WO2018122804A1

WO2018122804A1 - Road traffic anomaly detection method using non-isometric time/space division

Info

Publication number: WO2018122804A1
Application number: PCT/IB2017/058534
Authority: WO
Inventors: 杜豫川; 邓富文; 严军
Original assignee: 同济大学; 许军; 上海同济检测技术有限公司
Priority date: 2016-12-30
Filing date: 2017-12-30
Publication date: 2018-07-05
Also published as: GB202100341D0; GB201909409D0; WO2018122802A1; GB2582531B; GB202009833D0; GB2582531A; GB2569924A; CN109923595A; GB2588556B; CN109716414A; GB201711408D0; GB2587588B; GB201909407D0; WO2018122806A1; WO2018122585A1; GB201909406D0; CN109716414B; CN109844832B; GB201905907D0; GB2569924A8

Abstract

A road traffic anomaly detection method using non-isometric time/space division, comprising: establishing time/space sub-segments; preprocessing past tracking data; preprocessing real-time tracking data; analyzing past tracking data and training an RNN model; analyzing real-time tracking data and extracting features; detecting anomalies; indicating the degree of severity of the anomalies; and evaluating system performance. The method utilizes road network density or peak hour traffic as the basis for the division of time/space ranges, and uses a large amount of past tracking data to train an RNN model. The difference between RNN model prediction values and actual values reflects differences in traffic states. The present method achieves real-time smart detection of road traffic anomalies and incidents on city roads, increases detection reliability, and is low in cost.

Description

Road traffic anomaly detection method based on non-equidistant space-time division

Technical field

The invention belongs to the technical field of traffic detection. In particular, the present invention relates to a method for real-time detection of urban road traffic anomalies. Through the on-board GNSS positioning device of the floating car, the spatial position information of different time can be obtained. After data preprocessing, map matching and data fusion, the RN model is trained based on the historical data of the specific speed and time of the journey; the predicted value and real time according to the RNN model The difference between the actual values of the journey speed can effectively identify urban road traffic anomalies. Background technique

Traffic anomaly detection is an important part of urban traffic management and one of the core functions of intelligent transportation systems. Traffic anomalies mainly include traffic accidents, vehicle dumping, falling objects, damage or malfunction of road traffic facilities, and other special events that cause traffic flow disturbances. Such incidents are prone to traffic congestion, reduced road capacity, and severely affect the normal operation of the entire road traffic system. Through traffic anomaly detection, traffic managers can timely understand traffic anomaly information and take appropriate inducement and control measures to reduce the adverse effects of traffic anomalies.

Traffic anomaly detection can be divided into manual mode and automatic mode. Manual methods include patrol cars, emergency telephone reporting, and video surveillance. Due to the human and material resources and poor real-time performance, traffic management needs cannot be met. The automatic method relies on the automatic event detection (AID, Automated Incidence Detection) algorithm. The basic principle is to identify traffic anomalies by detecting changes in road traffic at different locations. Currently used AID algorithms include pattern recognition algorithms (such as Califorma algorithm, Monica algorithm), statistical prediction algorithms (such as exponential smoothing, Kalman filtering), traffic flow model algorithms (such as McMaster algorithm), and intelligent recognition algorithms. (such as artificial neural networks, fuzzy logic algorithms).

However, the current detection methods have disadvantages such as high requirements on facilities, high computational complexity, and inability to make further judgments on the situation of abnormal conditions. The invention utilizes the trajectory data returned by the taxi and the bus GNSS positioning device to establish a historical traffic state database and a real-time traffic state database, and analyzes the traffic anomaly events by analyzing the difference of the traffic flow characteristics reflected by the two. The method has the characteristics of good real-time performance, parallel processing, high recognition rate and low requirements for detection facilities, and is suitable for detecting urban road traffic anomalies in a data environment with real-time floating vehicle positioning data.

At present, for the monitoring of traffic anomalies, the following representative technologies are available:

A US patent application, US 20160148512, discloses a composition principle and implementation method of a traffic anomaly detection and reporting system. The system consists of a sensor, a communication module, a mobile processing module, and a user interaction module. The sensor is used to collect relevant data around the vehicle; the communication module is used for transmitting the vehicle data and receiving data of the surrounding vehicle; the mobile processing module is for processing and analyzing the data of the relevant vehicle in a certain area and generating a traffic event report; user interaction The module is able to provide traffic incident reports like a user. The scheme is a traffic anomaly detection technology based on the vehicle and vehicle communication network, which can use various types of information collected by sensors to identify abnormal events. However, since the sensor and the communication unit need to be separately installed and debugged, the implementation is difficult; the processing capacity of the mobile processing unit is limited; and the mobile and fixed message receiving end is required, and the system itself has a failure probability and the reliability is not good.

A Chinese patent application, CN 104809878 A, discloses a method for detecting abnormal state of urban road traffic using bus GPS data. The scheme obtains the link delay time index according to the GPS historical data, obtains the instantaneous speed, the cycle average speed, the weighted moving average speed and the multi-vehicle average speed according to the current GPS data, and uses the gauge variable analysis algorithm to detect the abnormality. This program does not need new Increase inspection facilities and facilitate implementation. However, the characterization of the traffic situation is too simplistic, and it is impossible to analyze the characteristics and causes of traffic anomalies. There is no basis for the division of traffic scenarios, and the influence of weather and other factors on traffic situation changes cannot be considered. Summary of the invention

In order to explain the contents of the present invention more clearly, the technical terms involved will first be explained as follows:

Floating car: Also known as the probe car. Refers to buses and taxis that have on-board positioning devices and are driving on city roads.

GNSS: Global Navigation Satellite System. Including GPS, GLONASS, GALILEO and Beidou satellite navigation systems.

Space-time sub-zone: A zone divided by two dimensions, time and space, reflected in a certain space within a certain period of time. Divide a day into several time segments, such as 0:00-0: 10, 0: 10-0:20..., called the time sub-region; divide the implementation area of urban road traffic anomaly detection into several spatial segments, for example Longitude 121.58° E-121.59 ⁰ E, latitude 31.16° N-31.17° N, that is, the spatial sub-region; the space-time segment formed by the intersection of any one time sub-region and any one spatial sub-region, called the spatiotemporal sub-region For example, the longitude of 121.58° E-121.59 ⁰ E, latitude 31.16° N-31.17 ⁰ N is a time-space segment of 0:00-0:10.

Historical trajectory data: Historical trajectory data is trajectory data accumulated over a long period of time and stored in a database. Historical trajectory data is dynamically changing data that needs to be updated in a timely manner and periodically reprocessed and analyzed to ensure the accuracy of historical traffic feature extraction. The data for each time-space sub-area can be processed in parallel to increase efficiency. In the present invention, it may be simply referred to as historical data.

Real-time trajectory data: The real-time trajectory data is a trajectory data set within a time zone that is closest to the current time. In the present invention, it may be simply referred to as real-time data.

Traffic situation: A general term for the comprehensive situation of traffic operations within a certain period of time and within a certain space.

Traffic anomalies: traffic flow disturbances caused by traffic accidents, vehicle dumping, truck falling, road traffic facilities damage or malfunctions.

Abnormal traffic severity: The severity of traffic flow disorder is the difference in traffic flow characteristics after traffic flow and traffic anomalies in normal conditions.

Traffic Anomaly Index: A measure of the severity of traffic. The range is 0~10. The larger the value, the more serious the traffic anomaly.

