A traffic situation analysis system based on deep learning

Abstract

The present invention relates to the field of intelligent transportation, and in particular to a traffic situation analysis system based on deep learning. The present invention constructs the situation analysis module and the simulation module to estimate and predict the traffic status of the entire road network in real time, and explore the traffic operation rules of specific road sections. Through big data analysis technology and machine learning algorithms, key traffic indicators are extracted to improve prediction accuracy. At the same time, the simulation module is used to simulate traffic control measures to provide users with a scientific basis, improve traffic conditions, improve road traffic efficiency, and ensure traffic safety. At the same time, it provides technical support for traffic management and decision-making, and promotes the sustainable development of the transportation industry.

Description

Technical Field

[0001] The present invention relates to the field of intelligent transportation, and in particular to a traffic situation analysis system based on deep learning. Background Art

[0002] With the continuous development of my country's economy and society, the mileage of roads open to traffic and the number of motor vehicles have continued to increase. With the development of big data, the database of traffic big data constructed for various traffic scenarios has been initially improved. From the actual situation analysis, in daily life, traffic congestion in fixed time periods (morning and evening rush hours) and the number of road traffic accidents are also showing an increasing trend. The current traffic safety situation in my country is still very serious.

[0003] Traditional traffic situation analysis systems mainly rely on fixed sensor networks for data collection and processing, and have problems such as incomplete data coverage, slow update speed, and lack of real-time performance. This directly leads to insufficient real-time performance and low accuracy of past traffic flow forecasts. At the same time, there is a lack of situation analysis systems that support simulation and simulation, which leads to poor visualization and interactivity of the system, lack of support for road control, and inability to meet the needs of rapid response to emergencies. Summary of the invention

[0004] The purpose of the present invention is to overcome the problems of insufficient real-time performance, low accuracy and inability to perform decision simulation in the prior art, and to provide a traffic situation analysis system based on deep learning.

[0005] In order to achieve the above-mentioned object of the invention, the present invention provides the following technical solutions:

[0006] A traffic situation analysis system based on deep learning, comprising a real-time data acquisition module, a data storage module, a situation analysis module, a simulation module and a display module;

[0007] The real-time data acquisition module is used to obtain real-time comprehensive traffic data; the real-time comprehensive traffic data includes traffic data, road basic data and environmental data;

[0008] The data storage module is used to store historical traffic comprehensive data; the historical traffic comprehensive data includes traffic data, road basic data, environmental data and corresponding traffic situation data;

[0009] The situation analysis module is trained by the historical traffic comprehensive data and is used to predict the traffic situation analysis data of the corresponding area according to the real-time traffic comprehensive data; the traffic situation analysis data includes short-term traffic flow prediction results, long-term traffic situation prediction results and congestion prediction results;

[0010] The simulation module is used to simulate traffic flow under different conditions and different control measures;

[0011] The display module is used to output the traffic situation analysis data corresponding to the real-time traffic comprehensive data and the simulation results of the simulation module;

[0012] The situation analysis module is respectively connected to the real-time data acquisition module, the data storage module and the simulation module in communication; the display module is respectively connected to the situation analysis module and the simulation module in communication.

[0013] As a preferred solution of the present invention, the simulation module includes a simulation model and a simulation platform;

[0014] The simulation model includes: a macroscopic model and a microscopic model;

[0015] The macro model is used to describe the overall characteristics of the road network;

[0016] The microscopic model includes a car-following model and a lane-changing model, which are used to describe the behavior and interaction of a single vehicle; further, the car-following model is used to describe the motion law and behavior of a vehicle when following other vehicles; the lane-changing model simulates the lane-changing behavior of a vehicle by adjusting the lane-changing willingness parameter of the vehicle;

[0017] The simulation platform is used to simulate the road network environment, and call the situation analysis module or the pre-built simulation model to analyze the simulation scene to generate traffic flow simulation results.

[0018] As a preferred solution of the present invention, the macro model regards the traffic flow as a continuous medium and is defined by the flow conservation equation; further, the expression of the flow conservation equation is:

[0019] ,

[0020] Among them, ρ is the traffic density, q is the traffic flow, t and x represent the time variable and location space variable respectively;

[0021] Furthermore, the expression of the flow-density relationship is:

[0022] q = ρ v(ρ),

[0023] where v(ρ) is the velocity-density function;

[0024] Furthermore, the expression of the velocity-density function is:

[0025] ,

[0026] Among them, V is the maximum velocity, ρ max is the maximum density.

[0027] As a preferred solution of the present invention, the simulation platform includes the following working steps:

[0028] (1) Building a real-time simulation environment

[0029] A simulation environment for a road network is built through a geographic information system, and a physical environment model is established in the simulation environment. The constructed simulation model is then imported into a traffic simulation software. Finally, a real-time data interface is established to input sensor and GPS data into the simulation platform in real time, and the input data is processed in real time to update traffic status and environmental parameters; the traffic status includes flow, speed and / or density; the environmental parameters include weather data;

[0030] (2) Reliability analysis

[0031] Setting the simulation step size and synchronization mechanism so that the macro model and the micro model are consistent in time, and obtaining the reliability of the simulation result by calculating the root mean square error based on the detector data;

[0032] (3) Result verification

[0033] Analyze the distribution characteristics of simulation data in time and space through spatiotemporal data analysis technology, verify the spatiotemporal consistency of the results, and select a corresponding number of spatiotemporal points according to the correlation of the spatiotemporal consistency during simulation;

[0034] The spatiotemporal consistency verification includes the following steps:

[0035] ① Constructing space-time matrix

[0036] The space-time matrix is ​​composed of simulation data, and its construction includes the following process: obtaining traffic flow simulation data of different time periods and spatial positions within a certain time range, and then arranging the simulation data into a space-time matrix; wherein the rows of the space-time matrix represent time slices; the columns of the space-time matrix represent spatial positions; the matrix elements in the space-time matrix represent the traffic flow at the corresponding spatial position at the corresponding time slice;

[0037] ② Draw a spatiotemporal heat map

[0038] The spatiotemporal heat map uses traffic volume as a heat map division index to display the spatiotemporal variation characteristics of traffic flow, and the color bar from dark to light represents the traffic volume from large to small; further, the vertical axis of the spatiotemporal heat map represents different time slices; the horizontal axis represents different spatial positions;

[0039] ③ Spatial-temporal correlation analysis

[0040] The traffic status correlation between different time slices and spatial locations can be visually observed through the color changes of the correlation matrix heat map, and the correlation between two time slices or spatial locations can be quantified through the Pearson correlation coefficient;

[0041] If the correlation is higher than the preset value, a smaller number of preset points will be selected for simulation analysis during simulation;

[0042] If the correlation is lower than the preset value, a larger number of preset points will be selected for simulation analysis during simulation.

