectively integrate and analyze these data and mine the driving risk information contained therein has become an urgent problem to be solved.

[0004] In the study of driving risk, near-crash and crash events are generally used to characterize driving risk in the process of selecting characterization indicators. However, under normal driving conditions, crash is a low-probability event, and near-crash is a conflict state that requires rapid and drastic operations to avoid accidents. The frequency of occurrence is also low, so there are problems such as insufficient sample size and inability to better evaluate driving risks. Therefore, high-risk events marked by high acceleration / deceleration rates or other motion characteristics - driving safety events are proposed and used as risk characterization indicators for driving risk research.

[0005] However, in the existing research on driving risks using driving safety events, there is a lack of detailed distinction between different road types. There are significant differences in traffic flow characteristics and driving risks among different types of urban roads, such as expressways, intersections, and general roads. Expressways usually have higher speeds, but relatively smooth traffic flows; intersections are prone to collision accidents due to the intersection of traffic and traffic signal control; general roads have complex traffic environments, mixed pedestrians and non-motor vehicles, and higher driving risks. Therefore, the analysis of driving safety events based on road types is of great significance to improving the accuracy and effectiveness of traffic management.

[0006] In addition, traditional traffic risk assessment methods mostly focus on static data analysis and lack in-depth research on the temporal and spatial distribution patterns of traffic safety incidents. However, the occurrence of traffic safety incidents often has obvious temporal and spatial characteristics, such as specific time periods (morning and evening rush hours) and specific areas (schools, business districts) are high-incidence areas for accidents. Therefore, traffic risk analysis methods combined with temporal and spatial distribution patterns can help identify high-risk time periods and areas, formulate targeted prevention and intervention measures, and improve the level of traffic safety management.

[0007] Therefore, the urban road traffic risk analysis method that considers the temporal and spatial distribution of traffic safety incidents is not only a beneficial supplement to the traditional traffic risk assessment method, but also an important basis for realizing the application of intelligent transportation and autonomous driving technology.

[0008] Although a large number of sensor technologies and data collection methods have been applied in the current modern urban traffic system, the existing traffic risk assessment and management methods still have many problems and shortcomings: (1) Data silos and integration difficulties. Current traffic data are often scattered in different systems and devices, forming "data silos". These data include vehicle operation data, road monitoring videos, traffic accident records, etc., lack effective integration and analysis methods, resulting in low data utilization and difficulty in forming a comprehensive traffic risk assessment system database; (2) The current risk research based on driving safety events does not clearly distinguish between road types. Different types of roads have significant differences in traffic flow characteristics and driver behavior, but the existing methods for traffic risk assessment based on driving safety events often fail to fully consider the differences in road types, resulting in insufficient accuracy and pertinence of the assessment results. There is currently insufficient research on the spatiotemporal law analysis based on driving safety events. Existing research mostly focuses on the analysis of driving safety events in a single dimension (time or space), and fails to fully combine the spatiotemporal distribution law for comprehensive evaluation. This one-sided analysis method makes it difficult to fully identify traffic risks, affecting the effectiveness of prevention and intervention measures. (3) Research on deep learning technology in the field of driving risks still needs to be explored. Although deep learning technology has made significant progress in image recognition, natural language processing and other fields, its application in traffic risk assessment is still in the exploratory stage. How to use deep learning technology to efficiently process and accurately analyze complex traffic data is still a technical problem that needs to be solved urgently. Summary of the invention

[0009] The main purpose of this invention is to propose a traffic risk identification and analysis method that comprehensively considers road types, spatiotemporal distribution patterns and artificial intelligence technology, and clearly uses driving safety events as risk substitute indicators, which can not only improve data utilization efficiency and analysis accuracy, but also provide a scientific basis for traffic management and decision-making.

[0010] The technical solution adopted by the present invention is:

[0011] The present invention provides an urban road traffic risk analysis method considering the temporal and spatial distribution of traffic safety events, comprising the following steps:

[0012] The natural driving experiment collects the natural driving data of the vehicle in the target area within a certain period of time, including the first-person perspective video of the vehicle, the three-axis acceleration data sequence of the vehicle, the longitude and latitude coordinate sequence, and the speed sequence data;

[0013] Preprocess the vehicle's natural driving data;

[0014] Classify the images of the first-person perspective video of the vehicle in the preprocessed natural vehicle driving data through a pre-built urban road type classification model, and divide the natural vehicle driving data into several different urban road type data sets according to the classification results;

[0015] The driving safety event extraction model built in advance is used to extract driving safety events from different urban road type data sets to obtain driving safety event data under different urban road types;

[0016] Spatiotemporal analysis was conducted on traffic safety incident data on different road types, and peak risk periods and high-risk areas on different road types were statistically analyzed.

