CN114139984B - Urban traffic accident risk prediction method based on flow and accident cooperative sensing - Google Patents
Urban traffic accident risk prediction method based on flow and accident cooperative sensing Download PDFInfo
- Publication number
- CN114139984B CN114139984B CN202111470417.0A CN202111470417A CN114139984B CN 114139984 B CN114139984 B CN 114139984B CN 202111470417 A CN202111470417 A CN 202111470417A CN 114139984 B CN114139984 B CN 114139984B
- Authority
- CN
- China
- Prior art keywords
- time
- node
- flow
- accident
- traffic
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 47
- 206010039203 Road traffic accident Diseases 0.000 title claims abstract description 40
- 239000011159 matrix material Substances 0.000 claims abstract description 32
- 238000013528 artificial neural network Methods 0.000 claims abstract description 19
- 230000002159 abnormal effect Effects 0.000 claims abstract description 9
- 238000012549 training Methods 0.000 claims abstract description 9
- 230000036962 time dependent Effects 0.000 claims abstract description 8
- 230000004927 fusion Effects 0.000 claims abstract description 5
- 238000006467 substitution reaction Methods 0.000 claims abstract description 4
- 230000006870 function Effects 0.000 claims description 15
- 238000004364 calculation method Methods 0.000 claims description 11
- 230000005856 abnormality Effects 0.000 claims description 5
- 230000008447 perception Effects 0.000 claims description 4
- 208000027418 Wounds and injury Diseases 0.000 claims description 3
- 230000006378 damage Effects 0.000 claims description 3
- 230000034994 death Effects 0.000 claims description 3
- 231100000517 death Toxicity 0.000 claims description 3
- 208000014674 injury Diseases 0.000 claims description 3
- 230000000694 effects Effects 0.000 abstract description 13
- 230000003068 static effect Effects 0.000 abstract description 3
- 238000000605 extraction Methods 0.000 abstract 1
- 238000002474 experimental method Methods 0.000 description 6
- 238000006243 chemical reaction Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 238000010801 machine learning Methods 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 238000013527 convolutional neural network Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000013136 deep learning model Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- YHXISWVBGDMDLQ-UHFFFAOYSA-N moclobemide Chemical compound C1=CC(Cl)=CC=C1C(=O)NCCN1CCOCC1 YHXISWVBGDMDLQ-UHFFFAOYSA-N 0.000 description 2
- 238000007781 pre-processing Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000003066 decision tree Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 230000000977 initiatory effect Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 238000000638 solvent extraction Methods 0.000 description 1
- 230000002195 synergetic effect Effects 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0635—Risk analysis of enterprise or organisation activities
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/10—Services
- G06Q50/26—Government or public services
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Human Resources & Organizations (AREA)
- Data Mining & Analysis (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Economics (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Strategic Management (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Tourism & Hospitality (AREA)
- Marketing (AREA)
- Educational Administration (AREA)
- Entrepreneurship & Innovation (AREA)
- Development Economics (AREA)
- General Business, Economics & Management (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Primary Health Care (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Game Theory and Decision Science (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Traffic Control Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
A city traffic accident risk prediction method based on flow and accident cooperative sensing belongs to the technical field of traffic accident early warning. The method aims at solving the problem that the prediction effect of constructing the map neural network by using the static map adjacent matrix is not good enough because the existing traffic accident risk prediction does not have the depth combined with traffic flow information. The method comprises the steps that a node time-varying feature matrix is obtained by a node time-varying feature input module; respectively carrying out feature extraction on the node time-varying feature matrix by an event dynamic graph neural network and an abnormal flow dynamic graph neural network to obtain accident risk features and flow risk features; capturing time-dependent information by a time-dependent module to obtain a space-time mixing characteristic; and carrying out feature fusion on the time-space mixed features and the global time-varying feature matrix by the collaborative sensing module to obtain an accident risk prediction value, and calculating by combining the accident risk substitution value to obtain a loss function for model training. The traffic accident risk prediction method is used for traffic accident risk prediction.
Description
Technical Field
The invention relates to an urban traffic accident risk prediction method based on flow and accident cooperative sensing, and belongs to the technical field of traffic accident early warning.