[0043] As a preferred embodiment of the present invention, the simulation platform further includes a missing trajectory reconstruction method, comprising the following steps:

[0044] Obtain observation trajectories, extract time, space and motion features from each observation trajectory, and then calculate the degree of temporal overlap between any two trajectories; determine that several observation trajectories whose temporal overlap reaches a preset threshold belong to the same vehicle;

[0045] According to the temporal and spatial position relationship between the missing trajectory and the observed trajectory, the trajectory reconstruction problem is divided into four cases, and the trajectory reconstruction is performed according to the classification of the current missing trajectory; the four cases include the missing future trajectory of the rear vehicle, the missing future trajectory of the front vehicle, the missing past trajectory of the rear vehicle, and the missing past trajectory of the front vehicle;

[0046] The reconstructed trajectory is smoothly connected with the observed trajectory through a trajectory correction method based on proportional scaling.

[0047] As a preferred embodiment of the present invention, the calculation formula for the degree of overlap in time is:

[0048]

[0049] Among them, Ot is the temporal overlap, [ta1, ta2] and [tb1, tb2] are the time ranges of the two trajectories respectively.

[0050] As a preferred solution of the present invention, trajectory reconstruction is performed according to the trajectory reconstruction problem classification to which the current missing trajectory belongs, including:

[0051] ① The future trajectory of the following vehicle is missing:

[0052] ,

[0053] in, is the position of the following car at time m+1 in the future, is the position of the preceding vehicle at time m+1 in the future; is the desired following distance, , T is the safety time head distance, L is the vehicle length; are the speeds of the rear and front vehicles at time m respectively;

[0054] ②The future trajectory of the preceding vehicle is missing:

[0055] ,

[0056] ③The past track of the following vehicle is missing:

[0057]

[0058] in, is the position of the following car at time m-1 in the past, is the position of the preceding vehicle at time m-1 in the past; are the speeds of the rear and front cars at the past m-1 moments respectively;

[0059] ④ The previous track of the preceding vehicle is missing:

[0060] .

[0061] As a preferred solution of the present invention, when the track is completely missing, the track reconstruction is achieved by a track insertion method based on the expected spacing of the IDM model, including the following steps:

[0062] The observed trajectory points before and after the missing trajectory area are set as Ppre and Ppost respectively, and the time interval between the two areas is t=Tpost-Tpre;

[0063] Evenly insert N time points within the time interval t to generate the initial interpolation time series Ti=Tpre+i×t / (N+1);

[0064] For each time point Ti, the speed and position of the vehicle are updated according to the IDM model formula;

[0065] According to the speed and position of the vehicle at each time point, the trajectory reconstruction result is output.

[0066] As a preferred embodiment of the present invention, the trajectory correction method based on proportional scaling comprises the following steps:

[0067] Calculate the scaling factor , where d i YesP i-1 and P i The distance between them, L is the preset calculation length;

[0068] Each point (x, y) in the reconstructed trajectory is scaled and corrected in equal proportion so that it gradually transitions to the starting point of the observed trajectory; further, the corrected (x i ',y i ') The coordinate point calculation formula is:

[0069] ,

[0070] Among them, the original reconstructed trajectory point (x i ,y i ) within the connection area;

[0071] The scaled area is smoothed by spline interpolation, and the coefficients are solved by solving the linear equations to obtain a smooth curve, which is expressed as follows:

[0072] ,

[0073] Among them, S i (x) is a smooth curve, a i 、b i 、c i S i The coefficient of (x) is, and the linear equations to be solved are;

[0074] ,

[0075] .

[0076] As a preferred embodiment of the present invention, the simulation platform also includes a vehicle trajectory lane division method, comprising the following steps:

[0077] (1) Determine the characteristic points of the road centerline

[0078] Analyzing the trajectory data of the vehicle by a clustering algorithm, identifying the clustering area of ​​the vehicle trajectory, and using the clustering area as the characteristic point of the road; the trajectory data includes a location and a timestamp;

[0079] (2) Correction and iteration of feature points

[0080] Calculate the local neighborhood of each feature point and count the distribution of vehicle trajectories in the neighborhood; adjust the position of the feature point to make it closer to the trajectory-dense area; iterate the above steps repeatedly until the position of the feature point converges or reaches the predetermined accuracy requirement;

[0081] (3) Extract center lines and lane lines based on kernel density estimation and divide lanes

[0082] Select kernel function and bandwidth parameters; perform kernel density estimation based on the feature points obtained in (2) and generate a density distribution map;

[0083] In the density distribution map, the path with the highest density is obtained and used as the center line of the road; and the density peak points are extracted through a path tracing algorithm and connected to generate the center line of the road;

[0084] On both sides of the road centerline, density estimation is performed in the vertical direction to obtain the path with the second highest density, which is used as the lane line; and the density peak points are extracted through the path tracking algorithm, which are connected to generate the extracted lane line;

[0085] Among them, the formula for kernel density estimation is:

[0086] ,

[0087] Where f(x) is the density estimate at point x; n is the number of data points; h is the bandwidth parameter; K is the kernel function, and x i is the horizontal coordinate of the i-th trajectory point in the vehicle trajectory data.