[0017] Following the above technical solution, the method further includes the steps of: performing spatial hot spot analysis on traffic safety event data under different urban road types, and statistically identifying hot spot areas in the spatial data.

[0018] Following the above technical solution, spatial hot spot analysis specifically identifies the significance of spatial clustering by comparing the local average with the global average, and finds out all the hot spots where significant traffic safety incidents occur.

[0019] Following the above technical solution, the steps of preprocessing the natural driving data of the vehicle specifically include:

[0020] Perform outlier processing to filter out vehicles with speeds greater than 90km / h and triaxial acceleration greater than 10m / s 2 and data with latitude and longitude coordinates of zero;

[0021] The vehicle's three-axis acceleration data sequence, longitude and latitude coordinate sequence, and vehicle speed sequence data are interpolated and resampled multiple times, and the sampling rate of each set of sequence data is kept consistent with the image frame rate of the vehicle's first-person perspective video to obtain the pre-processed vehicle's natural driving data.

[0022] Following the above technical solution, the driving safety event extraction model specifically finds the time points with the largest vehicle driving volatility in each continuous data segment in different urban road type data sets, and uses these time points as the centers of each driving safety event, and extracts data of a preset time length before and after the center of all driving safety events as corresponding driving safety event fragments.

[0023] Following the above technical solution, the vehicle driving volatility is determined by the vehicle horizontal acceleration, and the horizontal acceleration calculation method is as follows:

[0024] Among them, a′ x is the longitudinal acceleration of the vehicle, a′ y is the lateral acceleration of the vehicle.

[0025] According to the above technical solution, urban road types include urban expressways, urban road intersections and general sections of urban roads.

[0026] Following the above technical solution, the urban road type classification model is specifically constructed using the cross-layer manifold invariant pruning method CLMIP. During training, the model input is the first-person historical image of the vehicle passing through the urban road mark, and the model output is the data segment with the road type label.

[0027] Following the above technical solution, the driving safety event extraction model is specifically based on the vehicle operation history data set and is constructed using the CNN-GRU module and the XGBoost classifier.

[0028] The present invention also provides an urban road traffic risk analysis system taking into account the temporal and spatial distribution of traffic safety events, comprising:

[0029] The data acquisition module is used to collect the natural driving data of the vehicle in the target area within a certain period of time through the natural driving experiment, including the first-person perspective video of the vehicle, the three-axis acceleration data sequence of the vehicle, the longitude and latitude coordinate sequence, and the vehicle speed sequence data;

[0030] A preprocessing module, used for preprocessing the natural driving data of the vehicle;

[0031] An urban road type recognition module is used to classify the images of the vehicle first-person perspective video in the preprocessed vehicle natural driving data through a pre-built urban road type classification model, and divide the vehicle natural driving data into a number of different urban road type data sets according to the classification results;

[0032] A driving safety event extraction module is used to extract driving safety events from different urban road type data sets through a pre-built driving safety event extraction model to obtain driving safety event data under different urban road types;

[0033] The spatiotemporal analysis module is used to conduct spatiotemporal analysis on traffic safety event data on different road types, and to count the peak risk periods and high-risk areas on different road types.

[0034] The spatiotemporal analysis module includes a high-risk period analysis submodule, which is used to analyze the time periods of traffic safety incident data on different types of roads. According to the time span, the frequency of traffic safety incidents per hour in a day is counted and the peak risk period is found.

[0035] The spatiotemporal analysis module also includes a spatial hotspot analysis submodule, which is used to perform spatial hotspot analysis on traffic safety incident data under different urban road types and statistically identify hotspot areas in spatial data.

[0036] The present invention also provides a computer storage medium, which stores a computer program that can be executed by a processor. The computer program executes the urban road traffic risk analysis method that considers the temporal and spatial distribution of traffic safety events as described in the above technical solution.

[0037] The beneficial effects of the present invention are as follows: the urban road traffic risk analysis method considering the spatiotemporal distribution of driving safety events comprehensively considers the diversity of urban road types, processes and analyzes the driving data collected by different sensors of a single vehicle, and judges and extracts driving safety events under different urban road types based on deep learning technology, and then analyzes the spatiotemporal laws of driving safety events to find out high driving risk areas and high driving risk time periods of different urban road types.

[0038] Furthermore, the present invention adopts a variety of advanced technical means to significantly improve the accuracy and practicality of traffic risk analysis; by acquiring and integrating vehicle operation history data of different time spans and regions and video data collected by on-board cameras, it solves the problem of data isolation and difficulty in comprehensive analysis in traditional methods; natural driving experiments are used to collect vehicle natural driving data over a longer period of time, combined with data preprocessing techniques such as outlier removal and interpolation resampling, to ensure data consistency and reliability, providing a solid foundation for subsequent analysis.