Traffic environment: The sum of all external influences and forces acting on road traffic participants. This includes road conditions, transportation facilities, landforms, meteorological conditions, and traffic activities of other transportation participants.

Map Matching: The process of associating geographic coordinates with a city road network.

Peak hourly traffic: The maximum hourly traffic flow in a city's road section.

Response variable: A variable that changes according to an independent variable, also called a dependent variable.

RNN: Recurrent Neural Network is an artificial neural network with nodes connected in a loop. The internal state of this network can show dynamic timing behavior. It can use its internal memory to process input sequences of arbitrary timing. .

Elman-RNN: An RNN network structure, see A RNN that learns to count) □

Training process: The process of optimizing neural network parameters through iterative calculations to reduce the model error of the neural network on the training data set. The object of the present invention is to establish a scheme based on a floating vehicle trajectory recording system, using historical GNSS positioning data and real-time GNSS positioning data, combined with traffic environment information to identify road traffic anomalies. In order to achieve the above object, the present invention provides the following technical solutions:

The premise of the implementation of the present invention is: a floating car (taxis, bus, etc.) equipped with a GNSS track recorder; with large-scale storage, A data center that computes, real-time task processing capabilities.

The scope of application of the present invention is: Urban roads (including ground roads and elevated roads) through which the above-mentioned floating vehicles pass.

The implementation steps of the present invention include:

1) Determine the spatio-temporal range of the test and establish the spatio-temporal sub-area.

Based on actual application requirements, determine the time range and spatial extent for which traffic anomaly events need to be detected. The time range can be set to all day, that is, 0:00-24:00; it can also be set to a specific time period. For example, to detect the traffic abnormal time during the period from 17:00 to 20:00, the time will be detected. The range is set from 17:00 to 20:00. Here is just a special example. There are many other situations, which are not explained here. The spatial scope can be set as a certain city area according to the administrative division, such as Beijing, Shanghai, Huangpu District, etc. It can also be set as a certain urban functional area according to the urban spatial structure, such as a central business district and industrial area of a certain city.

The establishment of the spatiotemporal sub-area refers to dividing the detected time range into a number of smaller time segments, and dividing the detected spatial range, that is, the implementation area of the urban road traffic anomaly detection, into a plurality of smaller spatial segments. For the establishment of spatiotemporal sub-areas, a variety of empirical division methods can be used, including equidistant space-time division method and non-equidistant space-time division method.

2) Data preprocessing.

The GNSS positioning data is used for data cleaning, data integration, data conversion, and data reduction to improve the structure of the data. GNSS, the Global Navigation Satellite System Positioning System, is a space-based radio navigation and positioning system that provides all-weather three-dimensional coordinates and speed and time information anywhere on the Earth's surface or near-Earth space. It mainly includes GPS (Global Positioning System) in the United States, GLONASS (Global Navigation Satellite System) in Russia, GALILEO in the European Union and China's Beidou satellite navigation system. It also includes QZSS in Japan and IRNSS in India. Such as regional navigation and positioning systems, as well as satellite positioning enhancement systems such as WASS in the United States and MSAS in Japan. In order to establish a unified data distribution standard among different navigation and positioning system devices, the National Ocean Electronics Association of the United States has developed a unified NEMA (National Marine Electronics Association) communication protocol to regulate GNSS data broadcasting. Therefore, each member system in GNSS, such as GPS, GLONASS, etc., although established and maintained by different countries and organizations, has a consistent data distribution format, so there is no need to transform the data format.

Within the selected space, there are many vehicles equipped with GNSS positioning equipment, such as taxis, buses, freight cars, private cars, etc. Based on the current application status of urban traffic data, in actual applications, urban taxis are often used as floating vehicles as data sources for traffic anomaly detection systems.

The collected GNSS positioning information contains some unreasonable information. In order to ensure the accuracy of the traffic abnormal state detection and discrimination results, it is first necessary to identify the abnormal data to ensure the reliability of the data. These anomalies include: data that falls outside the time and space of detection, and spatial position jumps that are clearly out of reasonable range. The so-called "space position jump beyond the reasonable range" is illustrated by the following example. If the positioning point uploaded by a floating vehicle positioning device is recorded as A at 10:30:00 on a certain day, the positioning point uploaded by the floating vehicle positioning device at time 10:30:30 is recorded as B, the distance between position A and position B. It is 1500 meters, then the speed of the floating car is calculated to be at least 180km/h, which is beyond the common sense, so it is an abnormal spatial position jump, which should be eliminated in data processing.

3) Quick map matching.

After pre-processed GNSS positioning data, it is necessary to combine the urban road network data, map the GNSS positioning points to the city map through the map matching algorithm, establish the matching relationship between the positioning points and the road segments, and correct the error caused by the positioning drift.

At present, the electronic maps of various geographical regions are relatively detailed. Such electronic maps can be derived from the city's geographic information system, and of course can also be derived from other methods and routes. These electronic maps detail the urban road information, and several sections can be obtained by dividing. By matching the anchor points to the road segments by means of distance, angle, etc., the positioning information is matched to the actual geographical environment. 4) The representation of the floating vehicle path and the matching of different vehicle paths.

The path of the vehicle may not be unique given a set of starting and ending points. The complex urban traffic network consists of several sections, which are numbered, for example, Ll, L2, etc. Roads may have two different directions of travel. In this case, two different directions of travel should be represented as two different sections, given different sections.

For a given starting point and ending point, the intersection of the road segments in the urban road network can usually be used. Knowing the path of a floating car, it is now necessary to select the same path as the floating car path from the path information that has been sent by other floating cars, so as to obtain the same path group between the starting point and the ending point.

5) Data sampling.

The positioning data of the floating car includes information such as position coordinates, instantaneous vehicle speed, and recording time. In the urban road traffic anomaly detection method based on floating car data proposed in this patent, data sampling refers to screening part of the data from all floating car data for subsequent analysis and processing, and the screening is based on the computing power of the data center and Pre-proposed accuracy requirements were made. Different data sampling methods can be used based on different calculation capabilities and accuracy requirements. For example, when the computing power of the data center is strong and the accuracy of the detection is high, all the floating vehicle positioning data can be treated as a processing object, and comprehensive processing analysis is performed; and when the computing power of the data center is limited, it is assumed The current data center can process 500 data for each spatial sub-area within 1 minute, but the actual situation is that in 2000, each floating sub-area can generate 2000 floating-vehicle positioning data, so it can be from 2000 data. Randomly extract 500 data for analysis, so as to obtain processing results with limited accuracy within the computing power of the data center.

Depending on how the floating car data is used, it is possible to sample different properties of the floating car data, such as the travel speed and travel time. The urban road traffic anomaly detection method based on floating car data proposed in this patent uses the travel speed as the basis for urban road traffic anomaly detection. Therefore, data sampling refers to sampling the speed of the journey.

6) Historical trajectory data analysis and R N model training.

The so-called historical trajectory data refers to the floating vehicle trajectory data accumulated in long-term urban road traffic operations. Using historical floating vehicle trajectory data, an urban road traffic feature model can be established to reflect the general characteristics of urban traffic operations. The urban road traffic feature model mentioned here can refer to certain specific indicators, such as average speed, weighted average speed, etc.; it can also refer to various statistical models, such as the probability distribution of travel speed. In many previous models, a single indicator was used to indicate the traffic characteristics of a certain section or area (such as the historical average speed). Although this method is simple, the accuracy is not high and the sensitivity is poor, which often cannot be used in the detection of traffic anomalies. good effect. Therefore, this patent proposes to train the RNN model with the variation of traffic characteristic variables for each spatiotemporal sub-region. The R N model can give the predicted value of the traffic characteristic variable at the next moment when the real value of the traffic characteristic variable is given for some time period.