[0088] Compared with the prior art, the present invention has the following beneficial effects:

[0089] The present invention estimates and predicts the traffic status of the entire road network in real time by constructing the situation analysis module and the simulation module, and explores the traffic operation rules of specific road sections. Through big data analysis technology and machine learning algorithms, key traffic indicators are extracted to improve prediction accuracy. At the same time, the simulation module is used to simulate traffic control measures to provide users with a scientific basis, improve traffic conditions, improve road traffic efficiency, and ensure traffic safety. At the same time, it provides technical support for traffic management and decision-making, and promotes the sustainable development of the transportation industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] Figure 1 This is a schematic diagram of the structure of a traffic situation analysis system based on deep learning according to Embodiment 1 of the present invention;

[0091] Figure 2 This is a schematic diagram of time overlap calculation in a traffic situation analysis system based on deep learning according to Embodiment 2 of the present invention;

[0092] Figure 3 Schematic diagram A of trajectory reconstruction classification in a traffic situation analysis system based on deep learning according to Embodiment 2 of the present invention;

[0093] Figure 4 Schematic diagram B of trajectory reconstruction classification in a traffic situation analysis system based on deep learning according to Embodiment 2 of the present invention;

[0094] Figure 5 This is a schematic diagram of trajectory reconstruction when the trajectory is completely missing in a traffic situation analysis system based on deep learning according to Embodiment 2 of the present invention;

[0095] Figure 6 This is a schematic diagram of trajectory smoothing calculation in a traffic situation analysis system based on deep learning according to Embodiment 2 of the present invention;

[0096] Figure 7 This is a schematic diagram of the steps of determining the characteristic points of the road centerline in a traffic situation analysis system based on deep learning according to Embodiment 2 of the present invention;

[0097] Figure 8 This is a schematic diagram of the steps of correcting feature points in a traffic situation analysis system based on deep learning according to Embodiment 2 of the present invention;

[0098] Fig. 9 This is a schematic diagram of the steps of iterating feature points in a traffic situation analysis system based on deep learning according to Embodiment 2 of the present invention;

[0099] Fig.10 This is a schematic diagram of lane division in a traffic situation analysis system based on deep learning described in Example 2 of the present invention. DETAILED DESCRIPTION

[0100] The present invention is further described in detail below in conjunction with test examples and specific implementation methods. However, this should not be understood as the scope of the above subject matter of the present invention being limited to the following embodiments, and all technologies realized based on the content of the present invention belong to the scope of the present invention.

[0101] Example 1

[0102] like Figure 1 As shown, a traffic situation analysis system based on deep learning includes a real-time data acquisition module, a data storage module, a situation analysis module, a simulation module and a display module.

[0103] The real-time data acquisition module is used to obtain real-time comprehensive traffic data; the real-time comprehensive traffic data includes traffic data, road basic data and environmental data.

[0104] The data storage module is used to store historical traffic comprehensive data; the historical traffic comprehensive data includes traffic data, road basic data, environmental data and corresponding traffic situation data.

[0105] The situation analysis module is trained by the historical traffic comprehensive data and is used to predict the traffic situation analysis data of the corresponding area based on the real-time traffic comprehensive data; the traffic situation analysis data includes short-term traffic flow prediction results, long-term traffic situation prediction results and congestion prediction results.

[0106] The simulation module is used to simulate traffic flow under different conditions and different control measures.

[0107] The display module is used to output the traffic situation analysis data corresponding to the real-time comprehensive traffic data and the simulation results of the simulation module.

[0108] The situation analysis module is respectively connected to the real-time data acquisition module, the data storage module and the simulation module in communication; the display module is respectively connected to the situation analysis module and the simulation module in communication.

[0109] Example 2

[0110] This embodiment is a specific implementation scheme of a traffic situation analysis system based on deep learning described in Example 1. The traffic situation analysis system includes a real-time data acquisition module, a data storage module, a situation analysis module, a simulation module and a display module.

[0111] The real-time data acquisition module is used to obtain real-time comprehensive traffic data; the real-time comprehensive traffic data includes traffic data, road basic data and environmental data. The traffic data in the real-time data acquisition module is generated by obtaining the vehicle passing data of the toll station and the gantry, and the road monitoring video data.

[0112] The data storage module is used to store historical traffic comprehensive data; the historical traffic comprehensive data includes traffic data, road basic data, environmental data and corresponding traffic situation data.

[0113] The situation analysis module is trained by the historical traffic comprehensive data and is used to predict the traffic situation analysis data of the corresponding area according to the real-time traffic comprehensive data; the traffic situation analysis data includes short-term traffic flow prediction results, long-term traffic situation prediction results and congestion prediction results. The situation analysis module is the core part of this embodiment, and its goal is to achieve the estimation and prediction of traffic status in the entire road network. The module uses the collected various historical traffic comprehensive data to build an algorithm model, and then estimates and predicts the traffic status through the algorithm model. By converting the real-time traffic comprehensive data into valuable information, a scientific basis is provided for traffic management and decision-making. Specifically, the situation analysis module aims to achieve high-precision, fine-grained short-term estimation and prediction of traffic flow, and to predict, alarm, impact assessment, and effect simulation of emergencies including congestion events, and provide decision support for road safety operation during holidays. The situation analysis module includes a short-term traffic flow prediction unit, a long-term traffic situation prediction unit, and a congestion prediction unit.

[0114] The traffic flow short-term prediction unit includes the following training steps:

[0115] S1: Obtain historical traffic comprehensive data, pre-process it and generate a training sample set;

[0116] S2: constructing a short-term traffic flow prediction model, training the model through the training sample set, and outputting the short-term traffic flow prediction model after the model converges;

[0117] The traffic flow short-term prediction model includes a graph convolutional network module, an attention mechanism module and a long short-term memory network module.

[0118] The long-term traffic situation prediction unit includes the following training steps:

[0119] S1: Obtain historical traffic comprehensive data, pre-process it and generate a training sample set;

[0120] S2: constructing a long-term traffic situation prediction model, training the model through the training sample set, and outputting the long-term traffic situation prediction model after the model converges;

[0121] The traffic flow short-term prediction model includes a SARIMAX module.

[0122] The congestion prediction unit includes the following training steps:

[0123] S1: Obtain historical traffic comprehensive data, pre-process it and generate a training sample set;

[0124] S2: constructing a congestion prediction model, training the model through the training sample set, and outputting the congestion prediction model after the model converges;

[0125] The congestion prediction model includes.

[0126] Furthermore, the situation analysis module divides the road in a grid-like manner, and each grid represents a road section of a certain length. The sub-units that have been trained are used to process the real-time comprehensive traffic data in each grid, and key indicators such as vehicle flow, average vehicle speed, and road clearing status are extracted. By analyzing these indicators, the traffic status in each grid is estimated. At the same time, combined with the traffic data of the entire road network, the algorithm model is used to predict the traffic flow in the future, and the prediction content includes vehicle flow, average speed, and road clearing status.