[0039] Furthermore, the present invention also constructs an urban road type classification model through the cross-layer manifold invariant pruning method CLMIP method. The model can accurately distinguish different urban road types, laying a foundation for the analysis of traffic safety incidents on different types of roads; by classifying the preprocessed data, a data set of different road types is obtained, and a traffic safety event extraction model is constructed using the CNN-GRU module and the XGBoost classifier. The model can extract features from high-frequency kinematic data to achieve efficient classification and extraction of traffic safety incidents; by statistically analyzing the time periods of traffic safety incidents, the risk peak periods in different time periods of a day are identified, and combined with the spatial hotspot analysis method, the high-incidence areas of traffic safety incidents are determined; this analysis method of the spatiotemporal distribution law can not only comprehensively evaluate the traffic risks under different road types, but also provide a scientific basis for traffic management departments to formulate more accurate and effective traffic safety management measures.

[0040] In summary, the present invention provides an innovative traffic risk assessment method by comprehensively utilizing multi-source data, deep learning technology, and spatiotemporal distribution law analysis. This method not only improves the efficiency and accuracy of data processing and analysis, but also provides solid technical support for traffic safety management and scientific decision-making. By real-time monitoring and early warning of high-risk periods and areas, the present invention helps to improve the overall safety and operating efficiency of the traffic system, realizes dynamic risk management of driving safety incidents, and provides important support for the development of intelligent transportation and autonomous driving technology.

[0041] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0043] FIG1 is a flow chart of a method for analyzing urban road traffic risks taking into account the spatiotemporal distribution of traffic safety events according to an embodiment of the present invention.

[0044] Figure 2 A schematic diagram of road type classification using the CLMIP method according to an embodiment of the present invention;

[0045] Figure 3 A time regularity distribution diagram of driving safety events according to an embodiment of the present invention;

[0046] Figure 4A spatial hotspot distribution map of driving safety events according to an embodiment of the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0048] It should be noted that the illustrations provided in the embodiments of the present invention are only used to illustrate the basic concept of the present invention in a schematic manner. Therefore, the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout type may also be more complicated.

[0049] In the present invention, it is also necessary to explain that, if the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. appear, the orientation or position relationship indicated is based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application. In addition, if the terms "first" and "second" appear, they are only used for description and distinction purposes, and cannot be understood as indicating or implying relative importance.

[0050] Example 1

[0051] like Figure 1A As shown, the urban road traffic risk analysis method considering the spatiotemporal distribution of traffic safety events in this embodiment includes the following steps:

[0052] S1. Collecting vehicle natural driving data within a certain period of time in the target area through natural driving experiments, including vehicle first-person video, vehicle three-axis acceleration data sequence, latitude and longitude coordinate sequence, and vehicle speed sequence data;

[0053] S2, preprocessing the natural driving data of the vehicle;

[0054] S3, classifying the images of the first-person perspective video of the vehicle in the preprocessed natural vehicle driving data through a pre-constructed urban road type classification model, and dividing the natural vehicle driving data into a number of different urban road type data sets according to the classification results;

[0055] S4. Extracting driving safety events from different urban road type data sets through a pre-built driving safety event extraction model to obtain driving safety event data under different urban road types;

[0056] S5. Conduct spatiotemporal analysis on the traffic safety incident data on different road types, and calculate the peak risk periods and high-risk areas on different road types.

[0057] Step S5 can specifically analyze the time periods of traffic safety incident data on different road types, and count the frequency of traffic safety incidents per hour in a day according to the time span to find out the peak risk period; it can also identify the significance of spatial clustering by comparing the local average value of traffic safety incidents with the global average value, and find out all hot spots of traffic safety incidents with significance.

[0058] This embodiment comprehensively considers the diversity of urban road types, processes and analyzes the driving data collected by different sensors of a single vehicle, and judges and extracts driving safety events under different urban road types based on deep learning technology. It then analyzes the spatiotemporal laws of driving safety events to find out high driving risk areas and high driving risk time periods for different urban road types.

[0059] Step S2 specifically includes the following steps:

[0060] S21: Perform outlier processing to exclude vehicles with a speed greater than 90 km / h and a three-axis acceleration greater than 10 m / s 2 and data with latitude and longitude coordinates of zero;

[0061] S22, the vehicle three-axis acceleration data sequence, longitude and latitude coordinate sequence, and vehicle speed sequence data are interpolated and resampled multiple times, and the sampling rate of each set of sequence data is kept consistent with the image frame rate of the vehicle's first-person perspective video to obtain the pre-processed vehicle natural driving data.

