Method, system and device for judging and positioning traffic accident on non-monitored road section
Technical Field
The invention belongs to the field of intelligent traffic, and relates to a machine learning method for diagnosing and positioning traffic accidents.
Background
Along with the development of smart cities, road monitoring facilities are increasingly complete, and vehicle accidents in the working range of monitoring equipment can be quickly captured by a deep learning solution based on computer vision. However, the complexity of the road network determines that it is not possible to deploy monitoring equipment everywhere on the road. Especially for urban trunks, elevated roads and the like which carry a large number of vehicles to pass through in special time periods, once a traffic accident occurs, traffic jam can be directly caused. If the time sequence data collected by the limited road monitoring equipment can be used for diagnosing and positioning accidents on non-monitored road sections, the traffic police can be helped to quickly intervene in accident sites to process and dredge traffic, and important value is provided for creating intelligent road networks.
The current traffic accident alarm solution can be divided into two types: alarm piles or other sensing equipment paved on the road side, and vehicle-mounted accident detection alarm equipment. The former still requires equipment to be laid on a complex road network, the latter requires additional equipment to be installed on a vehicle, and both hardware-based solutions are costly and inconvenient to popularize. The solution of the method is based on the existing monitoring points of the road network, adopts a machine learning method to mine the speed and flow data of the monitoring points, captures the data abnormity, and carries out positioning inference on the trigger points of the abnormal data by combining the relative spatial position, road information, time information and the like among the monitoring points. Low cost and convenient practical popularization and application.
Some researchers use machine learning and big data to predict city anomalies, for example, the track of taxi drivers is used to predict city anomalies. The traffic abnormity is captured according to the change of the route selection modes of drivers, and keywords are further extracted from related microblogs to explain reasons of the abnormity. The methods are essentially different from the solution scheme of the text, and the text adopts the traffic of the road monitoring points and the vehicle speed information for judgment, and the taxi driving track, the driver route and the microblog information are not communicated. Meanwhile, the method focuses on rapid positioning of accidents, the urban road is divided and modeled spatially according to the monitoring points, and the characteristic change caused by abnormality is captured by the nearest monitoring point.
Disclosure of Invention
1. Objects of the invention
The invention can quickly judge and position the abnormal condition of the road of the whole road network based on the data collected by the existing road detection equipment.
2. The technical scheme adopted by the invention
The invention provides a method for judging and positioning traffic accidents on non-monitored road sections, which comprises the following steps:
dividing the road network into road sections based on the road network monitoring points and the relative spatial positions thereof, and monitoring the traffic flow, the speed, the traffic flow leaving from the road sections and the speed of vehicles entering the road sections at two ends; performing bidirectional modeling of a road section: dividing the road network into road sections, vehicles entering the road sections and vehicles leaving the road sections according to the monitoring points, and dispersing the whole road network model into mutually independent units by model diagnosis and positioning according to the monitoring data at the two ends of the road sections;
monitoring points at two ends of a road section, wherein traffic accidents occurring at a certain moment and position indicate the speed and the flow characteristic information of the road section to be transmitted to the monitoring points at the two ends, and the transmission speed and the influence condition are influenced by road characteristics; and learning road information of a specific road section through a neural network model training parameter: inputting the variation vectors of the vehicle speed of the monitoring points at two ends at different moments and the variation vectors of the flow at different moments, and mapping the data of four time sequences to a [0,1] interval according to the first n moments in a standardized manner; and recording the current monitoring state.
Selecting m x 4 channel data within the range from the current time n to m before as input in the monitoring process, wherein m is less than n; extracting time sequence characteristics by adopting an RNN (radio network); the depth is 2 layers; for the eigenvector output y output by the double-layer RNN model, which is [ y11, y12, … and y1m ], whether an accident occurs or not is judged through two fully-connected layers and a Sigmoid activation function output1 which is Sigmoid (W1 y + b1), output2 which is Sigmoid (W2 y + b2) dab which is used for judging the distance from an accident occurrence point to a point, and dab which is the distance between detection points ab.
The invention provides a system for judging and positioning traffic accidents on non-monitored road sections, wherein a road network model comprises the following components:
a model preprocessing module; dividing the road network into road sections based on the road network monitoring points and the relative spatial positions thereof, and monitoring the traffic flow, the speed, the traffic flow leaving from the road sections and the speed of vehicles entering the road sections at two ends; performing bidirectional modeling of a road section: dividing the road network into road sections, vehicles entering the road sections and vehicles leaving the road sections according to the monitoring points, and dispersing the whole road network model into mutually independent units by model diagnosis and positioning according to the monitoring data at the two ends of the road sections;
a model building module; monitoring points at two ends of a road section, wherein traffic accidents occurring at a certain moment and position indicate the speed and the flow characteristic information of the road section to be transmitted to the monitoring points at the two ends, and the transmission speed and the influence condition are influenced by road characteristics; and learning road information of a specific road section through a neural network model training parameter: inputting the variation vectors of the vehicle speed of the monitoring points at two ends at different moments and the variation vectors of the flow at different moments, and mapping the data of four time sequences to a [0,1] interval according to the first n moments in a standardized manner; and recording the current monitoring state.
An accident monitoring module: selecting m x 4 channel data within the range from the current time n to m before as input in the monitoring process, wherein m is less than n; extracting time sequence characteristics by adopting an RNN (radio network); the depth is 2 layers; for the eigenvector output y output by the double-layer RNN model, which is [ y11, y12, … and y1m ], whether an accident occurs or not is judged through two fully-connected layers and a Sigmoid activation function output1 which is Sigmoid (W1 y + b1), output2 which is Sigmoid (W2 y + b2) dab which is used for judging the distance from an accident occurrence point to a point, and dab which is the distance between detection points ab.