[0127] Furthermore, this embodiment also uses data cleaning technology to remove invalid and abnormal data, uses feature extraction technology to extract key features, and uses machine learning algorithms for classification, regression and clustering to optimize the training sample set for model training. In addition, model parameter adjustment and optimization technology can also be used to further optimize the parameters of the algorithm model to improve prediction accuracy and stability.

[0128] The simulation module includes a simulation model and a simulation platform; the simulation model includes: a macro model and a micro model.

[0129] The macro model is used to describe the overall characteristics of the road network; the macro model regards the traffic flow as a continuous medium and is defined by the flow conservation equation; further, the expression of the flow conservation equation is:

[0130] ,

[0131] Among them, ρ is the traffic density, q is the traffic flow, t and x represent the time variable and location space variable respectively;

[0132] Furthermore, the expression of the flow-density relationship is:

[0133] q = ρ v(ρ),

[0134] where v(ρ) is the velocity-density function;

[0135] Furthermore, the expression of the velocity-density function is:

[0136] ,

[0137] Among them, V max is the maximum velocity, ρ max is the maximum density.

[0138] The microscopic model includes a car-following model and a lane-changing model, which are used to describe the behavior and interaction of a single vehicle.

[0139] The lane-changing model is used to simulate the lane-changing behavior of the vehicle by adjusting the lane-changing willingness parameter of the vehicle. Based on the driver's behavior rules and traffic environment, the safe distance between vehicles, speed differences, road markings and other factors are considered, and the lane-changing behavior of the vehicle is simulated by adjusting the lane-changing willingness parameter of the vehicle. This embodiment realizes the control of vehicle lane changing by adjusting the following parameters:

[0140] (1) Strategic Lane-Change Desire

[0141] Cooperative lane-changing willingness refers to the degree to which a vehicle tends to adjust its lane-changing behavior based on the behavior of surrounding vehicles. When surrounding vehicles exhibit cooperative lane-changing behavior, the vehicle may increase its willingness to change lanes in order to better adapt to traffic flow dynamics. In SUMO, cooperative lane-changing willingness is usually configured through the parameter lcCooperative. This parameter indicates the degree to which a vehicle tends to perform cooperative lane-changing, that is, the probability that the vehicle tends to change lanes based on the behavior of surrounding vehicles.

[0142] (2) Maximum Speed ​​Difference Lane-Change Desire

[0143] The maximum speed lane change intention refers to the vehicle's tendency to change lanes based on the speed difference between the current lane and the target lane. In SUMO, the maximum speed lane change intention can be configured through the parameter lcSpeedGain. This parameter represents the relationship between the vehicle's willingness to change lanes and the speed difference between the current lane and the target lane.

[0144] (3) Minimum Lane-Change Gap

[0145] The minimum lane change gap refers to the minimum distance requirement between the vehicle and the vehicle in front and behind when changing lanes. When the vehicle changes lanes, this minimum distance requirement must be met. In SUMO, the minimum lane change gap can be configured through the parameter lcMinGap.

[0146] (4) Lane-Change Delay

[0147] Lane change delay refers to the response delay of the vehicle to change lanes, that is, the time interval between the vehicle's decision to change lanes and the actual execution of the lane change operation. In SUMO, the lane change delay can be configured through the parameters lcKeepRight and lcKeepLeft, which represent the lane change delay when the vehicle keeps the right and left lanes respectively.

[0148] The following model is used to describe the movement and behavior of a vehicle when following other vehicles. Based on factors such as vehicle speed, acceleration, and distance between vehicles, the following behavior between vehicles is simulated by calibrating following parameters such as vehicle headway and minimum following interval, that is, how the vehicle adjusts its speed according to the speed and distance of the vehicle in front. In this embodiment, the IDM and Gipps models are mainly used. The vehicle in the IDM model will adjust its own acceleration according to the difference between its desired speed, safe distance, and the speed of the vehicle in front. In the free flow state, the vehicle tends to travel at the desired speed; in the following state, the vehicle will maintain a safe distance and accelerate or decelerate according to the speed and distance of the vehicle in front. Compared with the IDM, the Gipps model is more realistic and takes into account more factors in real road conditions, such as the driver's reaction time, acceleration limits, etc. Specifically:

[0149] IDM Model

[0150] 1) Model Assumptions

[0151] ① Acceleration is a strictly decreasing function of speed. In addition, if there is no restriction from other vehicles, the vehicle will accelerate to the desired speed v0; acceleration is an increasing function of the distance s from the vehicle in front; acceleration decreases when approaching the vehicle in front; the minimum distance (bumper-to-bumper distance) s0 must be maintained between the vehicle in front, but if the distance is already less than s0, the vehicle will not move backwards.

[0152] ② The distance in the equilibrium state is not less than the safety distance s0+vT, where s0 is the minimum distance and T is the time interval.

[0153] ③ Under normal circumstances, the braking action is "soft". The deceleration gradually increases to a comfortable value b (comfortable value), and when a stable car following is reached, it will smoothly decrease to 0; in dangerous conditions, the deceleration exceeds the comfortable value, and the remaining braking will remain at b.

[0154] ④The transition between different driving modes is smooth.

[0155] ⑤The model should be as simple as possible. Each model parameter should only describe one aspect of driving behavior.

[0156] 2) Three driving states

[0157] ① Accelerate from rest on a free road. The vehicle starts at the highest acceleration a. The higher the speed, the lower the acceleration. When the speed gradually approaches the desired speed v0, the acceleration gradually approaches 0. The larger the value of δ, the slower the acceleration decreases as it approaches v0.

[0158] ② Follow a car in front, and the distance between you and the car in front is roughly the safe distance s0+vT.

[0159] ③When approaching a very slow or stationary vehicle, the deceleration usually does not exceed the safe deceleration b.

[0160] 3) IDM model formula

[0161] ,

[0162] in, is the acceleration of the vehicle; a is the maximum acceleration of the vehicle; is the desired speed of the vehicle, and the common value on highways is 120 km / h; δ is the acceleration index, and the common value on highways is 4; v ɑ is the current speed of the vehicle; s * is the expected following distance; ∆v ɑ is the speed difference between the vehicle and the preceding vehicle; s ɑis the distance between the vehicle and the vehicle in front. The acceleration equation mainly includes two parts: It is used to measure the difference between the current vehicle speed and the expected vehicle speed, and promote vehicle acceleration. It is used to measure the difference between the current vehicle distance and the expected vehicle distance and promote vehicle braking.

