CN109345832A - A kind of urban road based on depth recurrent neural network is overtaken other vehicles prediction technique - Google Patents
A kind of urban road based on depth recurrent neural network is overtaken other vehicles prediction technique Download PDFInfo
- Publication number
- CN109345832A CN109345832A CN201811351157.3A CN201811351157A CN109345832A CN 109345832 A CN109345832 A CN 109345832A CN 201811351157 A CN201811351157 A CN 201811351157A CN 109345832 A CN109345832 A CN 109345832A
- Authority
- CN
- China
- Prior art keywords
- vehicles
- data
- vehicle
- overtaking
- neural network
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/01—Detecting movement of traffic to be counted or controlled
- G08G1/0104—Measuring and analyzing of parameters relative to traffic conditions
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/01—Detecting movement of traffic to be counted or controlled
- G08G1/0104—Measuring and analyzing of parameters relative to traffic conditions
- G08G1/0125—Traffic data processing
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/01—Detecting movement of traffic to be counted or controlled
- G08G1/017—Detecting movement of traffic to be counted or controlled identifying vehicles
- G08G1/0175—Detecting movement of traffic to be counted or controlled identifying vehicles by photographing vehicles, e.g. when violating traffic rules
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/60—Type of objects
- G06V20/62—Text, e.g. of license plates, overlay texts or captions on TV images
- G06V20/625—License plates
Landscapes
- Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Analytical Chemistry (AREA)
- Chemical & Material Sciences (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Multimedia (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Traffic Control Systems (AREA)
Abstract
The invention belongs to machine learning field, overtake other vehicles rate prediction technique more particularly to a kind of urban road based on depth recurrent neural network.Using the Vehicle License Plate Recognition Systems such as electronic police and bayonet, is compared by the time of upstream and downstream Car license recognition, more can accurately obtain overtake other vehicles relationship of the vehicle between section.The model has higher precision, stronger generalization ability compared with traditional neural network.The present invention is capable of the trend of overtaking other vehicles of more acurrate predicted city road, has not only ensured traffic safety to a certain extent, but also provide decision support for regulatory authorities.
Description
Technical field
The invention belongs to machine learning fields, more particularly to a kind of urban road based on depth recurrent neural network
Rate of overtaking other vehicles prediction technique.
Background technique
Overtake other vehicles always social question of common concern, especially large area of urban road is overtaken other vehicles phenomenon, is not only buried
Security risk, or even jeopardize the security of the lives and property of people.Realize to urban road overtake other vehicles problem it is timely manage be current
The problem of urgent need to resolve.The stream of overtaking other vehicles of accurate predicted city specific road section will provide forcefully decision support for law enforcement agency,
It greatly ensure that the stability and security of traffic order in urban road.It is most common in driving procedure that urban road, which is overtaken other vehicles,
Behavior, for overtaking other vehicles, problem domestic and foreign scholars have carried out extensive research.
Lyapunov function design intelligent vehicle lane-change is chosen based on Backstepping control algolithm and is overtaken other vehicles in trip peak etc.
Contrail tracker realizes the tracing control to vehicle lane-changing and track of overtaking other vehicles.Liu Jiang is by overtaking other vehicles to two-lane highway
Behavior test research, obtains behavioural characteristic of the different drivers in overtaking process.Zhu Xiujuan etc. is simulated using driving simulator
Dynamic overtaking process judges whether certain behaviors of driver are safe.Afshin etc. has been counted according to Iranian casualty data and has been overtaken other vehicles
The ergonomic characteristics of accident occur for process.It is super on two-lane highway that Eleni I analyzes male and female driver
The vehicle duration, and establish the passing time model for considering driver's gender and speed.
In recent years, deep learning gets more and more people's extensive concerning as a kind of new machine learning algorithm, various depth
The frame of habit applies to artificial intelligence field, such as image classification, speech recognition, natural language processing, computer game, cancer
Detection etc..Deep learning also embodies stronger and stronger application value in terms of traffic big data excavation.Such as Zhang needle
The prediction of expressway travel time is proposed a kind of based on stack autocoder model.It is obtained by freeway toll station
Pick-up is averaged hourage per hour, and the training pattern in such a way that historical data is using unsupervised learning uses gradient descent method
The parameter of adjustment and Optimization Prediction layer, when finally predicting travelling in following one hour using the past three hours hourages
Between.The root-mean-square error of this method is 13.6%, better than traditional BP neural network prediction model.Kumar et al. proposes one
Kind is based on seasonal ARMA model (SARIMA).The model considers season compared to traditional ARIMA model
The influence of section and early evening peak to vehicle flowrate, the model are used only a small amount of input data and carry out short-term forecast to traffic flow.