The invention provides a device for judging and positioning traffic accidents on non-monitored road sections, which comprises a processor and a memory, wherein the processor is used for processing the traffic accidents;
one or more processors, and
one or more programs stored in the memory and configured to be executed by the one or more processors, the programs, when executed by the processors, implementing the methods.
3. Advantageous effects adopted by the present invention
(1) The invention adopts the monitoring data to judge and position the accident of the section which is not monitored;
(2) the invention can realize accident monitoring and positioning on the basis of the road section monitoring model.
In conclusion, the invention can quickly find and position traffic accidents on the road sections with limited road traffic capacity, process the traffic accidents in time, increase the road traffic capacity and relieve the peak road pressure.
Drawings
FIG. 1 is a schematic diagram of a road and modeling.
Fig. 2 is a schematic diagram of the propagation of characteristic parameters in road space and time when an accident occurs on a road section.
Fig. 3 is a diagram of a model for extracting time series characteristics.
FIG. 4 is a flow chart of the internal calculation for each neuron (amplitude f1/f 2).
Detailed Description
The technical solutions in the examples of the present invention are clearly and completely described below with reference to the drawings in the examples of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments of the present invention without inventive step, are within the scope of the present invention.
The present invention will be described in further detail with reference to the accompanying drawings.
Example 1
The invention provides a method for judging and positioning abnormity occurring in a road section by adopting traffic flow and speed data of a road section close to a monitoring point.
Mainly solves two technical problems:
firstly, feature extraction and diagnosis are carried out on the nearest monitoring points
Based on the existing road network monitoring points and the relative spatial positions thereof, performing segmentation modeling on the road network; the abnormal condition of the road section is subjected to feature extraction and diagnosis by the nearest monitoring point;
secondly, judging the abnormal traffic condition and positioning the accident point
And establishing a corresponding machine learning algorithm, judging whether abnormal conditions affecting traffic occur in the road section, and determining the position of the accident point.
1. Road network model
The road network is divided into road sections based on the road network monitoring points and the relative spatial positions of the road network monitoring points, and the traffic flow, the speed, the leaving traffic flow and the speed of vehicles entering the two ends of the road sections are monitored. The following points are observed in the modeling process:
a. and (4) respectively modeling in two directions by considering the traffic flow direction, namely the same road. As shown in a section of road on the upper graph of fig. 1(a), 1 and 2 are monitoring points for monitoring flow and vehicle speed, the bidirectional driving section is modeled by two sections: 1_2,2_ 1. The traffic flow and the speed of the vehicle entering and exiting the road section at the position 1 and entering and exiting the road section at the position 2 are respectively monitored.
b. The road network is divided into road sections according to monitoring points and the modes of road section vehicle inflow and vehicle outflow. And diagnosing and positioning the abnormal traffic condition on each road section through the monitoring data at the two ends of the road section through a model. The X-5 section is common to the 3-5 and 4-5 sections as shown in fig. 1(b), and thus traffic abnormality on the X-5 section is diagnosed in the monitoring models of both sections.
Due to the diffusion of road network influence caused by special events such as city road network complexity and traffic accidents, the accuracy of positioning abnormity on a city scale is insufficient. After the modeling method is used for segmentation, the abnormity of each road section is only responsible for inflow and outflow end data of the road section, which is equivalent to the fact that the whole road network model is scattered into units which are independent from each other, and only a space-time model shown in figure 2 needs to be considered when an accident monitoring model is established. Considering time dispersion, a model of diffusion of traffic flow and vehicle speed from accident sources:
P[d][t]=f(d-d0,t-t0,P[d][t-1],roadfactor…)
wherein P represents monitoring information such as vehicle speed v, vehicle flow q and the like. In the road section from the monitoring point a to the monitoring point B shown in fig. 2, a traffic accident occurring at the time t0 and the position d0 may cause the speed V of the road section, the traffic Q characteristic information is propagated to the monitoring point, the propagated speed and the influence condition are influenced by the road characteristic roadfactor, and the road characteristic includes road condition information such as the number of lanes, the speed limit of a lane, and the like. The road information of a specific road section can be learned in training parameters, so that the model solution of the whole road network accident monitoring is converted into the following machine learning problem, and the input is as follows: [ Va0, Va1 … Van … ] shows the change of the vehicle speed at the monitoring point a at the time points 0 and 1 … n …, and [ Qa0, Qa1 … Qan ] shows the change of the flow rate at the monitoring point a at the time points 0 and 1 … n …. [ Vb0, Vb1 … Vbn … ] shows the change of the vehicle speed at monitoring point b at time 0,1 … n …, [ Qb0, Qb1 … Qbn ] shows the change of the flow rate at monitoring point b at time 0,1 … n …, specifically:
1. the four time series data are mapped to a [0,1] interval according to the first n time standardization, and the following graph formula is shown:
time V is 0-n
aThe maximum value of (a) is,
time V is 0-n
aIs measured.
2. The four time sequence data after mapping 0-1 are as follows, and the right superscript n is recorded as the monitoring state at the current moment:
3. and selecting m-4 channel data within the range from the current time n to m before as input in the monitoring process, wherein m is less than n.
4. The invention adopts a standard RNN network to extract time sequence characteristics, and the model is shown in figure 3. In box is a standard 2-layer depth RNN model. The internal calculation process for each cell, i.e. neuron (amplitude f1/f2), is shown in FIG. 4.
5. For the eigenvector output y output by the double-layer RNN model, which is [ y11, y12, … and y1m ], whether an accident occurs or not is judged through two fully-connected layers and a Sigmoid activation function output1 which is Sigmoid (W1 y + b1), output2 which is Sigmoid (W2 y + b2) dab which is used for judging the distance from an accident occurrence point to a point, and dab which is the distance between detection points ab.
The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.