[0163] Gipps Model

[0164] 1) Safe speed model

[0165] ① Model assumptions

[0166] a) Braking is always performed with a continuous deceleration b, without any distinction between comfort and (physically possible) maximum deceleration.

[0167] b) There is a constant reaction time ∆t.

[0168] c) Even if the leading vehicle suddenly decelerates to a complete stop (worst case), the distance between the following vehicle and the leading vehicle should not be less than the minimum interval s0.

[0169] ②Derivation of safe speed formula

[0170] a) The braking distance of the vehicle ahead is , which describes the distance traveled by the leading vehicle in the worst case when it decelerates to a stop at a constant deceleration, where v1 is the speed of the leading vehicle.

[0171] b) Stopping distance of the following vehicle , where v∆t is the reaction distance, which describes the distance the following vehicle travels from the moment the driver realizes that he needs to slow down until he stops, and v is the speed of the following vehicle.

[0172] c) Interval formula: conditions that s should satisfy

[0173] ,

[0174] d) Safe speed

[0175] ,

[0176] ③Characteristics reflected by the model:

[0177] a) Each step interval of the model is ∆t

[0178] b) If the speed of the following vehicle is greater than v safe -a∆t or v0-a∆t is small, then in the next step, the vehicle speed will directly reach min(v0,v safe ).

[0179] c) In other cases, the following vehicle will use constant acceleration until it reaches v safe or v0.

[0180] 2) Steady state equilibrium (uniform and stable state):

[0181] v(t+∆t)=v1=v,

[0182] ,

[0183] Furthermore, the macro model and micro model use the same database to update and share data in real time, and realize bidirectional data transmission through a standardized API interface. Define the location and boundary conditions of the macro area and the micro area, convert the output (flow, density) of the macro model into the input (vehicle position and speed) of the micro model, and use the decomposition method to decompose the macro flow into individual vehicles and assign them to specific lanes and positions in the micro model; the micro model generates detailed vehicle trajectories and operation data based on the behavior of specific vehicles and feeds them back to the macro model for updating the traffic flow characteristics of the macro model.

[0184] The simulation platform is used to simulate the road network environment, and call the situation analysis module or the pre-built simulation model to analyze the simulation scene to generate traffic flow simulation results. The simulation platform includes the following working steps:

[0185] (1) Building a real-time simulation environment

[0186] A simulation environment of a road network is built through a geographic information system (GIS), and a physical environment model (including traffic signs, weather, etc.) is established in the simulation environment. The constructed simulation model is then imported into the traffic simulation software. Finally, a real-time data interface is established to input sensor and GPS data into the simulation platform in real time, and the input data is processed in real time to update traffic status and environmental parameters; the traffic status includes flow, speed and / or density; the environmental parameters include weather data;

[0187] (2) Reliability analysis

[0188] Setting the simulation step size and synchronization mechanism so that the macro model and the micro model are consistent in time, and obtaining the reliability of the simulation result by calculating the root mean square error based on the detector data;

[0189] (3) Result verification

[0190] The distribution characteristics of simulation data in time and space are analyzed by spatiotemporal data analysis technology to verify the spatiotemporal consistency of the results; the following steps are included:

[0191] ④ Constructing the space-time matrix

[0192] The space-time matrix is ​​composed of simulation data. The specific construction process is: obtain traffic simulation data within a certain time range, traffic flow information in different time periods and spatial locations, and then organize the simulation data into a space-time matrix; wherein the rows of the space-time matrix represent time slices, which are divided into morning peak period, evening peak period and flat peak period; the columns of the space-time matrix represent spatial locations, which are divided into ramp entrances and exits and hub interchanges. The matrix elements in the space-time matrix represent the traffic flow at the corresponding spatial location at the corresponding time slice;

[0193] ⑤ Draw a spatiotemporal heat map

[0194] The spatiotemporal heat map uses traffic volume as a heat map division index to display the spatiotemporal variation characteristics of traffic flow, and the color bar from dark to light indicates that the traffic volume is from large to small; further, the vertical axis of the spatiotemporal heat map represents different time slices, indicating the variation trend of traffic status in different time periods; the horizontal axis represents different spatial positions, indicating the variation characteristics of traffic status at different positions;

[0195] ⑥ Spatial-temporal correlation analysis

[0196] The traffic status correlation between different time slices and spatial locations can be visually observed through the color changes of the correlation matrix heat map, and the correlation between two time slices or spatial locations can be quantified through the Pearson correlation coefficient;

[0197] If the correlation is higher than the preset value, it means that the traffic state change trends of the two time slices or spatial locations are similar. When performing simulation, a preset smaller number of points are selected for simulation analysis, that is, representative points are selected for data output, indicating the traffic operation characteristics of all time-space points with high correlation;

[0198] If the correlation is lower than the preset value, it means that their changing trends are different. When performing simulation, a preset larger number of points are taken for simulation analysis, that is, more points need to be refined for result output to accurately describe the changes in traffic status;

[0199] Furthermore, given two time slices or spatial location traffic status data sequences X and Y, the calculation formula of the Pearson correlation coefficient r is:

[0200]

[0201] Where X i and Y i Respectively represent the traffic state value (flow) at the i-th time point or spatial position; and are the means of the sequences X and Y respectively; n is the number of sampling points at different time and space positions.

[0202] Identifying anomalies and consistency `High correlation means that the traffic status change trends of two time slices or spatial locations are similar, while low correlation means that their change trends are different.

[0203] Furthermore, the simulation platform also includes a missing trajectory reconstruction method, comprising the following steps:

[0204] (1) Obtain observation trajectories, extract temporal, spatial, and motion features from each observation trajectory, and then calculate the temporal overlap between any two trajectories; several observation trajectories whose temporal overlap reaches a preset threshold are determined to belong to the same vehicle.

[0205] like Figure 2 As shown in the figure, an improved car-following model is used for trajectory association, trajectory restoration is performed based on the spatiotemporal position relationship between the missing trajectory and the observed trajectory, and the vehicle ID is used to determine whether different trajectories belong to the same vehicle.

[0206] Furthermore, the calculation formula for the degree of temporal overlap is:

[0207] ,

[0208] Among them, Ot is the temporal overlap, [ta1, ta2] and [tb1, tb2] are the time ranges of the two trajectories respectively.