Background
In recent years, with the promotion of industrialization and the improvement of the living standard of people, the scale and population of cities are continuously increased, and the daily travel of people is greatly facilitated by a plurality of transportation means and developed urban road networks. However, occasional traffic accidents, abnormal traffic congestion and serious road network failures also bring huge life and property losses to people. How to reduce the occurrence of traffic accidents has become one of the important challenges in the field of global public transportation security.
The occurrence of traffic accidents has higher contingency and randomness, and is influenced by superposition of various factors such as vehicles, drivers, pedestrians, roads, weather and the like. Because of objective condition limitations, people cannot collect all influencing factors and data when an accident happens together for modeling and analysis, so that it is difficult to individually predict when and where an accident happens to a vehicle. But has relative smoothness and periodicity for certain specific areas, and therefore the number of incidents that occur over a certain fixed length of time remains generally steady and inherently regular. Therefore, the accident occurrence and distribution rules of the area can be mined by collecting a large amount of historical space-time data and traffic accident data of the area, so that the traffic accident risk in the future time area can be predicted.
At present, the methods for researching the traffic accident risk problem mainly comprise a traditional machine learning method, a deep learning method and the like. Compared with the traditional machine learning model, the deep learning model effect is greatly improved, but the CNN-based and RNN-based models are concentrated on local correlation, are limited by local invariance and the limited length of an aggregator, and therefore lack strong exploratory capacity for the global. The GNN-base model can aggregate information similar to road network characteristics and learn the occurrence mode of accidents in two dimensions in time space at the same time, but most of the prior works only input traffic as common characteristics of nodes, do not deeply input important information of the traffic as an initiating factor of the accidents, and neglect the importance of the traffic. The traffic in the road network and the occurrence mode of traffic accidents have certain correlation, the correlation caused by the traffic and the road network accidents is not captured in the prior work, massive traffic data per se contain information of human group movement, a complex interaction relationship exists between the traffic and the events in the road network, the drastic change of the traffic reflects the change level of the road network to a great extent, and the traffic in the accident-prone area is large.
Disclosure of Invention
Aiming at the problems that no depth is combined with traffic flow information in the existing traffic accident risk prediction, a map neural network is constructed by using a static map adjacency matrix, and the prediction effect is not good enough, the invention provides the urban traffic accident risk prediction method based on the cooperative sensing of flow and accidents.
The invention relates to a city traffic accident risk prediction method based on flow and accident cooperative sensing, which comprises the following steps,
When a time segment t is obtained by a node time-varying feature input module, node time-varying feature matrixes corresponding to the first continuous m time segmentsV is a node area;
node time-varying feature matrix by event dynamic graph neural network Extracting traffic accident perception characteristics to obtain accident risk characteristics based on flow distribution similarity;
Node time-varying feature matrix based on abnormal flow dynamic graph neural network Extracting traffic flow sensing characteristics to obtain flow risk characteristics based on flow anomaly;
Capturing time-dependent information of accident risk features and flow risk features by a time-dependent module to obtain space-time mixing features;
The collaborative perception module performs feature fusion on the time-space mixed feature and the global time-varying feature matrix { G t,Gt-1...Gt-m+1 }, and an accident risk prediction value of the time segment t is obtained
Accident risk prediction value according to time segment t by output moduleAnd accident risk replacement valueCalculating to obtain a loss function for model training; the trained model is used for predicting the accident risk of the next time slice;
The accident risk substitution value The calculation method of (1) is as follows:
In the middle of The value is replaced for the risk of accident for the ith node area,For the accident risk realism value of the ith node area, α 1 is a first weight, T is a time window, Δ is a minimum value greater than 0, and α 2 is a second weight.
According to the urban traffic accident risk prediction method based on the cooperative sensing of the flow and the accident,
The node area V is: v= { V 1,v2...,vN},vi represents the i-th node area, i=1, 2,3, … …, N is the total number of node areas;
the node area is a grid area obtained by rasterizing the target city area.
According to the urban traffic accident risk prediction method based on the cooperative sensing of the flow and the accident,
Accident risk true value of ith node area in time slice tThe calculation method of (1) is as follows:
Where R (v i) is all incidents occurring in node region v i at time segment t, inS j is the j-th incident in all incidents R (v i), p j is the number of injuries to the j-th incident in all incidents R (v i), q j is the number of deaths to the j-th incident in all incidents R (v i), R j is the number of vehicles involved in the j-th incident in all incidents R (v i), and a 3 is a third weight.