Jeong etc. utilizes automatic vehicle positioning data, is come using regression model, artificial neural network (ANN) and time series models
Predict bus arrival time.The result shows that ANN model is better than time series models and regression model in terms of precision of prediction.
Zhang Weiwei etc. selects shot and long term Memory Neural Networks (LSTM), the input length of fixed LSTM concealed nodes number, model, test
Different hidden layer number of nodes and consider spatial coherence under the conditions of LSTM estimated performance, and with traditional BP neural network,
ARIMA model, KNN method are compared and are analyzed, the results showed that, LSTM has preferably fitting and Training Capability.
Deep learning in terms of construction of high-tech traffic system using more and more extensive, but deep learning is predicted in urban road
In application it is not extensive.To sum up, to realize more intelligent, precision urban road management, the present invention is based on Suzhou works
The industry garden street Xing Hu-modern times big road segment segment electronic police data propose that one kind is passed based on GRU with the theoretical frame of deep learning
The prediction model of overtaking other vehicles for returning neural network (Gated Recurrent Unit, GRU), for specification driving behavior and to hand over
The alert decision assistant that provides supports have great importance.
Summary of the invention
It overtakes other vehicles prediction the present invention provides a kind of urban road based on depth recurrent neural network, innovatively merges,
It is intended to sufficiently solve the problems, such as precisely to predict.
The present invention adopts the following technical scheme:
Step S1, data are acquired.The data source that the present invention uses is based on Car license recognition detected by electronic police
Information.Electronic police is mounted on each crossroad, pavement mouthful, main main and side road inlet and outlet etc. and can all install, when vehicle is sailed
Stop line is crossed, high-definition digital video camera photographed vehicle tail photo, and electronic police system can be handled according to photo, extracted
Record information is transmitted on data server by information of vehicles.System inputs real-time electronic police number from traffic signal control
According to signal sequence data, weight analysis and processing license plate identification data and time data.
Step S1.1, intersection A and B four direction are equipped with electronic police, and cover all lanes.When vehicle passes through
When stop line, electronic police will record the license plate number and timestamp t of vehicle.With the section L from intersection A to BabFor, one
It is t that vehicle, which is driven out to the stop line time from the intersection A,a, the stop line time is driven out to the crossing B as tb, the journey time of section AB
Tab=tb-ta。
Step S1.2, the vehicle captured by electronic police in intersection A and B, referred to as matching vehicle.Match vehicle it
Between the carry out overtaking analysis that can be convenient.Vehicle 1 and vehicle 2, if t1b< t2bAnd t1a> t2a.Then think vehicle 1 in section
LabIt has been more than vehicle 2.
Step S2, number of overtaking other vehicles is counted.Overtaking other vehicles with lane group (LaneGroup, LG) is that unit is counted, institute of the present invention
Meaning lane group refers to by the combination in the import lane of same signal controlling on lamp group.Only just carried out in the vehicle of same lane group
It overtakes other vehicles comparison, because the current signal phase stage is different, the mutual relationship of overtaking other vehicles is not counted in the vehicle of different lane groups
It overtakes other vehicles statistics.Overtaking other vehicles for a certain lane group certain time period is counted, such as 7 points to 8 points overtake other vehicles.Algorithm is as follows:
1. taking 7 points to 8 points 15 minutes matched datas, each car has this lane timestamp and upstream crossing timestamp, so
It is from morning to night ranked up according to this lane timestamp afterwards.
2. 15 minutes vehicles compare for each car and below, if the upstream timestamp ratio of back vehicle from
Oneself is early, then it is assumed that is to overtake other vehicles, records Overtaken Vehicle licence plate.
3. if the upstream and downstream time tolerance of a vehicle is more than 15 minutes, more than its vehicle not counting overtaking other vehicles.Because it
It may be in intermediate waiting.
Step S3, data prediction.Normalized has been carried out to all data, has seen following formula:
In formula: X indicates data of the present invention of overtaking other vehicles, XminIndicate the minimum value inputted in data of overtaking other vehicles; XmaxIt indicates
Maximum value in data of overtaking other vehicles;X ' expression overtake other vehicles data be normalized after data.What secondly the present invention analyzed mainly overtakes other vehicles
Object is overtaking other vehicles between common vehicle, it is therefore desirable to filter out some special cars.Special car includes police car, ambulance, army
Vehicle, foreign nationality's vehicle etc..It is poor by using the Link Travel Time overtaken other vehicles between passed vehicle, it can not be completely authentic and valid anti-
The cut-in situation for reflecting vehicle needs to reject some abnormal datas, and several situations may generate abnormal data below:
1. vehicle in the driving process of midway, stops because of certain cause specifics, (vehicle casts anchor, midway carrying etc.) is produced
The raw time delays for being higher than normally travel vehicle;
2. the branch that vehicle driving never installs detection device on the way leaves detection interval, lead to electronic police system not
License plate identification data can be matched to;
3. vehicle secondary or multiple trip in section, causes same license plate data repeatedly to go out in electronic police system
It is existing.