[0209] (2) According to the spatiotemporal relationship between the missing trajectory and the observed trajectory, the trajectory reconstruction problem is divided into four cases, and the trajectory is reconstructed according to the classification of the current missing trajectory; the four cases include the missing future trajectory of the following vehicle, the missing future trajectory of the leading vehicle, the missing past trajectory of the following vehicle, and the missing past trajectory of the leading vehicle;

[0210] According to the trajectory reconstruction problem classification to which the current missing trajectory belongs, the trajectory is reconstructed. Further, represents the position of the nth vehicle (the following vehicle) at time m, Indicates the position of the n-1th vehicle (the front vehicle) at time m+1 (in the future), represents the position of the nth vehicle at time m-1 (in the past), such as Figure 3 and Figure 4 As shown in Figure 1, the trajectory reconstruction problem is divided into four cases: the future trajectory of the rear vehicle is missing, the future trajectory of the front vehicle is missing, the past trajectory of the rear vehicle is missing, and the past trajectory of the front vehicle is missing.

[0211] use Figure 2 The method is used to reconstruct the trajectories of these four situations, including:

[0212] ① The future trajectory of the following vehicle is missing:

[0213] ,

[0214] in, is the position of the following car at time m+1 in the future, is the position of the preceding vehicle at time m+1 in the future; is the desired following distance, which is usually a function of speed and safety time head distance. , T is the safety time head distance, L is the vehicle length; are the speeds of the rear and front vehicles at time m respectively;

[0215] ②The future trajectory of the preceding vehicle is missing:

[0216] ,

[0217] ③The past track of the following vehicle is missing:

[0218]

[0219] in, is the position of the following car at time m-1 in the past, is the position of the preceding vehicle at time m-1 in the past; are the speeds of the rear and front cars at the past m-1 moments respectively;

[0220] ④ The previous track of the preceding vehicle is missing:

[0221] .

[0222] Furthermore, when the trajectory is completely missing, such as Figure 5 As shown, trajectory reconstruction is achieved by a trajectory insertion method based on the expected spacing of the IDM model, including the following steps:

[0223] The observed trajectory points before and after the missing trajectory area are set as Ppre and Ppost respectively, and the time interval between the two areas is t=Tpost-Tpre;

[0224] Evenly insert N time points within the time interval t to generate the initial interpolation time series Ti=Tpre+i×t / (N+1);

[0225] For each time point Ti, the speed and position of the vehicle are updated according to the IDM model formula;

[0226] According to the speed and position of the vehicle at each time point, the trajectory reconstruction result is output.

[0227] (3) Through the trajectory correction method based on proportional scaling, the reconstructed trajectory and the observed trajectory are smoothly connected. Figure 6 As shown, for the reconstructed trajectory points Scale proportionally to , to solve the problem of non-smooth connection between the reconstructed trajectory and the detected vehicle trajectory. Further, the trajectory correction method based on proportional scaling includes the following steps:

[0228] Calculate the scaling factor , where d i YesP i-1 and P i The distance between them, L is the preset calculation length;

[0229] Each point (x, y) in the reconstructed trajectory is scaled and corrected in equal proportion so that it gradually transitions to the starting point of the observed trajectory; further, the corrected (x i ',y i ') The coordinate point calculation formula is:

[0230] ,

[0231] Among them, the original reconstructed trajectory point (x i ,y i ) within the connection area;

[0232] The scaled area is smoothed by spline interpolation, and the coefficients are solved by solving the linear equations to obtain a smooth curve, which is expressed as follows:

[0233]

[0234] Among them, S i (x) is a smooth curve, a i 、b i 、c i S i The coefficient of (x).

[0235] Furthermore, in order to determine each interval [x i ,x i+1 S in ] i (x) The following conditions must be met:

[0236] Interpolation condition: S i (x)=y i ; S i+1 (x)=y i+1 ;

[0237] Continuity condition: S i ′(x+1)=S i+1 ′(x+1); S i ′′(x+1)=S i+1 ′′(x+1).

[0238] Therefore, for each i=0,1,…,n-1, the linear equations to be solved are:

[0239] Interpolation conditions:

[0240]

[0241] Continuity conditions:

[0242] .

[0243] Through the above linear equations, S can be solved i The coefficient a of (x) i 、b i 、c i .

[0244] Furthermore, the simulation platform also includes a vehicle trajectory lane division method, comprising the following steps:

[0245] (1) Determine the characteristic points of the road centerline

[0246] The vehicle trajectory data is analyzed by clustering algorithms (such as K-means, DBSCAN), the clustered areas of the vehicle trajectories are identified, and the clustered areas are used as feature points of the road; the trajectory data includes location (latitude and longitude or plane coordinates) and timestamp.

[0247] like Figure 7 As shown, according to the vehicle trajectory point distribution map, densely distributed trajectory points are found as weight points, and the characteristic points of the road centerline are determined according to the distribution positions of the weight points.

[0248] (2) Correction and iteration of feature points

[0249] Calculate the local neighborhood of each feature point and count the distribution of vehicle trajectories in the neighborhood; adjust the position of the feature point to make it closer to the trajectory-dense area; iterate the above steps repeatedly until the position of the feature point converges or reaches the predetermined accuracy requirement.

[0250] Further, such as Figure 8 As shown in FIG, the road feature points are further corrected based on the extracted positions of the two feature points T0 and T1 being parallel to the lane (perpendicular to the vertical line). Fig. 9 As shown, according to the position distribution of the vehicle trajectory points and the feature points T, multiple iterations are performed to obtain the divided lane lines LC, LL, RL, and RC.

[0251] (3) Extract center lines and lane lines based on kernel density estimation and divide lanes

[0252] Select a kernel function (such as a Gaussian kernel) and bandwidth parameters; perform kernel density estimation based on the feature points obtained in (2) to generate a density distribution map;

[0253] In the density distribution map, the path with the highest density is obtained and used as the center line of the road; and the density peak points are extracted through a path tracing algorithm and connected to generate the center line of the road;

[0254] On both sides of the road centerline, density estimation is performed in the vertical direction to obtain the path with the second highest density, which is used as the lane line; and the density peak points are extracted through a path tracing algorithm (such as the shortest path algorithm), and then these points are connected in order to generate and extract clear and accurate lane lines;

[0255] That is Fig.10 As shown in the figure, the extracted lane lines are used to construct geometric polygons. The two sides of each lane line can be regarded as the boundaries of the polygon. The boundaries on both sides are connected to form closed areas, which are the geometric polygons of the lane. Once these polygons are constructed, the trajectory of the vehicle can be matched with these polygons based on its position information. Each trajectory point (the location information of the vehicle at different times) will be assigned to the polygon corresponding to its position, so as to determine which lane the vehicle is traveling in during a specific time period.