The traffic characteristic variables that can be collected include the travel speed and travel time. This patent uses the travel speed to reflect the traffic characteristics. Therefore, the traffic characteristic variable refers to the travel speed.

7) Real-time trajectory data analysis and feature extraction.

The so-called real-time trajectory data refers to the trajectory data of the floating car in traffic operation in a period of time not far from the current time. Using real-time floating car trajectory data, you can grasp the dynamics of traffic characteristics and reflect the current characteristics of current traffic operations. This patent uses the travel speed of the current space-time sub-zone to describe the current traffic characteristics.

8) Anomaly detection.

The idea of system state anomaly detection was first proposed by Dennngng, that is, by monitoring the abnormality of the system used in the system audit record, it is possible to detect an event that violates security and may cause a system abnormality. Dennrng's model is independent of any particular system, application environment, system vulnerability, and fault type, and is therefore a general anomaly detection model. The model consists of five parts: subject, object, audit record, outline, exception record and activity rule. A contour is a body that is represented by a metric and a statistical model relative to the object. Normal behavior. Dennrng's model defines three metrics, namely event counter, interval timer, resource measurer, and proposes five statistical models, namely, operational model, mean and standard deviation models, multivariate models, Markov process models, and Time series model. The model proposed by Denning establishes the statistically-based normal behavioral feature profile of the system subject through the analysis of the system audit data. When the detection, the audit data in the system is compared with the normal behavioral feature profile of the established subject. Exceeding a certain threshold is considered an abnormal event. This model lays the foundation for anomaly detection, and many anomaly detection methods and systems developed in the future are developed on the basis of it.

In recent years, in the development of anomaly detection technology, more artificial intelligence methods have been introduced to improve the performance of anomaly detection. These methods of artificial intelligence mainly include data mining, artificial neural networks, and fuzzy evidence theory. Data mining methods are used to determine what features are most important in a large data set. This technique is used in anomaly detection mainly to seek a more concise definition of the normal mode, rather than simply enumerating all normal modes as in the conventional anomaly detection method. The introduction of data mining methods enables the detection system to generally include normal patterns not included in the training data simply by identifying the main features in the normal mode. The artificial neural network anomaly detection problem can be regarded as a general data classification problem. In the statistical anomaly detection mentioned above, user behavior data is divided into two categories according to certain statistical criteria: abnormal behavior and normal behavior. Because the statistical-based method has certain difficulties in extracting and abstracting the audit instance, it may cause large errors, and must rely on some probability distribution hypotheses. Generally, it is necessary to describe the measure of user behavior by experience and feeling, so the artificial neural network is introduced. Clustering method. The artificial neural network has the self-learning self-adaptive ability to train the neural network with sample points representing the normal user behavior. Through repeated learning, the neural network can extract the normal user or system activity patterns from the data and encode them into the network structure. In the detection, the audit data can be judged whether the system is normal by learning a good neural network. Because the anomaly evaluation criterion has certain ambiguity, the fuzzy evidence theory is introduced into the anomaly. For example, an intrusion detection framework model based on fuzzy expert system is established, which can better reduce the false alarm rate and false alarm rate.

This patent proposes an anomaly detection scheme based on cyclic neural network. In particular, it is an anomaly detection scheme using RNN (Circular Neural Network). The real-time trajectory data is input into the RN model based on historical data, and the predicted values are obtained, and then compared. The difference between the predicted value and the true value. The program uses a deep learning method that automatically updates the model over time and has strong adaptive capabilities.

9) Quantitative characterization of abnormal severity and release of abnormal information.

The severity of traffic anomalies should be released to the public in a clear and concise manner to avoid possible congested areas and improve the efficiency of urban traffic. The severity of the abnormal condition is characterized by the traffic anomaly index, ranging from 0-10, where 0 means no anomaly and 10 means a height anomaly.

The location of the anomaly is projected onto the electronic map and published publicly through the smart mobile device APP or the like.

10) System performance evaluation.

The evaluation of system performance refers to the evaluation of the accuracy of traffic abnormal state detection, and its evaluation indicators include false positive rate and false negative rate. The lower the false positive rate and the false negative rate, the better the performance of the system. In the step 1), the division of the spatiotemporal sub-area may specifically adopt the following method:

11) Isometric space-time division method. The segment size of the time dimension is determined. The time segment span is a fixed value, usually 30 mm is taken as a time segment; the segment size of the spatial dimension is determined, and the spatial segment span is a fixed value, and a spatial grid of 200 m×200 m is usually taken as a spatial segment.

12) Non-equidistant space-time division method. For urban central areas where the road network density is greater than 2km/km ² or the peak hour flow is greater than 1000 vehicles/hour, take 30min time segments and 200m X 200m space segments. For road network density less than 2km/km ² or peak hour traffic is less than 1000 sub-urban suburbs, taking 30min time segments and 400mX 400m space segments. The step 3) specifically includes the following steps:

31) Divide the space area to be processed into a grid of a certain size, and the range of each grid area can be expressed as ={(x _s , y _s )\x _s G[x _r , x _r+1 ), y _s &[y _r , y _r+1 )}, each grid area contains several road segments, and the set of these road segments is represented as R _s , and each road segment in the set of the road segments is represented as ij and is Each road segment is assigned a number;

32) Determine the grid area where the anchor point is located, and use the distance and azimuth angle to search for the section where the anchor point A is located in the set of road segments. ^ The matching scheme includes:

321) Single point matching scheme:

Searching for the section closest to point A, when the difference between the direction of travel of the point A and the direction angle of the section ij is less than the threshold, that is, | - | < is satisfied, the matching is completed, and the threshold may be 2.5°, y, ιο ° etc. If you do not satisfy | - |< , delete the link in the search space and continue searching for other segments until the condition is met. The matching method is shown in Figure 3.

322) Point sequence matching scheme:

This program is suitable for high frequency floating car data. The floating vehicle GNSS data acquisition frequency is expressed as f ₀ =\l , and the point POHO), Pfc+i, which is adjacent to A in time. ;) is defined as the 1-adjacent point of Α, P04-2i _Q ;), P 4+2i _Q ) is defined as the 2-adjacent point of A, and so on, then P(t _A -kk), Pfc+ is defined as A /t-adjacent point. When / _Q <lHz, take /t=l or 2. Take the distance between the distances A and A/t-the neighboring points and calculate the mean value of the driving direction angles of the neighboring points of A and A. If the |.- |< is satisfied, the matching is completed; otherwise, the other sections are searched. Until one is satisfied

33) Use the straight line equation of the road segment (if it is a curved road segment, it is roughly split into straight lines), calculate the projection coordinates of the GNSS positioning point on the road segment, and reduce the error caused by the GNSS positioning drift. The specific method uses the GNSS positioning point linear projection method as:

Determine the straight line equation of the road segment ^ (If the road segment is a curve, divide it into several straight line segments):

Yy, =k(xx,) where the slope is:

The projection line equation is: yy _A =- ^x - ky _A -ky _t +k ² x _t +x _A

Solve the projected coordinates p is:

k ² +l

k ² y _A + y _t +kx _A - x _j

y _P

After the map matching process, combining the timestamp data of the coordinates of the positioning point, matching the positioning point to the space-time sub-region ₍ the step 5) may specifically adopt one of the following methods:

51) Full sample plan for speed information. The total travel speed data of each sub-floating vehicle in a time and space sub-region constitutes the whole. Real The method is to calculate the travel speed of each vehicle in the space-time sub-region: ="^ ~ ² ' ³ + + + "'", where ί/ _{1 2} ... 4 - _ln is the first in the space-time sub-area ^ The distance from the second GNSS anchor point, ..., the distance between the -1 and the GNSS anchor points, -t _n is the first in the space-time sub-region, ... . .., the time stamp of the GNSS anchor points; the data in each spatio-temporal sub-area is not filtered to form a set ^ for subsequent processing.