Step S4, building depth recurrent neural networks model is for data prediction of overtaking other vehicles.
Depth recurrent neural networks model is GRU neural network, and the input of the GRU neural network is that history is overtaken other vehicles data,
The output of network is the data of overtaking other vehicles of future time instance, and for GRU neural network by memory unit come more new data, a memory is single
Member is made of update door and resetting door, and to control the flowing of intermediate result, specific step is as follows,
ft=sigm (Wf[ht-1, xt]+bf)
it=sigm (Wi[ht-1, xt]+bi)
Wherein, i, f are that the parameter in GRU neural network structure has each served as forgetting and more new historical is overtaken other vehicles data
Effect, W is to the weighting parameters for data of overtaking other vehicles in GRU neural network, and b is bias,Play interim control parameter
Effect, i.e.,To be calculated based on current input and preceding state, obtained by the loss function training of GRU neural network.
Compared with the existing technology, the method for the present invention has the advantage that
(1) potential rule in data of overtaking other vehicles can be more excavated compared with shallow Model using depth Recursive Networks.
(2) there is the Filtering system to effective information using depth recurrent neural network GRU more traditional Recursive Networks,
The disturbance of noise in data of overtaking other vehicles can be more eliminated, so that model has stronger robustness.
(3) extensive use for passing through the Vehicle License Plate Recognition Systems such as electronic police and bayonet, compared with traditional wagon detector
License plate comparison obtains numerous valuable data between can use upstream and downstream intersection, and the data for providing fining are overtaken other vehicles
Prediction
Detailed description of the invention
Fig. 1 is the flow chart of prediction model of overtaking other vehicles;
Fig. 2 is the evolution graph of deep neural network;
Fig. 3 is effect picture of the Grey System Model on test set;
Fig. 4 is section electronic police scheme of installation;
Fig. 5 is the architecture diagram of depth recurrent neural network;
Fig. 6 is full dose data line chart;
Fig. 7 is several scatter plots of overtaking other vehicles of any two moon.
Specific embodiment
To further illustrate that each embodiment, the present invention are provided with attached drawing.These attached drawings are that the invention discloses one of content
Point, mainly to illustrate embodiment, and the associated description of specification can be cooperated to explain the operation principles of embodiment.Cooperation
With reference to these contents, those of ordinary skill in the art will be understood that other possible embodiments and advantages of the present invention.Figure
In component be not necessarily to scale, and similar component symbol is conventionally used to indicate similar component.
Now in conjunction with the drawings and specific embodiments, the present invention is further described.
As shown in fig.1, the flow chart for prediction model of overtaking other vehicles.
Data source:
The data source that the present invention uses is based on Car license recognition information detected by electronic police.Electronic police peace
It can all be installed mounted in each crossroad, pavement mouthful, main main and side road inlet and outlet etc., when vehicle crosses stop line, high definition number
Word video camera photographed vehicle tail photo, and electronic police system can be handled according to photo, extracts information of vehicles, will record
Information is transmitted on data server.System inputs real-time electronic police data and signal sequence number from traffic signal control
According to, weight analysis and processing license plate identification data.
Step S1, data of overtaking other vehicles acquisition
As shown in figure 4, being section electronic police scheme of installation.Intersection A and B four direction are equipped with electronic police,
And all lanes of covering.When vehicle passes through stop line, electronic police will record the license plate number and timestamp t of vehicle.With from
The section L of intersection A to BabFor, a vehicle is driven out to the stop line time from the intersection A as ta, when being driven out to stop line to the crossing B
Between be tb, Link Travel Time Tab=tb-ta。
Step S1.2, the vehicle captured by electronic police in intersection A and B, referred to as matching vehicle.Match vehicle it
Between the carry out overtaking analysis that can be convenient.Vehicle 1 and vehicle 2, if t1b< t2bAnd t1a> t2a.Then think vehicle 1 in section
LabIt has been more than vehicle 2.