[0256] Among them, the formula for kernel density estimation is:

[0257] ,

[0258] Where f(x) is the density estimate at point x; n is the number of data points; h is the bandwidth parameter; K is the kernel function, and x i is the horizontal coordinate of the i-th trajectory point in the vehicle trajectory data.

[0259] The display module is used to output the traffic situation analysis data corresponding to the real-time comprehensive traffic data and the simulation results of the simulation module.

[0260] The simulation module described in this embodiment constructs a simulated road network and traffic environment by importing basic road data, vehicle types, traffic lights and other data. Users can set different traffic conditions and parameters in the simulation module, such as vehicle density, average speed, road conditions, etc., to simulate traffic flows under different circumstances. Through simulation operation, users can observe and analyze the operation of traffic flows on the simulated road section, including vehicle driving trajectories, speed changes, congestion conditions, etc.

[0261] The simulation module also supports users to set and simulate traffic control measures. For example, users can set control schemes such as dynamic speed limit, emergency lane opening, entrance and exit flow control, and observe the impact of these measures on traffic flow. By comparing the effects of different control schemes, users can select the best scheme to implement in order to improve traffic conditions and road traffic efficiency.

[0262] In addition, the simulation module also provides a wealth of data analysis and visualization tools. Users can analyze and display simulation results in the form of charts and reports provided by the system to better understand the traffic operation rules and control effects. These data can provide scientific basis and technical support for decision makers, helping them to better formulate and manage traffic strategies.

[0263] Traffic control technology is to develop a traffic flow control strategy by exploring the changing trend of road section operation. This project intends to adopt a method that combines traffic state estimation with model predictive control technology to propose a control strategy based on the historical operation law data of the road section and the future traffic development trend. Specifically, the control measures include toll station entrance and exit flow control strategy, fixed-point flow control strategy, road section speed limit strategy, and lane dynamic expansion strategy.

[0264] The toll station entrance and exit flow limiting strategy utilizes the traffic flow information (such as flow, speed, occupancy, etc.) of the toll station entrance and exit ramps and downstream lanes to implement the toll station entrance and exit ramp control strategy to alleviate traffic congestion problems; that is, flow limiting control is performed on the toll station entrance and exit ramps.

[0265] The fixed-point flow limiting strategy is a real-time fixed-point flow limiting strategy for the upstream and downstream of the congested section according to the traffic flow operation status of the simulated road section (such as controlling the upstream vehicles of the congested section to slow down), so as to alleviate the congestion problem of the expressway section; that is, controlling the upstream vehicles of the congested section to slow down or controlling the downstream vehicles of the congested section to speed up.

[0266] The road section speed limit strategy is to adopt a speed limit control strategy for the traffic flow of a specific congested section according to the traffic simulation operation status of the expressway section to ensure the driving safety of the section; that is, to adopt a speed limit control strategy for the traffic flow of a specific congested section.

[0267] The lane dynamic expansion strategy is to formulate open use permissions for emergency lanes according to the time and space scope based on the traffic status information of the expressway section, so as to improve its operating efficiency and service level and reduce the impact of emergency vehicles on traffic flow; that is, to formulate open use permissions for emergency lanes on specific sections according to the time and space scope.

[0268] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A traffic situation analysis system based on deep learning, characterized in that: It includes real-time data acquisition module, data storage module, situation analysis module, simulation module and display module; The real-time data acquisition module is used to obtain real-time comprehensive traffic data; the real-time comprehensive traffic data includes traffic data, road basic data and environmental data; The data storage module is used to store historical traffic comprehensive data; the historical traffic comprehensive data includes traffic data, road basic data, environmental data and corresponding traffic situation data; The situation analysis module is trained by the historical traffic comprehensive data and is used to predict the traffic situation analysis data of the corresponding area according to the real-time traffic comprehensive data; the traffic situation analysis data includes short-term traffic flow prediction results, long-term traffic situation prediction results and congestion prediction results; The simulation module is used to simulate traffic flow under different conditions and different control measures; The display module is used to output the traffic situation analysis data corresponding to the real-time traffic comprehensive data and the simulation results of the simulation module; The situation analysis module is respectively connected to the real-time data acquisition module, the data storage module and the simulation module; the display module is respectively connected to the situation analysis module and the simulation module; The simulation module includes a simulation model and a simulation platform; The simulation model includes: a macroscopic model and a microscopic model; The macro model is used to describe the overall characteristics of the road network; The microscopic model includes a car-following model and a lane-changing model, which are used to describe the behavior and interaction of a single vehicle; the car-following model is used to describe the motion law and behavior of a vehicle when following other vehicles; the lane-changing model simulates the lane-changing behavior of a vehicle by adjusting the lane-changing willingness parameter of the vehicle; The simulation platform is used to simulate the road network environment, and call the situation analysis module or the pre-built simulation model to analyze the simulation scene to generate traffic flow simulation results; The simulation platform also includes a missing trajectory reconstruction method, comprising the following steps: Obtain observation trajectories, extract time, space and motion features from each observation trajectory, and then calculate the degree of temporal overlap between any two trajectories; determine that several observation trajectories whose temporal overlap reaches a preset threshold belong to the same vehicle; According to the temporal and spatial position relationship between the missing trajectory and the observed trajectory, the trajectory reconstruction problem is divided into four cases, and the trajectory reconstruction is performed according to the classification of the current missing trajectory; the four cases include the missing future trajectory of the rear vehicle, the missing future trajectory of the front vehicle, the missing past trajectory of the rear vehicle, and the missing past trajectory of the front vehicle; The reconstructed trajectory is smoothly connected with the observed trajectory through a trajectory correction method based on proportional scaling. Among them, the calculation formula for the degree of overlap in time is: ， Among them, Ot is the temporal overlap, [ta1, ta2] and [tb1, tb2] are the time ranges of the two trajectories respectively.