[0062] Specifically, the driving safety event extraction model finds the time points with the largest vehicle driving volatility in each continuous data segment in different urban road type data sets, and uses these time points as the centers of each driving safety event. It extracts data of a preset time length before and after the center of all driving safety events as corresponding driving safety event fragments.

[0063] Furthermore, the vehicle driving volatility is determined by the vehicle horizontal acceleration, and the horizontal acceleration is calculated as follows:

[0064] Among them, a′ x is the longitudinal acceleration of the vehicle, a′ y is the lateral acceleration of the vehicle.

[0065] Furthermore, the urban road type classification model is specifically constructed using the Cross-Layer Manifold Invariance based Pruning method CLMIP (CLMIP for short). During training, the model input is the first-person historical image of vehicles passing through urban roads, and the model output is a data segment with road type labels. The model can accurately distinguish different types of urban roads, laying the foundation for the analysis of driving safety incidents on different types of roads; by classifying the preprocessed data, a data set of different road types is obtained.

[0066] Furthermore, the traffic safety event extraction model is specifically based on the vehicle operation history data set and is constructed using the CNN-GRU module and the XGBoost classifier. The model can extract features from high-frequency kinematic data to achieve efficient classification and extraction of traffic safety events; by statistically analyzing the time periods of traffic safety events, the risk peak periods in different time periods of the day are identified, and combined with the spatial hotspot analysis method, the high-incidence areas of traffic safety events are determined; this analysis method of spatiotemporal distribution rules can not only comprehensively evaluate traffic risks under different road types, but also provide a scientific basis for traffic management departments to formulate more accurate and effective traffic safety management measures.

[0067] Example 2

[0068] This embodiment is based on Embodiment 1, and provides implementation examples of various specific steps based on the embodiment 1.

[0069] See also Figure 1B As shown, the embodiment of the present application discloses a method for analyzing urban road traffic risks taking into account the spatiotemporal distribution of traffic safety events, including:

[0070] Step 1: Obtain a historical data set of vehicle operation in any area with a time span of no less than 1 year and a historical video data set collected from the first-person perspective of any vehicle-mounted camera. The historical data of vehicle operation includes a speed data sequence, a vehicle three-axis acceleration data sequence, and a driving safety event sequence data; the historical video data set collected from the first-person perspective of the vehicle contains images of different types of urban road scenes;

[0071] Step 2: Collect vehicle natural driving data in the target area for a long period of time through natural driving experiments to generate a natural driving experiment dataset, including vehicle first-person video images, vehicle three-axis acceleration sequence data, longitude and latitude coordinate sequence data, and vehicle speed sequence data; Step 3: Preprocess the acquired vehicle operation history dataset, vehicle first-person historical video dataset, and natural driving experiment dataset;

[0072] Step 4: Based on the historical video dataset collected from the first-person perspective of the vehicle, the urban road type classification model is constructed using the CLMIP method;

[0073] Step 5: Using the urban road type classification model, classify the first-person perspective video images of vehicles in the preprocessed natural driving experiment dataset, and divide the preprocessed natural driving experiment dataset into datasets of different road types;

[0074] Step 6: Based on the vehicle operation history dataset, a driving safety event classification model is constructed using the CNN-GRU module and the XGBoost classifier;

[0075] Step 7: Using the driving safety event extraction model, the driving safety event extraction is performed on each urban road type data set to obtain the driving safety event data sets under different urban road types and the combined total driving safety event data set;

[0076] Step 8: Count the timestamps of the driving safety event fragments in the driving safety event data set on different road types, and count the frequency of driving safety events every hour in a day according to the time span of the data set to find the peak risk period;

[0077] Step 9: Perform spatial hot spot analysis on the total traffic safety incident dataset to find statistically significant traffic safety incident hot spots;

[0078] Specifically:

[0079] In step 1, the vehicle operation history data and the historical video dataset collected from the first person perspective are obtained from open source datasets on various websites;

[0080] In step 2, the data set obtained by the natural driving experiment covers a period of 10 months. The acquisition process and description are as follows: the natural driving experiment collects the natural driving data of the vehicle in the target area for a long period of time, the front camera of the vehicle ADAS (Advanced Driver Assistance Systems) collects the first-person video data of the vehicle, and the vehicle gyroscope collects the three-axis acceleration data sequence of the vehicle. (longitudinal acceleration), (lateral acceleration), (vertical acceleration), GPS (Global Position System) collects longitude and latitude data series (Coordinate point), vehicle speed data sequence collected by vehicle CAN (Controller Area Network) signal The specific data collection equipment and data description are shown in Table 1.