According to the urban traffic accident risk prediction method based on the cooperative sensing of the flow and the accident,
When the time segment t is, the node time-varying feature matrix of the node area v i The method comprises the following steps:
wherein the method comprises the steps of For traffic of node region v i at time segment t,For the vehicle speed of the node area v i at time segment t,For the casualties of node area v i at time segment t,The vehicle is involved in the time segment t for the node region v i.
According to the urban traffic accident risk prediction method based on the cooperative sensing of the flow and the accident,
At time segment t, the global time-varying feature matrix G t of node region v i is:
Gt＝{Wt,Qt}，
w t is the weather signal of time segment t, and Q t is the timing signal of time segment t.
According to the urban traffic accident risk prediction method based on the cooperative sensing of the flow and the accident,
The calculation method of the flow distribution similarity comprises the following steps:
In the middle of For the flow distribution similarity of the node region v i and the node region v j at the time segment t, i, j=1, 2..n; for traffic of node region v i over consecutive m time segments, Traffic for node region v j for consecutive m time segments;
the corresponding traffic similarity adjacency matrix S t is:
according to the urban traffic accident risk prediction method based on the cooperative sensing of the flow and the accident,
The calculating method of the flow anomaly degree comprises the following steps:
In the middle of For the traffic anomalies of node region v i and node region v j at time segment t,For the traffic attenuations of node region v i and node region v j at time segment t,The flow distribution similarity of the node area v i and the node area v j in the time segment t-1 is obtained, and θ is a flow abnormality threshold;
the corresponding flow anomaly adjacency matrix Γ t is:
according to the urban traffic accident risk prediction method based on the cooperative sensing of the flow and the accident,
Flow attenuation degreeThe calculation method of (1) comprises the following steps:
according to the urban traffic accident risk prediction method based on the cooperative sensing of the flow and the accident,
The method for calculating the loss function comprises the following steps:
Loss is a Loss function, ω 1 is a first constant, ω 2 is a second constant, ω 2>ω1 >1, and top27 is a set of high risk regions.
The invention has the beneficial effects that: the invention takes the flow and the accident into consideration in a synergic way and takes the flow and the accident as the consideration factors of the final accident occurrence. The invention accommodates the flow-accident interaction relationship into the graph neural network on the basis of the data preprocessing of the label conversion strategy based on risk priori.
The invention adopts a brand-new traffic-accident cooperative sensing framework to perfect traffic accident risk prediction tasks, and provides a brand-new risk priori label conversion strategy in a data preprocessing part, compared with the traditional model, the invention takes the influence of traffic important factors, namely traffic, on accident risk into consideration more deeply.
In order to overcome the shortcomings of the traditional static GCN network, the invention uses two different adjacency matrix construction strategies to construct the dynamic GCN network, which are respectively used for capturing the characteristic modes of the event and the traffic in the induced accident.
The method has wide application range. Experiments prove that the finally obtained model has higher feasibility and accuracy in predicting the risk of traffic accidents, and can achieve the optimal prediction effect by taking the potential influence of flow fluctuation on the accidents into consideration from the time-space characteristics.
Drawings
FIG. 1 is a schematic flow chart of an urban traffic accident risk prediction method based on cooperative sensing of flow and accidents;
fig. 2 is a schematic diagram of urban area rasterization.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
It should be noted that, without conflict, the embodiments of the present invention and features of the embodiments may be combined with each other.
The invention is further described below with reference to the drawings and specific examples, which are not intended to be limiting.