2. A traffic situation analysis system based on deep learning according to claim 1, characterized in that: The macro model regards the traffic flow as a continuous medium and defines it through the flow conservation equation; the expression of the flow conservation equation is: ， Among them, ρ is the traffic density, q is the traffic flow, t and x represent the time variable and location space variable respectively; The expression of flow-density relationship is: q = ρ v(ρ), where v(ρ) is the velocity-density function; The expression of the velocity-density function is: ， Among them, V max is the maximum velocity, ρ max is the maximum density.

3. A traffic situation analysis system based on deep learning according to claim 1, characterized in that: The simulation platform includes the following working steps: (1) Building a real-time simulation environment A simulation environment for a road network is built through a geographic information system, and a physical environment model is established in the simulation environment. The constructed simulation model is then imported into a traffic simulation software. Finally, a real-time data interface is established to input sensor and GPS data into the simulation platform in real time, and the input data is processed in real time to update traffic status and environmental parameters; the traffic status includes flow, speed and / or density; the environmental parameters include weather data; (2) Reliability analysis Setting the simulation step size and synchronization mechanism so that the macro model and the micro model are consistent in time, and obtaining the reliability of the simulation result by calculating the root mean square error based on the detector data; (3) Result verification Analyze the distribution characteristics of simulation data in time and space through spatiotemporal data analysis technology, verify the spatiotemporal consistency of the results, and select a corresponding number of spatiotemporal points according to the correlation of the spatiotemporal consistency during simulation; The spatiotemporal consistency verification includes the following steps: Constructing the space-time matrix The space-time matrix is ​​composed of simulation data, and its construction includes the following process: obtaining traffic flow simulation data of different time periods and spatial positions within a certain time range, and then arranging the simulation data into a space-time matrix; wherein the rows of the space-time matrix represent time slices; the columns of the space-time matrix represent spatial positions; the matrix elements in the space-time matrix represent the traffic flow at the corresponding spatial position at the corresponding time slice; Plotting spatiotemporal heatmaps The spatiotemporal heat map uses traffic volume as a heat map division index to display the spatiotemporal variation characteristics of traffic flow. The color bar from dark to light represents the traffic volume from large to small. The vertical axis of the spatiotemporal heat map represents different time slices; the horizontal axis represents different spatial positions. Spatial-temporal correlation analysis The traffic status correlation between different time slices and spatial locations can be visually observed through the color changes of the correlation matrix heat map, and the correlation between two time slices or spatial locations can be quantified through the Pearson correlation coefficient; If the correlation is higher than the preset value, a smaller number of preset points will be selected for simulation analysis during simulation; If the correlation is lower than the preset value, a larger number of preset points will be selected for simulation analysis during simulation.

4. The traffic situation analysis system based on deep learning according to claim 1, characterized in that: Trajectory reconstruction is performed according to the trajectory reconstruction problem classification to which the current missing trajectory belongs, including: ① The future trajectory of the following vehicle is missing: ， in, is the position of the following car at time m+1 in the future, is the position of the preceding vehicle at time m+1 in the future; is the desired following distance, , T is the safety time head distance, L is the vehicle length; are the speeds of the rear and front vehicles at time m respectively; ②The future trajectory of the preceding vehicle is missing: ， ③The past track of the following vehicle is missing: in, is the position of the following car at time m-1 in the past, is the position of the preceding vehicle at time m-1 in the past; are the speeds of the rear and front cars at the past m-1 moments respectively; ④ The previous track of the preceding vehicle is missing: 。 5. A traffic situation analysis system based on deep learning according to claim 4, characterized in that: When the trajectory is completely missing, the trajectory reconstruction is achieved through the trajectory insertion method based on the expected spacing of the IDM model, including the following steps: The observed trajectory points before and after the missing trajectory area are set as Ppre and Ppost respectively, and the time interval between the two areas is t=Tpost-Tpre; Evenly insert N time points within the time interval t to generate the initial interpolation time series Ti=Tpre+i×t / (N+1); For each time point Ti, the speed and position of the vehicle are updated according to the IDM model formula; According to the speed and position of the vehicle at each time point, the trajectory reconstruction result is output.

6. A traffic situation analysis system based on deep learning according to claim 5, characterized in that: The trajectory correction method based on proportional scaling comprises the following steps: Calculate the scaling factor , where d i YesP i-1 and P i The distance between them, L is the preset calculation length; Each point (x, y) in the reconstructed trajectory is scaled and corrected in equal proportion so that it gradually transitions to the starting point of the observed trajectory; the corrected (x i ',y i ') The coordinate point calculation formula is: ， Among them, the original reconstructed trajectory point (x i ,y i ) within the connection area; The scaled area is smoothed by spline interpolation, and the coefficients are solved by solving the linear equations to obtain a smooth curve, which is expressed as follows: ， Among them, S i (x) is a smooth curve, a i 、b i 、c i S i The coefficient of (x) is, and the linear equations to be solved are; ， 。 7. The traffic situation analysis system based on deep learning according to claim 1, characterized in that: The simulation platform also includes a vehicle trajectory lane division method, comprising the following steps: (1) Determine the characteristic points of the road centerline Analyzing the trajectory data of the vehicle by a clustering algorithm, identifying the clustering area of ​​the vehicle trajectory, and using the clustering area as the characteristic point of the road; the trajectory data includes a location and a timestamp; (2) Correction and iteration of feature points Calculate the local neighborhood of each feature point and count the distribution of vehicle trajectories in the neighborhood; adjust the position of the feature point to make it closer to the trajectory-dense area; iterate the above steps repeatedly until the position of the feature point converges or reaches the predetermined accuracy requirement; (3) Extract center lines and lane lines based on kernel density estimation and divide lanes Select kernel function and bandwidth parameters; perform kernel density estimation based on the feature points obtained in (2) and generate a density distribution map; In the density distribution map, the path with the highest density is obtained and used as the center line of the road; and the density peak points are extracted through a path tracing algorithm and connected to generate the center line of the road; On both sides of the road centerline, density estimation is performed in the vertical direction to obtain the path with the second highest density, which is used as the lane line; and the density peak points are extracted through the path tracking algorithm, which are connected to generate the extracted lane line; Among them, the formula for kernel density estimation is: ， Where f(x) is the density estimate at point x; n is the number of data points; h is the bandwidth parameter; K is the kernel function, and x i is the horizontal coordinate of the i-th trajectory point in the vehicle trajectory data.

Images