52) Time-smooth sampling plan for speed information. Specify the length of the time segment and the upper limit of the number of segments of the same time; search for the velocity data in each time segment in a time-space sub-region. If the number of velocity data in the time segment exceeds the upper limit, the data of the upper limit is randomly added to the data to be processed. sample. The implementation method is to calculate the travel speed of each vehicle in the space-time sub-region: v _f W .. ""- ¹ '", where ₂ ... is between the first and second GNSS anchor points in the space-time sub-region Distance, ..., the first "-1 and _n "

The distance between the GNSS positioning points is the first time in the space-time sub-region, ..., the time stamp of the first GNSS positioning point; the specified time segment length t _p , the upper limit of the number of segments of the same time; ^ _{αχ ;} speed data within a search area at the time of each temporal sub-time segment, the time segments if the data rate exceeds the upper limit number of pieces; ^ «, randomly; ^ ^ was added and _∞ of data for subsequent processing. The step 6) specifically includes the following steps:

61) Determine the network structure of R N , where the basic structure of the Elman-RNN neural network is used. The network structure is shown in Figure 4 and Figure 5. The network specifically contains the following components:

611) Input layer: According to the characteristics of the neural network, the input layer is each instance of the historical data to be trained, since the input data here is a one-dimensional data stream, that is, the travel speed data in the space sub-area is formed in the time dimension. Time series data, therefore, the number of input layer neurons and the number of output layer neurons are set to 1.

612) Implicit layer: In the design of neural network, the number of hidden layers has not been determined. Generally, a large number of experiments are needed to determine the number of hidden neurons in the network model. The recommended value is 5~ 8.

613) Output layer: The purpose of establishing a neural network is to output a predicted value, that is, to predict the value of the next moment by the values of the first two moments. The output layer output value is the predicted value, so the number of output layers is set to 1.

614) Context layer (Context layer): In Elman-Network, the context layer is used to store the output of the hidden layer at the previous moment, so the number of context layer neurons is the same as the number of hidden layers.

615) Offset node: The input layer and output layer each contain an offset node, and the initial value is set to 0.

62) Set the basic parameters of the Elman-RNN neural network model:

The Elman network structure contains two parameters, namely the number of hidden layer neurons and the number of layers of the context layer. The number of neurons in the hidden layer The better results are selected through multiple experiments. The recommended setting is 5~8. Since a context layer saves the hidden layer output value at the previous time, it is recommended to set it to 1.

The parameter learning rate needs to be set when training the model using the back propagation algorithm. The size of the learning rate determines the degree of change of the weight in the iterative process of the neural network. The large learning rate can make the algorithm converge quickly, but it may fall into the local solution. The small learning rate makes the convergence of the algorithm slower, but it can guarantee convergence. To the global minimum, therefore, the value of the learning rate should generally be chosen to ensure the performance and stability of the model. The recommended setting is 0.01 to 0.8.

In the simulated annealing module of the algorithm, the initial temperature, the termination temperature, and the number of iterations of each temperature value need to be set. The value of the initial temperature has an important influence on the performance of the algorithm. The larger the initial temperature, the more iterations of the algorithm and the greater the probability of convergence to the minimum. But the time spent is also bigger. Similarly, when the initial temperature is low, the performance of the algorithm will be affected, but the time spent will be less. In practical applications, the initial temperature setting needs to be selected by repeated trial and error. The termination temperature is the lower limit temperature set by the termination algorithm during the temperature drop. The more iterations of each temperature value, the more the number of possible solutions, and the greater the likelihood of convergence to the global minimum. It is recommended to set the initial temperature to 10 ⁵ , the termination temperature to 10 - ² , and the number of iterations per temperature to 100.

63) Train the RN model. The input of the RN model is the time series data composed of the travel speed = { ν ₂ , ... , v _n } , and the output is the Elman-R N neural network model with the optimal parameters. The steps for model training are as follows:

631) Read data from a data file or database and combine the data into (input, output) pairs, ie ^^^, Ο^ι^,.,., ^^, for model training.

632) Create an Elman-RNN neural network, in which the number of neurons in the input layer is 1, the number of neurons in the hidden layer is 5-8, and the number of neurons in the output layer is 1. The context layer is hidden at a previous time. The output of the layer, the number of neurons is the same as the hidden layer. Since the output in the training data is positive, the Sigmoid activation function is used.

633) Initialize the R N model, set the input layer to the hidden layer, hide the layer to the output layer, and the weights of the connections between the neurons are random values, and set the offset units of the input layer and the hidden layer, respectively, and initialize to 0.

634) Set the main training algorithm to the back propagation algorithm, the standby algorithm to the simulated annealing algorithm, and use the hybrid strategy of the two algorithms to train the model. In the training process, if the error rate of the model does not decrease or decrease after a training session If the magnitude is less than the set minimum, the simulated annealing algorithm is used to train; if the greedy strategy is used, if the error rate of the model does not improve after a certain training, the update weight is discarded, and the weight of the model is reset to the previous training result.

635) Before starting training, save the network weight and error rate of the current model required by the greedy strategy. The error rate of the current model required to save the blending strategy. Start training. For the context layer and input layer data, multiply the corresponding weights, plus the offset, and calculate the output of the hidden layer by the activation function. The hidden layer data is multiplied by the corresponding weight, plus the offset, and the output of the model is obtained by activating the function. Calculate the difference between the actual output and the model output, and use the backpropagation algorithm to pass the error back to the hidden layer and the input layer to calculate the gradient. The weight increment is calculated based on the calculated gradient, the weight is updated, and the updated weight is saved. Copy the hidden layer output to the context layer.

64) After a single training, the model is processed using the following strategy:

641) Greedy Strategy: If the training error rate does not improve, the recovery weight and error rate are the last values.

642) Hybrid strategy: If the error rate of the model does not decrease after this training, or if the magnitude of the decline is less than the set minimum, then the simulated annealing algorithm is used for training. A training of the simulated annealing algorithm, the general steps are as follows: First calculate the error score of the current model, then add a random number "ί« to the ownership value and offset value of the current model, get new weights and offset values, where:

, , 0.5— Random

Add = * temp

startTemp

In the formula, Random is a random number, startTemp is the initial temperature, and temp is the current temperature. The updated model error score is calculated. If the new error score is smaller than the current error score, indicating that the new weight improves the performance of the model, save the new one. Weight, otherwise discard; multiply the current temperature by a fixed ratio ratio to lower the temperature:

Log stopTemp I startTemp)

Ratio = exp

Cycles - 1

Where stopTemp is the termination temperature and cycles is the number of iterations of a training session. Repeat the above process cycles.