Step S2, number of overtaking other vehicles is counted
Overtaking other vehicles with lane group (LaneGroup, LG) is that unit is counted, and the so-called lane group of the present invention refers to same
The combination in the import lane of signal controlling on lamp group.Only comparison of overtaking other vehicles, different lane groups are just carried out in the vehicle of same lane group
Vehicle because the current signal phase stage is different, the mutual relationship of overtaking other vehicles is not counted in statistics of overtaking other vehicles.
Overtaking other vehicles for a certain lane group certain time period is counted, such as 7 points to 8 points overtake other vehicles.Algorithm is as follows:
1. taking 7 points to 8 points 15 minutes matched datas, each car has this lane timestamp and upstream crossing timestamp, so
It is from morning to night ranked up according to this lane timestamp afterwards.
2. 15 minutes vehicles compare for each car and below, if the upstream timestamp ratio of back vehicle from
Oneself is early, then it is assumed that is to overtake other vehicles, records Overtaken Vehicle licence plate.
3. if the upstream and downstream time tolerance of a vehicle is more than 15 minutes, more than its vehicle not counting overtaking other vehicles.Because it
It may be in intermediate waiting.
Using absoluteness evaluation index --- rate is overtaken other vehicles as prediction object in section.
The section rate of overtaking other vehicles is the ratio that the vehicle number that section is overtaken other vehicles accounts for the observed vehicle number in section:
Step S3, data prediction
It is that the rate of overtaking other vehicles has carried out normalized to all data, sees following formula:
In formula: X indicates data of overtaking other vehicles (referring to the rate of overtaking other vehicles) of the present invention, XminIndicate the minimum inputted in data of overtaking other vehicles
Value;XmaxIndicate the maximum value overtaken other vehicles in data;X ' expression overtake other vehicles data be normalized after data.
What the present invention analyzed mainly overtake other vehicles, and object is overtaking other vehicles between common vehicle, it is therefore desirable to filter out some special vehicles
?.Special car includes police car, ambulance, military vehicle, foreign nationality's vehicle etc..By using the link travel overtaken other vehicles between passed vehicle
Time difference, the cut-in situation of reflection vehicle that can not be completely authentic and valid need to reject some abnormal datas, several below
Situation may generate abnormal data:
1. vehicle in the driving process of midway, stops because of certain cause specifics, (vehicle casts anchor, midway carrying etc.) is produced
The raw time delays for being higher than normally travel vehicle;
2. the branch that vehicle driving never installs detection device on the way leaves detection interval, lead to electronic police system not
License plate identification data can be matched to;
3. vehicle secondary or multiple trip in section, causes same license plate data repeatedly to go out in electronic police system
It is existing.
Step S4, building depth recurrent neural networks model is for data prediction of overtaking other vehicles
The depth recurrent neural network that the present invention uses is GRU (Gated Recurrent Unit) neural network, network
Input be historical data data of overtaking other vehicles, the output of network is that data i.e. problem of overtaking other vehicles of future time instance is one monotropic
The time series forecasting problem of amount.Depth is by memory unit (memory cell) using GRU come more new data, one
Cell is made of update door and resetting door, to control the flowing for characteristic information of overtaking other vehicles, uses sigmoid function that it is made to export 0
And the value between 1, therefore the characteristic information of overtaking other vehicles generated can flow between two doors according to this probability value, specifically,
Resetting door is responsible for overtake other vehicles information and the memory before that combination newly inputs, update door be responsible for determining to leave how much previous memory,
Determine the network intermediate result for data of overtaking other vehicles with great probability output.Specific step is as follows:
ft=sigm (Wf[ht-1, xt]+bf)
it=sigm (Wi[ht-1, xt]+bi)
Wherein, i, f are that the parameter in GRU network structure has each served as the effect for forgeing data of overtaking other vehicles with more new historical,
The history data spy that overtakes other vehicles refers to the data of overtaking other vehicles of a period, and a feature of Recursive Networks is exactly different moments
Network establishes connection, excavates the network associate characteristic between different moments, and then effectively eliminate hash, utmostly
Retain useful information.It can by network training because there may be some invalid datas or noises in raw network data
So that these parameters have the characteristics that identify effective information, reject garbage, wherein the training process of network be one not
It is open close to cross weighting, map so that the process that parameter is constantly adjusted by the change of objective function, wherein W is right in a network
The weighting parameters for data of overtaking other vehicles, b are bias.Play the role of interim control parameter, i.e.,Be based on current input and
Preceding state calculates, and is obtained by the loss function training of network.