The invention provides a city traffic accident risk prediction method based on flow and accident cooperative sensing, which is shown in a specific embodiment with reference to fig. 1, and comprises the following steps,
When the time segment t is obtained by the node time-varying feature input module 100, node time-varying feature matrixes corresponding to the first m time segments are obtainedV is a node area;
node time-varying feature matrix by event dynamic graph neural network 200 Extracting traffic accident perception characteristics to obtain accident risk characteristics based on flow distribution similarity;
Node time-varying feature matrix of abnormal flow dynamic graph neural network 300 Extracting traffic flow sensing characteristics to obtain flow risk characteristics based on flow anomaly;
Capturing time-dependent information of accident risk features and flow risk features by a time-dependent module 400 to obtain space-time mixing features; the time dependence module 400 is implemented by adopting a time convolution network TCN;
The collaborative awareness module 500 performs feature fusion on the time-space mixed feature and the global time-varying feature matrix { G t,Gt-1...Gt-m+1 }, and obtains the accident risk prediction value of the time segment t
Accident risk prediction value from time segment t by output module 600And accident risk replacement valueCalculating to obtain a loss function for model training; after training, the final model parameters are stored and imported to obtain the final prediction model. The trained model is used for predicting the accident risk of the next time slice;
The accident risk substitution value The calculation method of (1) is as follows:
In the middle of The value is replaced for the risk of accident for the ith node area,For the accident risk realism value of the ith node area, α 1 is a first weight, T is a time window, Δ is a minimum value greater than 0, and α 2 is a second weight.
In order to overcome the unbalance of accidents and accidents in the sample, the actual value of the accident risk is replaced and converted into the replacement value of the accident risk. In the model training process, if the accident risk true value is zero, the accident risk true value is replaced by a logarithmic function of the average value of the historical accidents in the whole sample of the area. Through the conversion, even if most areas have no accident in a certain time period, the label values of the areas have distinguishable accident potential risk probabilities after mapping, so that the neural network can better distinguish the risk differences of different areas, and the prediction effect of the model is improved. The model is trained by obtaining a loss function, namely completing the forward propagation process of the model, and then adopting the determined loss function.
In the present embodiment, first, urban areas are rasterized, each urban subarea is regarded as a node of a space-time diagram, and information such as accidents, traffic, and vehicle speed of the subarea is used as node characteristics of the subarea. And then according to the flow data of different time periods of each node, calculating a flow similarity matrix and a flow abnormality matrix as two adjacent matrixes of a time-space diagram, further constructing two dynamic graphic neural networks, respectively capturing different characteristic modes of new accidents possibly caused by events and flows, inputting the two characteristic modes into a time dependence module TCN for time dependence capturing, then fusing the characteristic modes with external characteristics, and finally predicting accident risk values of all areas of the city.
Time slices refer to the minimum length of time divisions of various types of data, which may be 1 day, 12 hours, 1 hour, etc. As the time division unit becomes smaller, the accident distribution becomes thinner and thinner, and the prediction difficulty becomes larger rapidly. Considering the timeliness of prediction and the difficulty of prediction, a time slice may be defined as 1 hour in this embodiment. The time window refers to how many consecutive time slices of data are used to predict the next time slice of data in the time series prediction task. The setting of the time window size affects the number of divisions of the valid samples and the complexity of the input data. While taking into account the periodicity of the data. For the above reasons, a time window of 12 may be defined.
The conversion of the accident risk replacement value ensures that most areas have no accident in a certain time period, but the label values of the accident risk replacement value contain distinguishable accident potential risk probabilities after mapping, so that the neural network can better distinguish the risk differences of different areas, thereby improving the prediction effect of the model.
In the event dynamic graph neural network EDGCN, the input is a node time-varying sequence and a graph adjacency matrix constructed based on the flow similarity. The more similar the traffic distribution is, the higher the priority of the feature delivery. Specifically, if an accident occurs in a certain sub-area at the time t, the graphic neural network at this time will preferentially transmit the accident risk information to those areas with high similarity to the traffic distribution of the area, because the probability of accident in the areas with similar traffic distribution conditions has a certain similarity at the same time (meaning that the global time-varying features are the same).
In the abnormal flow dynamic graph neural network ADGCN, a graph adjacency matrix constructed based on the flow abnormality degree is input as a node time-varying sequence. The flow distribution situation therefore occurs in the region of the large fluctuations in the last few time periods, with a higher priority for the characteristic transfer. Specifically, if the traffic flow condition of a certain subarea is severely changed at the time t, the graphic neural network at the moment can preferentially transmit the accident risk to the subareas which are similar to the traffic flow of the area or are greatly influenced by the traffic flow change of the subarea, and under the same time, the areas with the severely-fluctuated traffic flow are more likely to have accidents.
Further, as shown in fig. 2, the node area V is: v= { V 1,v2...,vN},vi represents the i-th node area, i=1, 2,3, … …, N is the total number of node areas;
the node area is a grid area obtained by rasterizing the target city area.