65) Stop condition judgment: If the degree of improvement to the model is less than the minimum value is greater than the set threshold, the algorithm terminates the steps 63), 64), 65) until the algorithm terminates, and the trained RNN model is obtained. See Figure 6 for the training process. The step 7) may specifically adopt one of the following methods:

71) Time window means. Using the travel speed data obtained by data sampling, calculate the current space sub-zone travel speed at the current time

1

The mean of the sub-area / _rt ) = "^∑V,, as a representation of the real-time traffic situation

7~ί

72) Rolling time domain mean method. Travel speed data obtained using the sampling data, the spatial sub-region to calculate a current travel speed in the nearest mean / ΣΣ Μ time sub-regions, wherein like M- 3 ~ 5, as the real time traffic situation characterizing _t

The step 8) specifically includes the following steps:

81) Sequence _Vl , v ₂ ,..., v of the point in which the time series data to be processed by the abnormal point detection is sorted in ascending order. In order to fit the data, the data is in the form of a two-dimensional point on the plane. The coordinates are expressed as (1),...,(",1„).

82) In order to detect the abnormal point data, the difference between the data predicted by the RNN model and the real data needs to be calculated. The specific implementation process is as follows:

821) It is known that the model fitted by historical data^ calculates the predicted value of the fitted model

822) Calculate the difference between the predicted value of the model and the true value: _c v _rt , v _pr \ = v _pr - _{rt The} step 9) specifically includes the following steps:

91) Normalize the difference in velocity distribution of each space-time sub-region to a normalized value of 0~1 _.

Diff^ - m {diff)

m x^diff^ - min diff^

92) Calculate the traffic anomaly index for each time and space sub-region

10;

93) Project the location of the region with an anomaly index higher than 5 onto the electronic map, and expose it to the public in the form of a smart mobile device APP to enable the driver to avoid potential congestion points and improve the efficiency of urban road traffic. The step 10) specifically includes the following steps:

101) Calculate the false negative rate of traffic anomaly:

α _λ = -^ χ 100%

102) Calculate the false positive rate of traffic anomaly:

α : 100%

In the above two formulas, the total number of missed events per unit time is the total number of false positive events per unit time. The total number of abnormal times that actually occurred during the unit time. The present invention has the following advantages over similar technologies in the same field:

(1) Make full use of existing floating vehicle operation data (GNSS trajectory data), through historical traffic feature extraction and real-time traffic situation Analysis, detecting changes in traffic conditions, real-time, low-cost, intelligent detection of urban road traffic anomalies;

(2) Establishing the R N model by changing the traffic characteristics, it can reflect the traffic operation law more comprehensively and comprehensively, avoiding the one-sidedness and instability of the traffic characteristics using a single index, and the reliability of detection is higher;

(3) Ability to adaptively adjust model parameters to reflect subtle changes in traffic patterns on large time scales;

(4) According to the test of actual data, the urban road traffic anomaly detection technology based on floating car data proposed by the present invention can realize the detection of abnormal events with high accuracy, the detection rate exceeds 90%, and the false negative rate is less than 15%. The false alarm rate is lower than 20%, and it has achieved good detection results, and can be applied to urban traffic intelligent management and service. DRAWINGS

The details and advantages of the present invention will become apparent and readily understood in conjunction with the following drawings in which:

Figure 1 shows a schematic diagram of the components and basic principles of the present invention;

Figure 2 is a schematic view showing the overall flow of the present invention in the implementation process;

3 is a schematic diagram showing an implementation manner of a fast map matching algorithm of the present invention;

Figure 4 is a block diagram showing the structure of an RNN in the present invention;

Figure 5 is a view showing the detailed structure of the RNN in the present invention;

Figure 6 shows a schematic flow chart of performing R N training;

Fig. 7 is a flow chart showing the abnormality detection of the present invention by comparing the RNN model with real data. Specific implementation

In order to more clearly and clearly clarify the objects, technical solutions and advantages of the present invention, the specific embodiments of the present invention are described in detail below.

As shown in Fig. 1, the overall system architecture of the present invention includes: an onboard GNSS track recorder mounted on a floating vehicle, a data center, a GNSS satellite, and a communication system. The GNSS here includes GPS, GLONASS, GALILEO, Beidou, IRNSS, QZSS and any similar navigation satellite positioning system. GNSS track recorders equipped with floating cars, buses, etc., record the position information of the vehicle at various points in time at a certain sampling frequency / (general requirements of 0.1 Hz), and pass the GPRS mobile communication network (also can be used) Wireless network communication technologies such as WCDMA and TD-LTE, but the cost will be increased accordingly) The location information will be sent to the data center in real time. The data center establishes a historical road traffic characteristic database through data preprocessing, data fusion, and through a specific algorithm; establishes a real-time traffic feature database for the recently received real-time data; and determines whether the current traffic feature is abnormal through the mapping relationship between the historical database and the real-time database And visualize the display through the processing terminal and generate a traffic anomaly event report.

The overall process of the scheme is shown in Figure 2, including the acquisition and storage of GNSS trajectory data, the establishment of spatiotemporal sub-areas, historical traffic feature extraction, real-time traffic feature extraction, and anomaly identification. Collecting and storing GNSS trajectory data is the data foundation of the whole scheme. Due to the huge amount of data, a distributed storage scheme should be adopted. For distributed storage, there are mature technologies, which are not the content of the present invention. The basic assumption of the establishment of a spatio-temporal sub-area is that it has the same traffic characteristics in a particular area and a specific time period. This assumption is generally applicable after long-term observation. Historical traffic feature extraction, the principle is to use the GNSS trajectory data to calculate the travel speed, use a large number of historical travel speed data in the same space-time sub-area, train the Elman-RNN model, and use the neural network model to characterize the change of traffic characteristics. Real-time traffic feature extraction, the principle is to process and analyze the travel speed data in the current time period, and obtain the traffic characteristic statistics. The anomaly identification is to compare the real value of the predicted range obtained by the Elman-RNN model of the historical trip speed training to determine whether a traffic anomaly event occurs.

According to the combination of the embodiments of the invention, the implementation is given below. Embodiment 1

Step 11. Using the equidistant space-time division method, determining the segment scale of the time dimension, the time segment span is a fixed value, usually taking 30 mm as a time segment; determining the segment scale of the spatial dimension, the spatial segment span is a fixed value, usually taking 200m×200m The spatial grid acts as a spatial fragment.

Step 12. Perform data preprocessing to perform data cleaning, data integration, data conversion, and data reduction on the GNSS positioning data to improve the structural degree of the data.

Step 13. Divide the spatial area to be processed into a grid of a certain size, and the range of each grid area can be expressed as ={(x _s , y _s )\x _s G[x _r , x _r+1 ) , y _s &[y _r , y _r+1 )}; determine the grid area where the anchor point is located, and use the distance and azimuth to search for the section where the anchor point is located; search for the nearest section of the point A, take the threshold ^= 2.5°, when the difference between the traveling direction angle satisfying point A and the direction angle of the road segment ^ is less than the threshold value, that is, | - | < is satisfied, and the matching is completed; if | - | < is not satisfied, the road segment is deleted in the search space and continues Search for other road segments until the conditions are met; use the straight line equation of the road segment (if it is a curved road segment, it is roughly split into straight lines), calculate the projection coordinates of the GNSS positioning point on the road segment, and reduce the error caused by the GNSS positioning drift. To: Determine the straight line equation of the road segment ^ (if the road segment is a curve, divide it into several straight line segments):

Yy, =k(xx,) where the slope is: k:

The projection line equation is:

Ky _A -ky _t +k ² x _t +x _A

Solve the projected coordinates p is:

k ² +\

k ² y _A + y _t +kx _A - x _j

y _P

k ² +l

After the map matching process, the anchor point is matched to the spatio-temporal sub-area in combination with the timestamp data of the coordinates of the positioning point.