Specifically, i is responsible for determining that the importance of new data intermediate result of overtaking other vehicles (new memory) has much,
Value is bigger, indicates more important, the size of value is obtained by network training, if ltIt is approximately equal to 0, new would not be passed to
memory.T-1 is responsible for transmitting how many h determinedt-1To ht.If ftIt is 1, by ht-1All pass to ht.On the contrary, being equal to 0, new
Memory is directly passed to ht.Addition is all introduced in the update from t to t-1.Sigmoid and tanh function value difference
For [0,1] and [- 1,1], they are used for overtake other vehicles data and the intermediate result progress Nonlinear Mapping after being weighted, defeated
Output layer is arrived out carry out network training with undated parameter, i.e. network actual parameter is trained by data weighting of overtaking other vehicles, mapping
It arrives.Above all of parameter item is obtained by network by objective function training, error back propagation, the parameter that training obtains
Effective information can utmostly be retained, so that the unlimited approaching to reality value of predicted value.
Data of the invention is the rate data of overtaking other vehicles of Xinghu Street, Suzhou Industrial Park-modern times big road segment segment, rate of overtaking other vehicles
Data are as shown in table 1.The section on 20 days April -2017 years on the 1st April in 2017 is selected to overtake other vehicles rate data as training set, with
The section on April 29, -2017 years on the 21st April in 2017 is overtaken other vehicles test set of the rate data as model, for verifying having for model
Effect property and practicability, as shown in Figure 6.Fig. 7 shows the scatter plot of optional bimestrial data of overtaking other vehicles, it can be seen that two
There is no correlations for number of overtaking other vehicles between month.Neural network is declined along the opposite direction of gradient in the training process, finally
One group of parameter is obtained, predicted value approaching to reality value can be made, normalization can make model decline searching optimal solution in gradient
Process become flat, it is easier to converge to optimal solution.Therefore, all data have carried out normalized.
In formula: X indicates data of overtaking other vehicles (referring to the rate of overtaking other vehicles) of the present invention;XminIndicate the minimum inputted in data of overtaking other vehicles
Value;XmaxIndicate the maximum value overtaken other vehicles in data;X ' expression overtake other vehicles data be normalized after data.Data of overtaking other vehicles are to pass through
What electronic police obtained, normalized effect: 1. in the training method of gradient decline, for accelerating convergence;2. in order to eliminate
The influence of dimension, such as the value is too large, and network is difficult to train.The input of depth recurrent neural network is the rate of overtaking other vehicles, and output is super
Vehicle rate, this two rate essence of overtaking other vehicles are the mappings of function to function, are mapped by nonlinear function sigmoid etc..So nothing
By whether normalizing and not affecting the numerical value meaning and property of the rate of overtaking other vehicles.
1 section 43-3 of table overtakes other vehicles data
Model-evaluation index
It overtakes other vehicles the estimated performance of rate prediction model to preferably assess section, the present invention uses mean square error (Mean-
Sqaure Error, MSE), two kinds of objective functions of average absolute value error (Mean Absolute Deviation, MAD).
In formula: p (i) is that section is overtaken other vehicles the true value of rate;It overtakes other vehicles the predicted value of rate for section;N is prediction verifying number
According to collection number.
The training process of neural network model
400 groups of data are acquired by electronic police, 300 groups are used to model, and 100 groups are verified.It is of the present invention
GRU network has 200 hidden layers, an input layer, an output layer.Compared with traditional recurrent neural network, Mei Geyin
Hiding layer unit is all made of GRU, and each unit has 3 doors, and input gate indicates whether that the data information of overtaking other vehicles that sampling is obtained adds
Enter, indicates to allow to be added if it is 1, if it is indicating to be added into if 0, if fallen between 0 to 1, only part number
According to that can pass through, filters out invalid data in this way, forget door and indicate whether to remove the data of overtaking other vehicles currently retained
Historical information, 0 indicate completely abandon, 1 is fully retained, and falls between 0 to 1, then it represents that only partial data is allowed to retain.
The out gate information that indicates whether will to overtake other vehicles is output to next hidden layer or output layer, and 0 to 1 number is taken to measure output
Information of overtaking other vehicles number, 0 indicate do not export, 1 indicate completely output.The learning rate of network iteration is 0.001.Using Adam mono-
Rank optimization algorithm substitutes traditional stochastic gradient descent process, network iterative process as shown in Figure 2, be deep neural network into
Change figure, wherein ordinate is error, and abscissa is frequency of training.