For urban region rasterization, the urban subregion may be divided into n=i×j grid regions, where I is the number of rows of the region division and J is the number of columns of the region division. Each node area may be set to a square size of 1.5km in side length.
Further, the accident risk realism value of the ith node area at time slice tThe calculation method of (1) is as follows:
Where R (v i) is all incidents occurring in node region v i at time segment t, inS j is the j-th incident in all incidents R (v i), p j is the number of injuries to the j-th incident in all incidents R (v i), q j is the number of deaths to the j-th incident in all incidents R (v i), R j is the number of vehicles involved in the j-th incident in all incidents R (v i), and a 3 is a third weight.
In this embodiment, accident data, traffic data, weather data, and the like may be divided into respective node areas according to time and place in advance and integrated into one trainable data set. The data in the dataset may correspond to a time dimension, a feature dimension, and a number of nodes, respectively.
As an example, a 1＝a2＝1,a3 =2.
Actual value of accident riskThe comprehensive risk value of an accident can be better represented, and the flexibility and the interpretability are higher.
Further, in time segment t, the node time-varying feature matrix of node region v i The method comprises the following steps:
wherein the method comprises the steps of For traffic of node region v i at time segment t,For the vehicle speed of the node area v i at time segment t,For the casualties of node area v i at time segment t,The vehicle is involved in the time segment t for the node region v i.
Further, at time segment t, the global time-varying feature matrix G t of node region v i is:
Gt＝{Wt,Qt}，
w t is the weather signal of time segment t, and Q t is the timing signal of time segment t.
The meteorological signals comprise weather conditions, wind speed, temperature and the like; the timing signals include timing related information such as month, week, season, etc.
In a real city, the traffic situation of each sub-area is in the process of changing continuously. Areas with similar road network conditions, such as two sub-areas that also contain several busy intersections, are likely to exhibit different levels of traffic at the same time. This is due to various reasons, such as different geographic locations where they are located, different laws of population flow in itself and in surrounding areas. However, the two sub-areas have similarity in the overall traffic fluctuation cycle, and the traffic flows of the two areas can still be considered to have certain similarity. Therefore, it is not possible to simply compare the traffic flow of only a certain time segment, but the traffic flow distribution of several time segments in the past should be compared dynamically. To capture this similarity, the concept of flow similarity between regions is defined.
Further, the method for calculating the flow distribution similarity comprises the following steps:
In the middle of For the flow distribution similarity of the node region v i and the node region v j at the time segment t, i, j=1, 2..n; for traffic of node region v i over consecutive m time segments, Traffic for node region v j for consecutive m time segments.
The corresponding traffic similarity adjacency matrix S t is:
wherein the method comprises the steps of Representing cosine similarity, using variablesAndAlternatively, the expression is:
Since the value of the vehicle flow is a non-negative value, thus (2) The value range of (C) is [0,1]. The larger the value, the more similar the traffic distribution of region v i,vj over the past m time segments.
The flow distribution conditions for the two sub-regions are also subject to severe variation and need to be screened. In order to screen out the region with severely reduced flow similarity, firstly defining the flow attenuation degree of the two sub-regions; flow attenuation matrixIs defined as:
The flow rate decay is to determine whether the flow rate similarity of two consecutive time slices of the two sub-regions is a decay. If the number is greater than 0, the similarity is reduced, and the number needs to be selected. If the similarity is smaller than 0, the similarity is improved, and the similarity is directly set to 0 without paying attention.
Based on the flow attenuation, a flow anomaly may be defined.
Further, the method for calculating the flow anomaly degree comprises the following steps:
In the middle of For the traffic anomalies of node region v i and node region v j at time segment t,For the traffic attenuations of node region v i and node region v j at time segment t,The flow distribution similarity of the node area v i and the node area v j in the time segment t-1 is obtained, and θ is a flow abnormality threshold;
the corresponding flow anomaly adjacency matrix Γ t is:
Further, the method for calculating the loss function includes:
Loss is a Loss function, ω 1 is a first constant, ω 2 is a second constant, ω 2>ω1 >1, and top27 is a set of high risk regions.
The present embodiment optimizes the loss function to solve the imbalance of the regional distribution under the premise of accident occurrence.