Step 14. The overall vehicle speed data of each of the secondary floating cars in a time and space sub-region constitutes an overall. Calculate the travel speed of each vehicle in the time-space sub-region: where ₂ ... is the distance between the first and second GNSS anchor points in the space-time sub-region f, ..., "-1 The distance from the nth GNSS anchor point is the first time in the space-time sub-region, ..., the time stamp of the GNSS anchor point; the data in each spatio-temporal sub-region is not filtered. , constitute a collection V _{ for subsequent processing.

Step 15. Establish an Elman-RNN neural network consisting of an input layer, an implicit layer, an output layer, a context layer, and a bias node. The number of neurons in the input layer is set to 1, and the number of neurons in the hidden layer is set to 5. The number of neurons in the output layer is set to 1, the number of neurons in the context layer is the same as the number of hidden layers, and is also set to 5, and the initial value of the offset node is set to 0; the initial parameters of each component are set, where The learning rate is set to 0.1, the initial temperature in the simulated annealing module is set to 10 ⁵ , the termination temperature is set to 10 - ² , and the number of iterations per temperature is set to 100. Using ^ as the input data, the neural network is trained and finally trained. Good RN model. Step 16. The real-time traffic data is averaged in the current time window / _rt ( _irt ) = "^5 , as the characterization step 17 of implementing the traffic situation, and the time series data subscripts to be processed by the abnormal point detection are sorted in ascending order. The sequence of points _Vl , v ₂ ,...,v„, known to fit the model based on historical data^, calculate the predicted value of the fitted model

(V); Calculate the difference between the predicted value of the model and the true value

Step 18. Normalize the difference in velocity distribution of each space-time sub-region to a normalized value of 0~1:

Diff^ -m {diff)

ξ' max - min [diff^

Calculate the traffic anomaly index of each time and space sub-region

10. Embodiment 2

Step 21: Using the equidistant space-time division method, determining the segment scale of the time dimension, the time segment span is a fixed value, usually taking 30 mm as a time segment; determining the segment scale of the spatial dimension, the spatial segment span is a fixed value, usually taking 200m×200m The spatial grid acts as a spatial fragment.

Step 22. Perform data preprocessing to perform data cleaning, data integration, data conversion, and data reduction on the GNSS positioning data to improve the structural degree of the data.

Step 23: Divide the spatial area to be processed into a grid of a certain size, and the range of each grid area can be expressed as ={(x _s , y _s )\x _s G[x _r , x _r+1 ) , y _s &[y _r , y _r+1 )}; determine the grid area where the anchor point is located, and use the distance and azimuth to search for the section where the anchor point is located; search for the nearest section of the point A, take the threshold ^= 2.5°, when the difference between the traveling direction angle satisfying point A and the direction angle of the road segment ^ is less than the threshold value, that is, | - | < is satisfied, and the matching is completed; if | - | < is not satisfied, the road segment is deleted in the search space and continues Search for other road segments until the conditions are met; use the straight line equation of the road segment (if it is a curved road segment, it is roughly split into straight lines), calculate the projection coordinates of the GNSS positioning point on the road segment, and reduce the error caused by the GNSS positioning drift. To: Determine the straight line equation of the road segment ^ (if the road segment is a curve, divide it into several straight line segments):

Yy, =k(xx,) where the slope is:

The projection line equation is:

Ky _A -ky _t +k ² x _t +x _A

Solve the projected coordinates p is:

k ² +l

k ² y _A + y _t +kx _A - x _j

y _P

After the map matching process, combined with the timestamp data of the coordinates of the positioning point, the positioning point is matched to the space-time sub-region ₍ Step 24: Calculate the travel speed of each vehicle in the space-time sub-zone: ="^ ~ ² ' ³ +++ "- ''", where ^ ... _1?! is the first and the first in the space-time sub-region The distance between two GNSS anchor points, ..., the distance between the -1 and the nth GNSS anchor points, ^... is the first in the space-time sub-region, ..... ., "Timestamp of a GNSS anchor point; specify the maximum number of clip segments at the same time segment length; ^ «; Search for velocity data within the time segment of a time-space sub-region, if the velocity data within the time segment The number exceeds the upper limit p and randomly f data is added to ^. .

Step 25: Establish an Elman-RNN neural network consisting of an input layer, an implicit layer, an output layer, a context layer, and a bias node. The number of neurons in the input layer is set to 1, and the number of neurons in the hidden layer is set to 5. The number of neurons in the output layer is set to 1, the number of neurons in the context layer is the same as the number of hidden layers, and is also set to 5, and the initial value of the offset node is set to 0; the initial parameters of each component are set, where The learning rate is set to 0.1, the initial temperature in the simulated annealing module is set to 10 ⁵ , the termination temperature is set to 10 - ² , and the number of iterations per temperature is set to 100. Using ^ as the input data, the neural network is trained and finally trained. Good RN model.

1

Step 26: The real-time traffic data is averaged in the current time window _rt ) = "^J , as a representation of the implementation of the traffic situation. Step 27, the point sequence data to be processed by the abnormal point detection is sorted in ascending order The sequence _Vl , v ₂ ,...,v„, known to fit the model based on historical data^, calculate the predicted value of the fitted model

(V); Calculate the difference between the predicted value of the model and the true value ^d tff .

Step 28: Normalize the difference in velocity distribution of each space-time sub-region to a normalized value of 0~1:

Diff^ -m {diff) Calculate the traffic anomaly index for each time and space sub-region

Embodiment 3

Step 31: Using a non-equidistant space-time division method, for a central area of the city where the road network density is greater than ² km/km ² or the peak hour flow is greater than 1000 vehicles/hour, a 30 min time segment and a 200 m×200 m spatial segment are taken, and the road network density is less than 2km/km ² or a suburb of a city with a peak hour flow of less than 1000 vehicles/hour, take a 30-minute time segment and a 400mX400m space segment.

Step 32: Perform data preprocessing, perform data cleaning, data integration, data conversion, and data reduction on the GNSS positioning data to improve the structural degree of the data.