Experimental result and analysis:
Trained network model is analyzed as follows: (1) to three kinds of networks (GRU, RNN, BP) respectively using square
Error (Mean-Sqaure Error, MSE), average absolute value error (Mean Absolute Deviation, MAD) are used as mesh
Scalar functions carry out network repetitive exercise, wherein average absolute value error is used to be significantly better than as the generalization ability of objective function
Network of the mean square error as objective function, as shown in figure 3, effect picture of the Grey System Model on test set.Wherein,
(a), (b) figure is training effect of the GRU Web vector graphic MAD and MSE as objective function respectively, (c), (d) figure be RNN respectively
Training effect of the Web vector graphic MAD and MSE as objective function, (e), (f) figure be BP Web vector graphic MAD and MSE conduct respectively
The training effect of objective function.This is because Some vehicles stop on a large scale caused by " vacation is overtaken other vehicles, and (vacation is overtaken other vehicles as more than stopping
Vehicle) " phenomenon, cause model to be caused to disturb there are more noise in data larger, cause network that cannot learn to potential
Rule, and then the generalization ability of network is caused to be greatly reduced.If using mean square error as objective function, can to make an uproar
The residual error of sound is further magnified, and when error back propagation and gradient decline, will deviate from being correctly oriented advance, predicted value
It cannot be approached to true value, cause large error.And average absolute value error will not make the residual error of noise as objective function
Excessively amplification, avoids the generation of this phenomenon to a certain extent.
As shown in table 2, for using absolute value error as in the model of objective function, GRU network performance is better than
RNN network performance, RNN network performance are better than BP network performance, and the absolute value error of GRU network is 0.1252, RNN network
Absolute value error is the absolute value error 0.142 of 0.128, BP network, the reason for this is that GRU has associative memory function as a kind of
Great effect can be played in network evolution with the depth Recursive Networks of Forgetting Mechanism, can network sufficiently be learnt
To useful information, the influence of garbage is eliminated.Without the BP neural network of function of associate memory, because not accounting for
Relationship on time dimension, it is larger so as to cause error.For traditional RNN network, although there is function of associate memory to consider
Relationship of the data of overtaking other vehicles on time dimension, but due to there is no Forgetting Mechanism, have no idea to eliminate data of overtaking other vehicles in vain, make
Network is obtained to be difficult to restrain.
The comparison of 2 network performance of table
With the continuous quickening of urbanization process, the cut-in situation occurred in urban road is serious to affect traffic order
Sequence has been broken the dynamic equilibrium in city road network, especially illegal passing and has been brought and great security risk.Due to traffic system
System is a kind of higher-dimension, complicated dynamical system, and the prediction for stream of overtaking other vehicles is a problem, and how to design suitable model is
The most important thing of forecasting traffic flow problem.The present invention uses deep learning method, proposes one kind based on GRU model to city
Road is overtaken other vehicles the method for rate prediction, and is compared and is analyzed with models such as traditional RNN network, BP networks.The result shows that
GRU model has better generalization ability, can be used for urban road and overtakes other vehicles the prediction of rate, for solving traffic congestion, road is gathered around
It blocks up, reduce traffic accident and provide assistant decision making support for traffic police and hold important in inhibiting.Pass through two groups of target letters
Number carries out network training and finds when verifying, and depth recurrent neural network can more excavate valuable in information flow of overtaking other vehicles obtain
Information, the framework of " door " can preferably reject useless information, retain useful information.Although traditional RNN model
There is certain memory function, but the influence of irrelevant information cannot be eliminated, BP network is easy to make due to the intensification of the network number of plies
The phenomenon that disappearing at gradient, thus perform poor in forecast of urban traffic, while in BP network, historical information
Between can not merge, and then cannot abundant mined information.To sum up, depth Recursive Networks have significant in terms of stream prediction of overtaking other vehicles
Effect, hence it is evident that be better than other models.
Although specifically showing and describing the present invention in conjunction with preferred embodiment, those skilled in the art should
Understand, is not departing from the spirit and scope of the present invention defined by the appended claims, it in the form and details can be with
The present invention is made a variety of changes, is protection scope of the present invention.
Claims (6)
- The prediction technique 1. a kind of urban road based on depth recurrent neural network is overtaken other vehicles, it is characterised in that: include the following steps,S1, data acquisition, acquisition pass through the information of vehicles and temporal information at crossing, judge whether to surpass according to collected data Vehicle;S2, number of overtaking other vehicles is counted, counts the data of overtaking other vehicles of certain a road section certain time period;S3, data prediction have carried out normalized to all data, using following formula:Wherein: X indicates data of overtaking other vehicles, XminIndicate the minimum value inputted in data of overtaking other vehicles, XmaxIndicate the maximum overtaken other vehicles in data Value, X ' expression overtake other vehicles data be normalized after data;S4, building depth recurrent neural networks model will be overtaken other vehicles data using history data prediction future of overtaking other vehicles.