If an accident occurs in a certain accident-prone area and the model predicts that no accident occurs, a penalty factor is additionally applied to the accident-prone area. Through punishment items, high risk areas are focused on, and therefore priori knowledge training is achieved, model effect optimization is promoted, and application value is improved.
The data sources in the invention comprise vehicle collision data, which can be acquired from New York police department; but also the type of accident vehicle and the main reasons for the collision.
The weather data can be obtained from a national weather data center, and the data dimension includes date and time, site longitude and latitude, temperature, humidity, visibility, wind direction, weather and the like.
In the present invention, different solutions are proposed for the double unbalance of data, the unbalance of accident occurrence and accident non-occurrence and the unbalance of accident distribution area. The unbalance of the accident distribution area adopts a threshold adjustment mode, namely a mode of adding weight to the high-incidence area to adjust the preference degree of the model. The definition of the accident-prone area is: the time period of the accident exceeds 15% of the whole history. There are 27 total. For these areas. In the training phase. By applying an additional penalty factor to the loss function. To make the model more focused on these areas.
Specific examples:
description of data:
total samples: 8760×660×12, hour×node×node characteristics;
each time the model is input: 8 x 660 x 12;
Model output: 660 x 1;
data set partitioning: 7:2:1, training: and (3) verification: testing;
Introduction of the index: MSE, MAE, ACC@27;
ACC@27 represents selecting 27 nodes (v i epsilon top 27) with high risk, and in a test stage, regarding the 27 nodes as two classification tasks, namely, accidents and accidents are not occurred to predict, wherein the prediction accuracy of the model on the high-frequency nodes directly influences the application value of the model.
Introduction of comparative model: because the data set is different from the original baseline model, the parameters of the original data set are referred to, and the embodiment carries out fine adjustment on the basis of the parameters, so that the parameters of the baseline model reach the optimal effect on the experimental data set, and the parameter description and the parameters of the baseline model are as follows:
(1) ARIMA: a common model for predicting time series generally has three coefficients: (q, d, q), the value of this experiment was (1, 2, 4);
(2) XGBoost: based on GBDT (gradient boosting decision tree), the excellent machine learning algorithm is realized, and in the experiment, L2 regularization coefficient is 0.002.
(3) LSTM: one common variant of recurrent neural networks. In this experiment, the number of layers was 2, the number of neurons was 256, and dropout=0.4.
(4) Hetero-ConvLSTM: is an accident prediction model combining spatial heterogeneity and time periodicity. The LSTM portion parameter setting is consistent with (3), CNN portion uses convolution kernels of 3×3 and 5×5 sizes, L2 regularization coefficient is 0.001, dropout=0.4.
(5) ST-GCN: the model provides space-time diagram convolution and can be well applied to prediction of space data such as traffic flow.
(6) AST-GCN: the model adopts a plurality of time scales to extract features on the basis of ST-GCN, and then performs attention fusion prediction. The time scale in this experiment was taken for 12 hours, 24 hours, 168 (7×24) hours, respectively.
The experimental results are shown in table 1:
TABLE 1
Conclusion of experiment: the effect of the traditional machine learning model is strong without deep learning model fitting capability, and compared with a deep learning method, ARIMA, XGBoost and the like, the effect is obviously poor; the overall effect of the GCN-based model is generally better; although the MSE of AST-GCN is very close to the model of the application, the model of the application has obvious advantages for predicting accident high-frequency nodes (often high-frequency nodes of flow information as well) and meets the expectations.
The invention provides a brand-new flow-event collaborative prediction framework, and aims to solve the zero expansion problem caused by less accident amount, firstly, the label conversion strategy and the loss function weight improvement are carried out based on priori knowledge, so that the distribution knowledge of the whole data set is better learned, meanwhile, through the decoupling information framework, the model learns two factors of the accident triggering event and the abnormal flow triggering event, and the triggering effect of the two factors on the urban abnormal event is fused, so that the model has stability, high efficiency and interpretability.
Finally, through verification, the model finally obtained by the method achieves the optimal effect on the real data set, and further illustrates the advantages of the method, and the method is an important step for exploring urban abnormal events.