Step 33: Divide the space area to be processed into a grid of a certain size, and the range of each grid area can be expressed as

4 ={( , ) ε[ _Γ , _{Γ +1} ), ^ e[n)};

The floating vehicle GNSS data acquisition frequency is expressed as f ₀ =\l. The point P(t _A -t ₀ ), Pfc+io) adjacent to A in time is defined as the 1-adjacent point of A, P(t _A - 2t ₀ ), ^04+23⁄4) is defined as the 2-adjacent point of A, and so on, then Ρθ4-/3⁄4;), defined as the /- neighbor of Α. When / _Q <lHz, take ^ =1 or 2. Take the distance between the neighboring points of the distances A and A and calculate the A and A

^The mean value of the driving direction angle of the neighboring point ^^, take the threshold ^=5°, if | .- |<4 is satisfied, the matching is completed; otherwise, the other road segments are searched until the condition is met. Use the straight line equation of the road segment (if it is a curved road segment, it is roughly split into straight lines), and calculate the projection of the GNSS positioning point on the road segment to reduce the error caused by the GNSS positioning drift. The specific method is:

Determine the straight line equation of the road segment (if the road segment is a curve, divide it into several straight line segments): yy^Hx-x,) where the slope is:

The projection line equation is: _A =- ^x - ky _A -ky _t +k ² x _t +x _A

Solve the projected coordinates p is:

k ² +\

-h , -kx,

k ² +l

After the map matching process, combining the timestamp data of the coordinates of the positioning point, the positioning point is matched to the space-time sub-region ₍ step 34, calculating the travel speed of each vehicle in the space-time sub-region: where ₂ ... 4 - 1, « is The distance between the first and second GNSS anchor points in the space-time sub-region, ..., the distance between the -1 and the n-th GNSS anchor points, ^... is the spatio-temporal sub-region Within the first, ..., the first "timestamp of a GNSS anchor point; specify the maximum length of the clip segment at the same time segment length; ^ «; search within a time-space sub-region within the time segment Speed data, if the number of speed data in the time segment exceeds the upper limit p, randomly take p data to join ^.

Step 35: Establish an Elman-RNN neural network consisting of an input layer, an implicit layer, an output layer, a context layer, and a bias node. The number of neurons in the input layer is set to 1, and the number of neurons in the hidden layer is set to 5. The number of neurons in the output layer is set to 1, the number of neurons in the context layer is the same as the number of hidden layers, and is also set to 5, and the initial value of the offset node is set to 0; the initial parameters of each component are set, where The learning rate is set to 0.05, the initial temperature in the simulated annealing module is set to 10 ⁵ , the termination temperature is set to 10 - ² , and the number of iterations per temperature is set to 200. Using ^ as input data for neural network training, and finally training Good RN model.

Step 36: Using the rolling time domain mean method, using the travel speed data obtained by data sampling, calculating the current space sub-zone travel speed

1

The mean ( _rt ) - ∑∑ _{v v of the} most recent M time sub-zones, where M = 3, is a representation of the real-time traffic situation.

M-N..

Step 37: The sequence _V1 , v ₂ , . . . , v of the points in which the time series data to be processed by the abnormal point detection is sorted in ascending order is known, and the model fitted according to the historical data is known to calculate the fitted model. Predictive value

(V); Calculate the difference between the predicted value of the model and the true value" = | -〃 "

Step 38: Normalize the difference in velocity distribution of each spatiotemporal sub-region to a normalized value of 0~1:

Calculate the traffic anomaly index of each time and space sub-region

Claims

Claim

1. A road traffic anomaly detection method for non-equidistant space-time division, comprising the following steps:

1) Establish space-time sub-area: divide the day into several time segments, each time segment is called a time sub-region; divide the implementation area of urban road traffic anomaly detection into several spatial segments, each spatial segment is called a space sub- Area; the intersection of any one of the time sub-areas and any one of the spatial sub-areas is called a spatio-temporal sub-area;

2) Preprocessing of historical trajectory data: processing the floating vehicle GNSS positioning historical data as the sampled vehicle speed data of the historical trajectory; Preprocessing of the real-time trajectory data: processing the floating vehicle GNSS positioning real-time data into the sampled vehicle speed data of the real-time trajectory;

3) Historical trajectory data analysis and RNN model training: The sampled vehicle speed data of the historical trajectory is organized and organized, and the RNN model is trained to obtain the historical vehicle speed characteristic model M _RNN ;

Real-time trajectory data analysis and feature extraction: using the sampled vehicle speed data of the real-time trajectory to calculate parameters that can reflect real-time traffic characteristics;

4) Anomaly detection: The difference between the historical vehicle speed characteristic model and the real-time traffic characteristic is measured by the difference index, and the difference between the historical and real-time traffic characteristics is obtained;

5) Quantitative characterization of abnormal severity: using the historical and real-time traffic feature difference values to calculate the traffic condition anomaly index;

6) System performance evaluation: Evaluate the accuracy of traffic abnormality detection and measure the stability of system operation.

2. The road traffic anomaly detecting method according to claim 1, wherein: step 1) adopting one of the following methods:

La) Non-equidistant space-time division method based on road network density: Based on road network density as a judgment index, when the road network density is greater than or equal to 2km/km ² , take 30min time segment and 200m X 200m space segment; When the density is less than 2km/km ² , take a 30 min time segment and a 400 m X 400 m spatial segment;

Lb) Non-equidistant space-time division method based on peak hour flow: based on peak hour flow as a judgment indicator, when the peak hour flow rate is greater than or equal to 1000 vehicles/hour, take 30mm time segment and 200mX 200m space segment; When less than 1000 vehicles/hour, take a 30-minute time segment and a 400-mX 400-meter space segment.

3. The road traffic anomaly detection method according to claim 1, wherein the preprocessing of the historical trajectory data in step 2) comprises:

2a) Data structuring: Data cleaning, data integration, data conversion, and data reduction of floating vehicle GNSS positioning history data to obtain structured GNSS positioning history data;

2b) Fast map matching: Combine the urban road network data, map the structured GNSS positioning historical data to the urban road network through the map matching algorithm, and establish the matching relationship between the positioning points and the road segments in the structured GNSS positioning historical data, Forming a matching relationship between the positioning point and the road segment in the structured GNSS positioning history data, and correcting the error caused by the positioning drift; 2c) calculating and sampling the traffic operation characteristic parameter: calculating according to the structured GNSS positioning history data The traffic operation characteristic parameter obtains the traffic characteristic data of the historical trajectory, and performs data sampling on the traffic characteristic data of the historical trajectory to obtain the sampled traffic characteristic data of the sampled historical trajectory.

4. The road traffic anomaly detection method according to claim 3, wherein the fast map matching in step 2b) comprises:

2b 1) Divide the space area to be processed into a grid of a certain size, and the range of each grid area can be expressed as 4 ={(x _s ,y _s )\x _s e[x _r ,x _r+1 ), &[y _r ,y _r+1 )}^ Each grid area contains several sections, which are R _s represent a collection of the links of each set of R _s and is expressed as given link number to each segment;

2b2) Determine the grid area where the anchor point is located, and use the distance and azimuth angle to search for the section where the anchor point A is located in the set of road segments.

2b3) Calculate the projection coordinates of the GNSS anchor point on the road segment using the GNSS anchor point linear projection method.

The road traffic anomaly detecting method according to claim 4, wherein the step 2b2) adopts one of the following methods:

2b21) Single point matching method: Search for the nearest road segment from a certain positioning point A. The implementation method is: For a certain road segment ij in the set of road segments, the difference between the traveling direction angle satisfying the point A and the direction angle of the road segment ij is smaller than the threshold value. Satisfy

| - |<, complete the match; if not satisfied with | - |<, continue to search for other road segments in the road segment collection until | - |<_;

2b22) Point sequence matching method: This scheme is applicable to high frequency floating car data; the time interval of floating car GNSS data of each two adjacent time is expressed as ί. , the floating vehicle GNSS data acquisition frequency is expressed as / _Q = l / i _Q , and the time record of a certain positioning point A is represented as a point P (t _{A -} t ₀ ) that is temporally adjacent to the positioning point A, Pfc+if is defined as 1-adjacent point of A, P(t _A -2h), P04+2i _Q ;) is defined as 2-adjacent point of a certain anchor point A, and so on, then Pi -kt^h is defined as a neighboring point of a certain positioning point A; when / _Q <lHz, take ^ =1 or 2; take a section of the distance from a certain positioning point A and the adjacent point of the positioning point ^ and calculate the positioning point and The mean value of the driving direction angle of the adjacent point ^4i, if |_|< is satisfied, the matching is completed; otherwise, the other road segments are searched until the |1|< is satisfied.