- The prediction technique 2. urban road as described in claim 1 based on depth recurrent neural network is overtaken other vehicles, it is characterised in that: The depth recurrent neural networks model of the step S4 is GRU neural network, and the input of GRU neural network is that history is overtaken other vehicles number According to the output of network is the data of overtaking other vehicles of future time instance, and GRU neural network is by memory unit come more new data, a memory Unit is made of update door and resetting door, and to control the flowing of intermediate result, specific step is as follows,ft=sigm (Wf[ht-1, xt]+bf)it=sigm (Wi[ht-1, xt]+bi)Wherein, i, f are that the parameter in GRU neural network structure has each served as the effect for forgeing data of overtaking other vehicles with more new historical, W It is to the weighting parameters for data of overtaking other vehicles in GRU neural network, b is bias,Play the role of interim control parameter, i.e.,To be calculated based on current input and preceding state, obtained by the loss function training of GRU neural network.
- The prediction technique 3. urban road as described in claim 1 based on depth recurrent neural network is overtaken other vehicles, it is characterised in that: The electronic police at crossing will record the license plate number and timestamp t of vehicle in the step S1, and formerly rear crossing A and B is by electronics The vehicle of police's capture, to match vehicle, vehicle 1 and vehicle 2, if t1b< t2bAnd t1a> t2a.Then think vehicle 1 in section AB has been more than vehicle 2.
- The prediction technique 4. urban road as claimed in claim 3 based on depth recurrent neural network is overtaken other vehicles, it is characterised in that: The t2b-t2a> 15 minutes, then it is assumed that the intermediate temporary parking of vehicle 2, vehicle 1 are not to overtake other vehicles.
- The prediction technique 5. urban road as described in claim 1 based on depth recurrent neural network is overtaken other vehicles, it is characterised in that: Data of overtaking other vehicles in the step S2 are that section is overtaken other vehicles rate, and the section rate of overtaking other vehicles is that the vehicle number that section is overtaken other vehicles accounts for section and seen Survey the ratio of vehicle number:
- The prediction technique 6. urban road as described in claim 1 based on depth recurrent neural network is overtaken other vehicles, it is characterised in that: Data prediction removes police car, ambulance, military vehicle and foreign nationality's vehicle in the step S3.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811351157.3A CN109345832B (en) | 2018-11-13 | 2018-11-13 | Urban road overtaking prediction method based on deep recurrent neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811351157.3A CN109345832B (en) | 2018-11-13 | 2018-11-13 | Urban road overtaking prediction method based on deep recurrent neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109345832A true CN109345832A (en) | 2019-02-15 |
CN109345832B CN109345832B (en) | 2021-06-29 |
Family
ID=65315368
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811351157.3A Active CN109345832B (en) | 2018-11-13 | 2018-11-13 | Urban road overtaking prediction method based on deep recurrent neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109345832B (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109920248A (en) * | 2019-03-05 | 2019-06-21 | 南通大学 | A kind of public transport arrival time prediction technique based on GRU neural network |
CN110880172A (en) * | 2019-11-12 | 2020-03-13 | 中山大学 | Video face tampering detection method and system based on cyclic convolution neural network |
CN110969848A (en) * | 2019-11-26 | 2020-04-07 | 武汉理工大学 | Automatic driving overtaking decision method based on reinforcement learning under opposite double lanes |
CN111598054A (en) * | 2020-06-19 | 2020-08-28 | 上海眼控科技股份有限公司 | Vehicle detection method and device, computer equipment and storage medium |
CN111815947A (en) * | 2020-05-27 | 2020-10-23 | 中汽数据有限公司 | Method for establishing lane change time model for natural driving vehicles on expressway |
CN113537555A (en) * | 2021-06-03 | 2021-10-22 | 太原理工大学 | Traffic sub-region model prediction sliding mode boundary control method considering disturbance |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102136190A (en) * | 2011-05-03 | 2011-07-27 | 上海理工大学 | Dispatching management system and method for event emergency response of urban bus passenger transport |
EP2610836A1 (en) * | 2011-12-30 | 2013-07-03 | Seat, S.A. | Device and method for the on-line prediction of the driving cycle in an automotive vehicle |
CN105389980A (en) * | 2015-11-09 | 2016-03-09 | 上海交通大学 | Short-time traffic flow prediction method based on long-time and short-time memory recurrent neural network |