[0081] Table 1 Description of experimental datasets

[0082]

[0083] In step 3, the preprocessing process includes:

[0084] The vehicle three-axis acceleration sequence in the vehicle operation history data obtained from the open source dataset

[0085] Speed ​​sequence and driving safety event data series The missing data are filled using linear interpolation, and the sampling rate is unified to 30Hz using cubic spline interpolation. The driving safety event data sequence refers to the data label sequence that characterizes whether a driving safety event has occurred. The input of the driving safety event extraction model mentioned here is labeled with driving safety events. Since the sequence is continuous and all have driving safety event labels, the driving volatility is the largest in order to find the center point of this driving safety event (for example, there is only one center point in the data segment with a driving safety event label within every 1.7s, and it is not necessarily at the 0.8s in the event process, that is, the driving volatility is the largest in order to find the center point of this driving safety event). The maximum volatility point is only used to find the time point when the event occurred, in order to prevent the occurrence of cross-time events in the event statistics process. The actual center is in which segment, which needs to be determined by calculating the maximum volatility to find the center point); the driving event center is selected according to the moment when the vehicle driving volatility in the driving safety event data sequence is the largest. The 25 time series points before the driving safety event center are used as the starting point, and the 25 time series points after the driving safety event center are used as the end point. This data sequence is a driving event segment. Since the sampling rate is 30Hz, the duration of a driving safety event segment is 1.7s. Driving volatility is calculated using horizontal acceleration as an indicator, and the calculation method is shown in formula (1);

[0086]

[0087] The preprocessed data set is a multidimensional vector of the form (N, 51, 4), which indicates that the Nth driving safety event contains 51 data frames, including 4 dimensions, longitudinal acceleration a′ x , lateral acceleration a′ y , vertical acceleration a′ z and vehicle speed v′.

[0088] The initial time series data obtained from the natural driving experiment was processed for outliers, and vehicles with speeds greater than 90 km / h and triaxial accelerations greater than 10 m / s were screened out. 2The data with zero latitude and longitude coordinates are converted into images from the first-person perspective of the vehicle, and the size is adjusted to 320×240. The vehicle three-axis acceleration sequence data, vehicle speed sequence data, and longitude and latitude sequence data are resampled through cubic spline interpolation, and the sampling rate of each set of sequence data is kept consistent with the image frame rate in the first-person perspective video of the vehicle, that is, adjusted to 30Hz. The preprocessed natural driving experiment data set includes: preprocessed three-axis acceleration data sequence A x1 ,A x2 ,A x3 ,…,A xn , A y1 ,A y2 ,A y3 ,…,A yn , A z1 ,A z2 ,A z3 ,…,A zn , the preprocessed longitude and latitude coordinate data sequence C1, C2, C3, ..., C n , the pre-processed vehicle speed data sequence V1, V2, V3, ..., V n .

[0089] The size of the vehicle first-person perspective image with road type labels obtained from the online dataset was adjusted to 320×240, where the road type labels are "urban expressway", "urban road intersection" and "general section of urban road".

[0090] In step 4, the first-person view image set of the vehicle camera with road type labels is used as the training set, and the urban road type classification model is constructed using the CLMIP method. The detailed schematic diagram of this method is shown in Figure 2 As shown. It should be noted that the hyperparameters introduced in the pruning process of the original deep CNN include the pruning rate ξ, the pruning step ζ and the threshold γ, which are set to 0.4, 2 and 0.25 respectively; the hyperparameters used in the fine-tuning process include the learning rate α, the batch size M and the weight decay λ, which are set to 0.01, 64 and 0.0005 respectively; when the training loss no longer decreases, the learning rate is reduced by 10 times; the contents of step 4 mainly include:

[0091] Initial data input. Mainly includes training set And the input of the initial parameter θ0, where x, y represent the input image and the corresponding road type label, respectively, and N is the number of samples in the batch. W0 and b0 are the initial weight and initial bias of the first layer respectively;

[0092] The initial convolutional neural network (CNN) model is pre-trained on the training set S until convergence. In this process, the loss function L is used to optimize the network parameter θ. The calculation process is shown in formulas (2) to (3):

[0093]

[0094]

[0095] Where C is the number of categories, y ij is a binary indicator (1 if class j is the correct classification for sample i, 0 otherwise), f(x i ) j is the probability that the model predicts that sample i belongs to category j, is the Frobenius norm of the weight of the lth layer, and λ is the weight of the regularization term; is the network parameter of layer l in iteration t, α is the learning rate, M is the size of the samples in the mini-batch, is the loss function L with respect to the parameters The gradient of