The road traffic anomaly detecting method according to claim 3, wherein the preprocessing of the historical trajectory data in step 2c) adopts one of the following methods:

2cl) Full sample method: The overall speed data of each sub-floating vehicle in a time-space sub-region is composed of the overall speed. The implementation method is to calculate the travel speed of each vehicle in the space-time sub-area: ν _ξ = (d ₁₂ +d ₂₃ +... + d _n _ _ln )/(t _n -t _x ), where ^...i is the distance between the first and second GNSS anchor points in the spatiotemporal sub-region, .... .., the distance between the n-1th and the nth GNSS anchor point, which is the first time in the space-time sub-region, ..., the time stamp of the first GNSS anchor point; The travel speed data in the area is not filtered, and constitutes a set ^ for subsequent processing;

2c2) Time-smooth sampling method: Specify the length of the time segment, set the upper limit of the number of segments of the same time; Search the velocity data in each time segment of a space-time sub-region, if the number of velocity data in the time segment exceeds the upper limit, randomly take the upper limit bar The number of data is used for subsequent processing. The implementation method is to calculate the travel speed of each vehicle in the space-time sub-region: ν _ξ = (d + d ₁₃ +... + d _n _ _ln ) / (t _n -? where ₂ . ..4—the distance between the first and second GNSS anchor points in the space-time sub-region f, ..., the distance between the -1 and the GNSS anchor points, in time and space The first time in the sub-area, ..., the time stamp of the GNSS anchor point; the specified time clip length is set to the upper limit of the number of clips at the same time P. Searching for a time-space sub-zone within the i-th time segment Speed data, if time data within the time segment P exceeds the upper limit number of pieces of data added randomly _∞ ^ and for subsequent processing.

The road traffic anomaly detecting method according to any one of claims 1 to 6, wherein the historical trajectory data analysis and the RNN model training in step 3) take the following steps:

3a) Establish the basic structure of the Elman-RNN neural network, including setting the number of neurons in the input layer, setting the number of neurons in the hidden layer, setting the number of neurons in the output layer, and setting the number of neurons in the context layer. Set the offset nodes of the input layer and the output layer and assign initial values;

3b) setting the basic parameters of the Elman-RNN neural network model, including the parameter learning rate of the back propagation algorithm for model training, the initial temperature and the termination temperature of the simulated annealing, and the number of iterations of each temperature;

3c) using the sampled vehicle speed data of the historical trajectory to perform training of the RNN model;

3d) Processing the model after a single training, including updating the weight of the neural network, evaluating the error rate of the model, and updating the strategy when the error rate has not decreased;

3e) performing the conditional judgment of the model training, if the degree of improvement of the model is less than the minimum value is greater than the set threshold, the algorithm terminates, and the RNN model M _{RNN is obtained} ;

Step 3) The real-time trajectory data analysis and feature extraction are performed by one of the following methods:

3f) Time window mean method: Using the sampled vehicle speed data of the real-time trajectory, calculate the mean value of the current space sub-zone travel speed in the current time sub-region // _rt (V _{i rt} ) = ~^3⁄4 _V , as real-time traffic characteristics Representation parameter

1~ί

3g) Rolling time-domain mean method: Using the sampled vehicle speed data of the real-time trajectory, calculating the mean value of the current space sub-zone travel speed in the most recent M time sub-regions / _rt (v _{i rt} ) = ~^£^ ,, Take 3~5 as the representation parameter of real-time traffic characteristics.

8. The road traffic anomaly detecting method according to claim 7, wherein the step 3a) comprises the following steps:

3al) Set the number of neurons in the input layer to 1;

3a2) Set the number of neurons in the hidden layer to 5~8;

3a3) Set the number of neurons in the output layer to 1;

3a4) setting the number of neurons in the context layer to be the same as the number of neurons in the hidden layer;

3a5) Set an offset node for each of the input layer and the output layer, and the initial value is set to 0;

Step 3b) contains the following steps:

3bl) Set the parameter learning rate when the backpropagation algorithm is used for model training to 0.01~0.8;

3b2) the initial temperature of the simulated annealing is set to ^105, simulated annealing end temperature is set to ^10-2;

3b3) Set the number of iterations per temperature to 100.

9. The road traffic anomaly detecting method according to claim 7, wherein the step 3c) comprises the following steps:

3d) Sorting the sampled vehicle speed data of the historical trajectory of each spatial sub-area by time, and synthesizing (input, output) pairs of the sampled vehicle speed data of the sorted historical trajectory, ie ( ), ^, ^), The form of ... , ^^, ^);

3c2) Create an Elman-RNN neural network, where the number of input layer neurons is 1, and the number of hidden layer neurons is 5, 6, and 7. Or 8, the number of neurons in the output layer is 1, the context layer saves the output of the hidden layer at the previous time, the number of neurons is the same as the hidden layer; using the Sigmoid activation function;

3c3) Set the input layer to the hidden layer, the hidden layer to the output layer, the weight of the connection between the neurons is a random value between 0~1; respectively set the input layer and the hidden layer of the offset unit and initialize to 0;

3c4) Using a back propagation algorithm and a simulated annealing algorithm to form a hybrid strategy to train the model.

10. The road traffic anomaly detecting method according to claim 7, wherein the step 3d) adopts one of the following methods:

3dl) Greedy strategy: If the error rate of the model does not decrease after a certain training, the recovery weight and error rate are the values before training; 3d2) Hybrid strategy: If the error rate of the model does not decrease after a certain training, or the magnitude of the decline is less than The set minimum value is trained using the simulated annealing algorithm; the training step of the simulated annealing algorithm is as follows: First, calculate the error score of the current model, and then the weight of the connection between the neurons of the current model and the value of the bias unit, Add a random number "ί«, get the new weight and the value of the offset unit, in the real add = {0.5 - Random) I startTemp emp, where Random is a random number, the range is greater than 0 is less than 1, startTemp is the initial Temperature, take 10 ⁵ , temp is the current temperature; calculate the updated model error score, if the new error score is smaller than the current error score, indicating that the new weight improves the performance of the model, then save the new weight, otherwise discard; The current temperature is multiplied by a fixed ratio ratio to lower the temperature: ratio = exp (log (stopTemp I startTemp) / (cycles - \)), wherein, stopTemp temperature to terminate, taking 10- ^2, cycles of a number of iterations of training, take 100; the process is repeated cycles times.

The road traffic abnormality detecting method according to any one of claims 1 to 6, wherein the step 4) the abnormality detecting comprises the following steps: 4a) pressing the subscript of the sampled vehicle speed data of the historical trajectory to be processed by the abnormality detecting The sequence of points sorted in ascending order vv ₂ ,... , v _n ; The data is expressed as a two-dimensional point in plane coordinates as (1 ;),.

4b) Calculate the predicted value of the RNN model i = M _RNN ( ), and calculate the difference between the predicted value of the model and the true value [^'^ ] = 1 "^'1; Step 5) Quantitative characterization of the abnormal severity includes:

5a) will be in each time and space sub-area

5b) Calculate the traffic anomaly index for each time-space subzone

10.