CN106971624A (en) * | 2017-04-25 | 2017-07-21 | 东软集团股份有限公司 | Overtake other vehicles the method and apparatus of early warning |
-
2018
- 2018-11-13 CN CN201811351157.3A patent/CN109345832B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102136190A (en) * | 2011-05-03 | 2011-07-27 | 上海理工大学 | Dispatching management system and method for event emergency response of urban bus passenger transport |
EP2610836A1 (en) * | 2011-12-30 | 2013-07-03 | Seat, S.A. | Device and method for the on-line prediction of the driving cycle in an automotive vehicle |
CN105389980A (en) * | 2015-11-09 | 2016-03-09 | 上海交通大学 | Short-time traffic flow prediction method based on long-time and short-time memory recurrent neural network |
CN106971624A (en) * | 2017-04-25 | 2017-07-21 | 东软集团股份有限公司 | Overtake other vehicles the method and apparatus of early warning |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109920248A (en) * | 2019-03-05 | 2019-06-21 | 南通大学 | A kind of public transport arrival time prediction technique based on GRU neural network |
CN110880172A (en) * | 2019-11-12 | 2020-03-13 | 中山大学 | Video face tampering detection method and system based on cyclic convolution neural network |
CN110969848A (en) * | 2019-11-26 | 2020-04-07 | 武汉理工大学 | Automatic driving overtaking decision method based on reinforcement learning under opposite double lanes |
CN110969848B (en) * | 2019-11-26 | 2022-06-17 | 武汉理工大学 | Automatic driving overtaking decision method based on reinforcement learning under opposite double lanes |
CN111815947A (en) * | 2020-05-27 | 2020-10-23 | 中汽数据有限公司 | Method for establishing lane change time model for natural driving vehicles on expressway |
CN111598054A (en) * | 2020-06-19 | 2020-08-28 | 上海眼控科技股份有限公司 | Vehicle detection method and device, computer equipment and storage medium |
CN113537555A (en) * | 2021-06-03 | 2021-10-22 | 太原理工大学 | Traffic sub-region model prediction sliding mode boundary control method considering disturbance |
Also Published As
Publication number | Publication date |
---|---|
CN109345832B (en) | 2021-06-29 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109345832A (en) | A kind of urban road based on depth recurrent neural network is overtaken other vehicles prediction technique | |
Gu et al. | Utilizing UAV video data for in-depth analysis of drivers’ crash risk at interchange merging areas | |
CN111768618B (en) | Traffic jam state propagation prediction and early warning system and method based on city portrait | |
Sukhadia et al. | Optimization of smart traffic governance system using artificial intelligence | |
CN110164128A (en) | A kind of City-level intelligent transportation analogue system | |
CN111462484A (en) | Congestion state determination method, device, equipment and computer readable storage medium | |
CN109523786A (en) | A kind of traffic accidents genetic analysis method | |
CN106652483A (en) | Method for arranging traffic information detection points in local highway network by utilizing detection device | |
CN107437339A (en) | Variable information advices plate control method for coordinating and system under a kind of information guidance | |
CN109887283A (en) | A kind of congestion in road prediction technique, system and device based on bayonet data | |
CN112767695A (en) | Real-time prediction method and system for traffic accident risk at signalized intersection | |
CN112967493A (en) | Neural network-based prediction method for vehicle passing intersection travel time | |
CN112906945A (en) | Traffic flow prediction method, system and computer readable storage medium | |
Zhu et al. | Spatio-temporal point processes with attention for traffic congestion event modeling | |
Long Cheu et al. | Mobile sensor and sample-based algorithm for freeway incident detection | |
Abdi et al. | A review of travel and arrival-time prediction methods on road networks: classification, challenges and opportunities | |
TW201822168A (en) | Vehicle moving direction predicting system and method using digital image recognition in combination with moving trace computation technology with application of big data computation technology | |
Zahid et al. | Predicting red light running violation using machine learning classifiers | |
Lum et al. | A study of stopping propensity at matured red light camera T-intersections | |
Li | High-resolution data-based methods for arterial traffic volume estimation | |
Gutierrez-Osorio et al. | Characterizing road accidents in urban areas of Bogota (Colombia): A data science approach | |
CN112613225B (en) | Intersection traffic state prediction method based on neural network cell transmission model | |
Gündogan et al. | Pattern recognition method for simplified coordinated traffic signal control | |
CN117273964B (en) | Intelligent vehicle insurance policy generation system and method for self-adaptive driving data | |
Lan et al. | A rolling-trained fuzzy neural network approach for freeway incident detection |
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 |