[0096] According to the reconstruction error of the data manifold, the redundant parameters in the network are removed. The data manifold descriptor of each layer is calculated, which represents the embedding of high-dimensional data in low-dimensional space. The simplified representation of the data is obtained by mapping the data from high-dimensional space to low-dimensional space. For the lth layer, the data manifold descriptor is The calculation method is shown in formula (4):

[0097]

[0098] in, is the feature map of the ith sample in layer l, ‖·‖ F is the Frobenius norm, M is the batch size;

[0099] For the fully connected layer, the output dimension is different from that of the convolutional layer. The convolutional layer changes from a three-dimensional tensor to a two-dimensional vector. The updated flow descriptor is calculated according to formula (5):

[0100]

[0101] Cosine similarity is used to measure the difference between manifolds across different dimensions. The calculation method is shown in formula (6):

[0102]

[0103] in, and are data manifolds of two different dimensions;

[0104] After obtaining the manifold descriptors of all types of layers in the network, the average reconstruction error ψ of the manifold in all samples before and after network pruning is calculated. The importance of the network weights is estimated based on the minimum-maximum normalized reconstruction error of the manifold and ranked according to the importance. The calculation method is shown in formula (7):

[0105]

[0106] The average reconstruction error is the average value of φ, which is expressed as the manifold multi-similarity of the pruned network;

[0107] Based on the above ranking criteria, the smaller weights in all convolutional layers and fully connected layers of the deep CNN are pruned. Model pruning requires at least Before the next fine-tuning of the pruned network, the weight matrix is ​​first updated by dot multiplication with the suppression mask. The update process is shown in formula (8):

[0108]

[0109] Where θ is the parameter set of the neural network at the beginning of the current pruning round, ° represents the element-wise operation for computing the Hadamard product, redundant weights are removed by multiplying by 0, and the remaining weights are retained by multiplying by 1;

[0110] Fine-tune the pruned network structure. Use mini-batch stochastic gradient descent (SGD) to fine-tune the pruned network to restore the accuracy that may be lost due to pruning. The fine-tuning process uses formula (6) as the update rule;

[0111] The fine-tuning process is terminated when the proportion of parameters removed from the model is greater than the pruning rate and the loss of the validation set is less than the threshold. The fine-tuning cycle process is shown in formula (9):

[0112]

[0113] Among them, γ is the hyperparameter threshold used to determine the usability of deep CNN;

[0114] After the fine-tuning process terminates, model training is complete.

[0115] In step 5, the urban road type classification model constructed in step 4 is used to divide the preprocessed natural driving experiment dataset into an urban expressway dataset, an urban road intersection dataset, and an urban road general section dataset, and each road type dataset is in the form of (N, 51, 4).

[0116] In step 6, the preprocessed vehicle operation history data is used as input, and the CNN-RGU feature engineering module is used to extract features from the high-frequency kinematic data, and the XGBoost classifier is used for classification to build a driving safety event extraction model. It should be noted that for the CNN-RGU module, when the training rounds are less than 120, the learning rate is 0.001, when the training rounds are greater than 120, the learning rate becomes 0.0001, the batch sample data size is 4000, and the optimizer is Adam; for the XGBoost module, the learning rate is 0.005, the minimum weight of the leaf node is 1.7, the maximum depth is 12, the regularization parameters λ1 and λ2 are 0.1 respectively, the sub-sample ratio is 0.7, and the sub-sample ratio of the column when constructing each tree is 0.7; the content of this step is as follows:

[0117] In the CNN module, for feature extraction, two one-dimensional convolutional layers are used to capture the relationship between each kernel. The convolutional layer is followed by maximum pooling to capture extreme driving behavior changes. The outputs of the convolutional layer and the pooling layer are shown in formulas (10) to (11):

[0118]

[0119] in, represents the output of the jth filter of the convolutional layer r at time t, F(·) is the activation function, and L r is the total number of filters in the convolutional layer r, w and b represent weights and biases respectively, represents the output of the i-th pooling layer at time t, which is used to capture extreme driving behavior changes;

[0120] In the GRU module, the update gate With reset gate These two gating units jointly determine the hidden state at each moment How to use the hidden state based on the current input and the previous moment Update, the calculation methods are shown in formula (12) and formula (13) respectively:

[0121]

[0122] Among them, σ(·) is the Sigmoid activation function, is the output of the (r-1)th GRU layer at time t, expressed as and Represent the weight matrices from the hidden state adjusted by the input and reset gate to the candidate hidden state, is the bias term, ⊙ represents the multiplication between elements;

[0123] The hidden state at the previous moment and the candidate hidden state are weightedly combined through the update gate to obtain the hidden state at the current moment. The calculation method is shown in formula (14):

[0124]

[0125] The output of the GRU is the feature fed into the fully connected layer for classification. The fully connected layer consists of four dense layers, each followed by batch normalization and dropout methods to improve efficiency and generalization. The last layer is a SoftMax classifier that determines whether the event is a driving safety incident or normal driving. The last moment h of the last GRU layer T The output vector is the extracted classification feature, which is the high-frequency velocity and three-axis acceleration data. This output is input into the XGBoost classifier as a new training set for classification and prediction training;

[0126] XGBoost is an ensemble learning algorithm based on gradient boosting. It minimizes a loss function by constructing multiple decision trees. Its objective function and final prediction result are shown in formulas (15) to (16):

[0127]

[0128] Where N is the number of samples, K is the number of categories, and y i is the true label of the i-th sample, is the probability that the model predicts that the i-th sample belongs to the k-th class, w x is the weight of the kth decision tree, and its value represents G i is the gradient of the ith sample, λ k is the regularization parameter, is the predicted label of the i-th sample, T is the number of weak classifiers, and a tk is the prediction score of the tth weak classifier for the kth class;

[0129] After the training of the XGBoost classifier is completed, the driving safety event classification model is finally obtained.

[0130] In step 7, the traffic safety event extraction model is used to extract traffic safety events from three urban road type data sets: urban expressways, urban road intersections, and general urban road sections. The input vector format is (N, 51, 4), and the traffic safety event data sets under different urban road types and the merged total traffic safety event data set are obtained.

[0131] In step 8, within the time span of 1 year in the data set, the frequency of traffic safety events every half hour within 24 hours of the three traffic safety event data sets is counted and the peak risk period is found. The results are as follows: Figure 3 shown.

[0132] In step 9, by calculating And normalize it to find all the hot spots of significant traffic safety events. The hot spot diagram is as follows: Figure 4 The calculation method is shown in formula (17):

[0133]

[0134] Among them, x j is the number of driving safety events at location j, w ij is the spatial weight between locations i and j, n is the total number of locations in the study area, S x is the sum of driving safety events at all locations, is the global average number of accidents;

[0135] The calculated The values ​​were normalized by Z-score, with the minimum Z-score corresponding to a P value less than 0.05 as the threshold, denoted as Z * , greater than Z * The hot spots are statistically significant, and the larger the Z score, the tighter the hot spot clustering. According to the above rules, find all the significant hot spots of traffic safety events.

[0136] Example 3

[0137] The urban road traffic risk analysis system considering the spatiotemporal distribution of traffic safety events in this embodiment is mainly used to implement the above method embodiment, and the system includes:

[0138] The data acquisition module is used to collect the natural driving data of the vehicle in the target area within a certain period of time through the natural driving experiment, including the first-person perspective video of the vehicle, the three-axis acceleration data sequence of the vehicle, the longitude and latitude coordinate sequence, and the vehicle speed sequence data;

[0139] A preprocessing module, used for preprocessing the natural driving data of the vehicle;

[0140] An urban road type recognition module is used to classify the images of the vehicle first-person perspective video in the preprocessed vehicle natural driving data through a pre-built urban road type classification model, and divide the vehicle natural driving data into a number of different urban road type data sets according to the classification results;

[0141] A driving safety event extraction module is used to extract driving safety events from different urban road type data sets through a pre-built driving safety event extraction model to obtain driving safety event data under different urban road types;

[0142] The spatiotemporal analysis module is used to conduct spatiotemporal analysis on traffic safety event data on different road types, and to count the peak risk periods and high-risk areas on different road types.

[0143] Furthermore, the spatiotemporal analysis module includes a high-risk period analysis submodule and a spatial hotspot analysis submodule, wherein the high-risk period analysis submodule is used to analyze the time periods of occurrence of traffic safety incident data on different types of roads respectively, and according to the time span, count the frequency of occurrence of traffic safety incidents per hour in a day and find out the peak risk period.

[0144] The spatial hotspot analysis submodule is used to perform spatial hotspot analysis on traffic safety incident data under different urban road types and statistically identify hotspot areas in spatial data.

[0145] Each module is mainly used to implement each step of the method embodiment, which will not be described one by one here.

[0146] The present application also provides a computer-readable storage medium, such as a flash memory, a hard disk, a multimedia card, a card-type memory (e.g., an SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a disk, an optical disk, a server, an App application store, etc., on which a computer program is stored, and when the program is executed by a processor, a corresponding function is realized. When the computer-readable storage medium of this embodiment is executed by a processor, the urban road traffic risk analysis method considering the spatiotemporal distribution of driving safety events of the method embodiment is realized.

[0147] It should be pointed out that, according to the needs of implementation, the various steps / components described in this application can be split into more steps / components, and two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of the present invention.

[0148] The order of execution of each step in the above embodiment does not mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present application.

[0149] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all these improvements and changes should fall within the scope of protection of the appended claims of the present invention.

Claims

1. A method for analyzing urban road traffic risk considering the spatiotemporal distribution of traffic safety events, characterized in that: The following steps are involved: The natural driving experiment collects the natural driving data of the vehicle in the target area within a certain period of time, including the first-person perspective video of the vehicle, the three-axis acceleration data sequence of the vehicle, the longitude and latitude coordinate sequence, and the speed sequence data; Preprocess the vehicle's natural driving data; Classify the images of the first-person perspective video of the vehicle in the preprocessed natural vehicle driving data through a pre-built urban road type classification model, and divide the natural vehicle driving data into several different urban road type data sets according to the classification results; The driving safety event extraction model built in advance is used to extract driving safety events from different urban road type data sets to obtain driving safety event data under different urban road types; Spatiotemporal analysis was conducted on traffic safety incident data on different road types, and peak risk periods and high-risk areas on different road types were statistically analyzed.

2. The urban road traffic risk analysis method considering the spatiotemporal distribution of traffic safety events according to claim 1 is characterized in that: The analysis of high-risk areas specifically identifies the significance of spatial clustering by comparing the local average value of traffic safety incidents with the global average value, and finds all hot spots where significant traffic safety incidents occur.

3. The urban road traffic risk analysis method considering the spatiotemporal distribution of traffic safety events according to claim 1 is characterized in that: The steps for preprocessing the vehicle's natural driving data specifically include: Perform outlier processing to filter out vehicles with speeds greater than 90km / h and triaxial acceleration greater than 10m / s 2 and data with latitude and longitude coordinates of zero; The vehicle's three-axis acceleration data sequence, longitude and latitude coordinate sequence, and vehicle speed sequence data are interpolated and resampled multiple times, and the sampling rate of each set of sequence data is kept consistent with the image frame rate of the vehicle's first-person perspective video to obtain the pre-processed vehicle's natural driving data.

4. The urban road traffic risk analysis method considering the spatiotemporal distribution of traffic safety events according to claim 1 is characterized in that: The driving safety event extraction model specifically finds the time points with the largest vehicle driving volatility in each continuous data segment in different urban road type data sets, and uses these time points as the centers of each driving safety event. It then extracts data of a preset time length before and after the center of all driving safety events as the corresponding driving safety event fragments.

5. The urban road traffic risk analysis method considering the spatiotemporal distribution of traffic safety events according to claim 4 is characterized in that: The vehicle's driving volatility is determined by the vehicle's horizontal acceleration, and the calculation method of the horizontal acceleration is as follows: Among them, a′ x is the longitudinal acceleration of the vehicle, a′ y is the lateral acceleration of the vehicle.

6. The urban road traffic risk analysis method considering the spatiotemporal distribution of traffic safety events according to claim 1 is characterized in that: Urban road types include urban expressways, urban road intersections and general sections of urban roads.

7. The urban road traffic risk analysis method considering the spatiotemporal distribution of traffic safety events according to any one of claims 1 to 6, characterized in that: The urban road type classification model is specifically constructed using the cross-layer manifold invariant pruning method CLMIP. During training, the model input is the first-person historical image of vehicles passing through urban roads, and the model output is data segments with road type labels.

8. The urban road traffic risk analysis method considering the spatiotemporal distribution of traffic safety events according to any one of claims 1 to 6, characterized in that: The driving safety event extraction model is specifically based on the vehicle operation history dataset and is constructed using the CNN-GRU module and XGBoost classifier.

9. An urban road traffic risk analysis system considering the temporal and spatial distribution of traffic safety events, characterized in that: include: The data acquisition module is used to collect the natural driving data of the vehicle in the target area within a certain period of time through the natural driving experiment, including the first-person perspective video of the vehicle, the three-axis acceleration data sequence of the vehicle, the longitude and latitude coordinate sequence, and the vehicle speed sequence data; A preprocessing module, used for preprocessing the natural driving data of the vehicle; An urban road type recognition module is used to classify the images of the vehicle first-person perspective video in the preprocessed vehicle natural driving data through a pre-built urban road type classification model, and divide the vehicle natural driving data into a number of different urban road type data sets according to the classification results; A driving safety event extraction module is used to extract driving safety events from different urban road type data sets through a pre-built driving safety event extraction model to obtain driving safety event data under different urban road types; The spatiotemporal analysis module is used to conduct spatiotemporal analysis on traffic safety event data on different road types, and to count the peak risk periods and high-risk areas on different road types.

10. A computer storage medium storing a computer program executable by a processor, wherein the computer program executes the urban road traffic risk analysis method considering the spatiotemporal distribution of traffic safety events as described in claims 1-8.

Images