Although the invention herein has been described with reference to particular embodiments, it is to be understood that these embodiments are merely illustrative of the principles and applications of the present invention. It is therefore to be understood that numerous modifications may be made to the illustrative embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the different dependent claims and the features described herein may be combined in ways other than as described in the original claims. It is also to be understood that features described in connection with separate embodiments may be used in other described embodiments.
Claims (5)
1. A city traffic accident risk prediction method based on flow and accident cooperative sensing is characterized by comprising the following steps of,
When the time segment t is obtained by the node time-varying feature input module (100), node time-varying feature matrixes corresponding to the previous m time segmentsV is a node area;
Node time-varying feature matrix by event dynamic graph neural network (200) Extracting traffic accident perception characteristics to obtain accident risk characteristics based on flow distribution similarity;
node time-varying feature matrix by abnormal traffic dynamic graph neural network (300) Extracting traffic flow sensing characteristics to obtain flow risk characteristics based on flow anomaly;
Capturing time-dependent information of accident risk features and flow risk features by a time-dependent module (400) to obtain space-time mixing features;
the collaborative awareness module (500) performs feature fusion on the time-space mixed feature and the global time-varying feature matrix { G t,Gt-1...Gt-m+1 }, and obtains an accident risk prediction value of the time segment t
Accident risk prediction value from time segment t by output module (600)And accident risk replacement valueCalculating to obtain a loss function for model training; the trained model is used for predicting the accident risk of the next time slice;
The accident risk substitution value The calculation method of (1) is as follows:
In the middle of The value is replaced for the risk of accident for the ith node area,Alpha 1 is a first weight, T is a time window, delta is a minimum value larger than 0, and alpha 2 is a second weight;
The node area V is: v= { V 1,v2...,vN},vi represents the i-th node area, i=1, 2,3, … …, N is the total number of node areas;
the node area is a grid area obtained by rasterizing the target city area;
accident risk true value of ith node area in time slice t The calculation method of (1) is as follows:
Where R (v i) is all incidents occurring in node region v i at time segment t, inS j is the j-th incident in all incidents R (v i), p j is the number of injuries to the j-th incident in all incidents R (v i), q j is the number of deaths to the j-th incident in all incidents R (v i), R j is the number of vehicles involved in the j-th incident in all incidents R (v i), and a 3 is a third weight;
When the time segment t is, the node time-varying feature matrix of the node area v i The method comprises the following steps:
wherein the method comprises the steps of For traffic of node region v i at time segment t,For the vehicle speed of the node area v i at time segment t,For the casualties of node area v i at time segment t,A vehicle for node region v i at time segment t;
At time segment t, the global time-varying feature matrix G t of node region v i is:
Gt＝{Wt,Qt}，
w t is the weather signal of time segment t, and Q t is the timing signal of time segment t.
2. The urban traffic accident risk prediction method based on the cooperative sensing of flow and accident according to claim 1, wherein,
The calculation method of the flow distribution similarity comprises the following steps:
In the middle of For the flow distribution similarity of the node region v i and the node region v j at the time segment t, i, j=1, 2..n; for traffic of node region v i over consecutive m time segments, Traffic for node region v j for consecutive m time segments;
the corresponding traffic similarity adjacency matrix S t is:
3. The urban traffic accident risk prediction method based on the cooperative sensing of flow and accident according to claim 2, wherein,
The calculating method of the flow anomaly degree comprises the following steps:
In the middle of For the traffic anomalies of node region v i and node region v j at time segment t,For the traffic attenuations of node region v i and node region v j at time segment t,The flow distribution similarity of the node area v i and the node area v j in the time segment t-1 is obtained, and θ is a flow abnormality threshold;
the corresponding flow anomaly adjacency matrix Γ t is:
4. The urban traffic accident risk prediction method based on cooperative sensing of flow and accident according to claim 3, wherein the flow attenuation degree The calculation method of (1) comprises the following steps:
5. The urban traffic accident risk prediction method based on the cooperative sensing of flow and accident according to claim 4, wherein the calculation method of the loss function comprises:
Loss is a Loss function, ω 1 is a first constant, ω 2 is a second constant, ω 2＞ω1 > 1, and top27 is a set of high risk regions.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111470417.0A CN114139984B (en) | 2021-12-03 | 2021-12-03 | Urban traffic accident risk prediction method based on flow and accident cooperative sensing |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111470417.0A CN114139984B (en) | 2021-12-03 | 2021-12-03 | Urban traffic accident risk prediction method based on flow and accident cooperative sensing |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114139984A CN114139984A (en) | 2022-03-04 |
| CN114139984B true CN114139984B (en) | 2024-08-20 |
Family
ID=80387795
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111470417.0A Active CN114139984B (en) | 2021-12-03 | 2021-12-03 | Urban traffic accident risk prediction method based on flow and accident cooperative sensing |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114139984B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115546733B (en) * | 2022-11-23 | 2023-04-18 | 北京数业专攻科技有限公司 | Group behavior characteristic prediction method and device based on mobile signaling |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102360525A (en) * | 2011-09-28 | 2012-02-22 | 东南大学 | Discriminant analysis-based high road real-time traffic accident risk forecasting method |
| CN109117987A (en) * | 2018-07-18 | 2019-01-01 | 厦门大学 | Personalized street accidents risks based on deep learning predict recommended method |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11715369B2 (en) * | 2016-08-15 | 2023-08-01 | University Of Southern California | Latent space model for road networks to predict time-varying traffic |
| CN109993970B (en) * | 2019-03-15 | 2020-09-29 | 西南交通大学 | Urban area traffic accident risk prediction method |
-
2021
- 2021-12-03 CN CN202111470417.0A patent/CN114139984B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102360525A (en) * | 2011-09-28 | 2012-02-22 | 东南大学 | Discriminant analysis-based high road real-time traffic accident risk forecasting method |
| CN109117987A (en) * | 2018-07-18 | 2019-01-01 | 厦门大学 | Personalized street accidents risks based on deep learning predict recommended method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114139984A (en) | 2022-03-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109923575B (en) | Device and method for measuring absolute and/or relative risk potential of vehicle accident | |
| CN113487066B (en) | Long-time-sequence freight volume prediction method based on multi-attribute enhanced graph convolution-Informer model | |
| CN109448361B (en) | Resident traffic travel flow prediction system and prediction method thereof | |
| CN109840660A (en) | A kind of vehicular characteristics data processing method and vehicle risk prediction model training method | |
| Mao et al. | Analysis of road traffic speed in Kunming plateau mountains: a fusion PSO-LSTM algorithm | |
| CN110956807A (en) | Highway flow prediction method based on combination of multi-source data and sliding window | |
| CN113988476A (en) | Dynamic assessment prediction method for road transportation safety risk | |
| CN117332909B (en) | Multi-scale urban waterlogging road traffic exposure prediction method based on intelligent agent | |
| Xu et al. | Urban short-term traffic speed prediction with complicated information fusion on accidents | |
| CN113868492A (en) | Visual OD (origin-destination) analysis method based on electric police and checkpoint data and application | |
| Lin et al. | Building autocorrelation-aware representations for fine-scale spatiotemporal prediction | |
| CN114139984B (en) | Urban traffic accident risk prediction method based on flow and accident cooperative sensing | |
| Iranmanesh et al. | Identifying high crash risk segments in rural roads using ensemble decision tree-based models | |
| Poddar et al. | Massively parallelizable approach for evaluating signalized arterial performance using probe-based data | |
| Dorosan et al. | Use of machine learning in understanding transport dynamics of land use and public transportation in a developing city | |
| CN112905856B (en) | Method for constructing high-speed traffic data set with space-time dependence | |
| CN114880852A (en) | Modeling analysis method and system based on social perception data | |
| Hoong et al. | Road traffic prediction using Bayesian networks | |
| Zhao et al. | Exploring the impact of trip patterns on spatially aggregated crashes using floating vehicle trajectory data and graph Convolutional Networks | |
| Zhou et al. | Stack ResNet for short-term accident risk prediction leveraging cross-domain data | |
| Babaei et al. | Spatial Data-Driven Traffic Flow Prediction Using Geographical Information System | |
| CN117037461A (en) | Road network traffic jam prediction method based on multi-weight graph three-dimensional convolution | |
| CN115796590A (en) | Method and device for predicting road fog accident risk, electronic equipment and medium | |
| CN115456238A (en) | Urban trip demand prediction method based on dynamic multi-view coupling graph convolution | |
| CN115424139A (en) | Residential area extraction method fusing remote sensing data and position big data |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |