Real-time estimation method for traffic flow of urban road network without detector based on transfer learning

Abstract

The application discloses a kind of based on transfer learning's city road network detectorless section flow real-time estimation method, belong to intelligent transportation system field.The method of the application includes city local road network similarity evaluation step and model transfer learning step, city local road network similarity evaluation step is by to the multi-dimension road network characteristics of influence road section flow quantitative analysis and weight assignment, establishes similarity evaluation system, to find the detector road network similar to the detectorless road section road network, model transfer learning step with the detectorless road section road network needing to be estimated as transfer learning target field, with the detector road network found by similarity evaluation system as source field, using fine-tuning method carries out the transfer learning of flow estimation model, to complete the flow real-time estimation of no data road section;The application fully excavates the influence law of city road network to flow, realizes high-precision flow estimation of no data road section with a brand-new data-driven thought.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intelligent transportation systems, and particularly relates to a city road network detector-free road section flow real-time estimation method based on transfer learning. BACKGROUND

[0002] In China, the continuous growth of urban vehicle number leads to road congestion becoming the traffic norm, and mastering urban traffic travel information and formulating reasonable traffic control measures is an effective way to solve traffic congestion. Traffic flow is a kind of travel information that needs to actually count the number of vehicles running on the road network and can reflect the traffic state in real time, and its acquisition is crucial. However, only part of the roads in the city are equipped with fixed detectors that can obtain traffic flow data, and the flow data of other roads without detector installation cannot be obtained. At present, the research on traffic state estimation mainly focuses on the point-level flow loss scene caused by detector network interruption, power shortage, and the flow segment-level loss scene caused by software or detector hardware failure under adverse weather, and the research on the slice-level data loss scene caused by the absence of detectors is very lacking. At present, the research on the slice-level data loss scene caused by the absence of detectors can be divided into statistical methods and machine learning methods.

[0003] In the research of real-time estimation of non-detector road section traffic based on statistical methods, the general method is to extract the space-time characteristics from the road sections with detectors, and then use the model established by the similar road sections to interpolate the missing data. Aslam in the paper "In City-scale traffic estimation from a roving sensor network" selects a group of road sections according to the characteristics of road facilities to evaluate the correlation, and then uses these road sections to establish a multiple linear regression (MLR) model to estimate the undetected amount. Similarly, based on the MLR model, Liu in the paper "Mining road network correlation for traffic estimation via compressive sensing" reveals the hidden patterns in the city network and measures the road section correlation to estimate the missing traffic state. In the research of real-time estimation of non-detector road section traffic based on machine learning methods, Zhan in the paper "Citywide traffic volume estimation using trajectory data" proposes a Bayesian network estimation model, which fuses taxi global positioning system (GPS) data and fixed detector data to construct a feature map, thereby realizing the estimation of city-scale traffic volume. After this research, Meng in the paper "City-wide traffic volume inference with loop detector data and taxi trajectories" continues to develop a semi-supervised learning-based model to estimate these non-detector road sections. Liu in the paper "Think like a graph: real-time traffic estimation at city-scale" studies a graph theory-based parallel computing model combined with a dynamic neural network to study the missing state estimation.

[0004] However, in the above examples, the research method based on statistical methods only considers the similarity of the attributes of the road itself when analyzing the road similarity, without considering the influence of the surrounding road sections and road connection entities on the road traffic, resulting in poor data interpolation effect and lack of persuasiveness. The research model based on machine learning is complex and has a large amount of data, which cannot achieve real-time and accurate estimation of road section traffic, so the above methods are not effective for real-time estimation of traffic of non-detector road sections in urban road network. SUMMARY

[0005] In order to overcome the deficiencies of the prior art, the present application provides a kind of real-time estimation method for urban road network detectorless section flow based on transfer learning, the object in the present application is all urban detectorless sections, and the traffic flow state of its section can be effectively estimated in real time.

[0006] The technical solutions specifically adopted by the present application are as follows:

[0007] A kind of real-time estimation method for urban road network detectorless section flow based on transfer learning, which comprises:

[0008] S1, for target urban road network, crossroads are taken as division points to divide roads into several sections and mark serial numbers, road sections are taken as nodes, and the direction of the connection between road sections is taken as the direction of edges, so as to abstract the urban road network into an urban road network graph; then, based on the urban road network graph, the local road network corresponding to each road section node in the graph is constructed, and the local road network of each road section node is a road network with all nodes and edge connection relationships within two hops of the center node as the center node;

[0009] S2, urban local road network similarity evaluation: taking the road network topology index, road network functional area attribute index, road network distance index and road network node attribute index as multi-dimensional evaluation indexes for comprehensively reflecting the influence of road network on road traffic flow, establishing an urban local road network similarity evaluation system through quantitative analysis and weight assignment of the multi-dimensional evaluation indexes; then, the similarity between the local road network of the detectorless section to be estimated and each local road network of the detector section is calculated using the urban local road network similarity evaluation system, so as to find the local road network of the detector section with the most similar traffic distribution to the local road network of the detectorless section to be estimated;

[0010] S3, model transfer learning: taking the local road network of the detectorless section to be estimated as the target field of transfer learning, and taking the local road network of the detector section found by the urban local road network similarity evaluation system as the source field of transfer learning, using the fine-tuning method to perform transfer learning on the traffic estimation model pre-trained on the source field, so as to complete real-time estimation of the traffic of the detectorless section by using the traffic estimation model obtained by transfer learning.

[0011] It should be noted that in the present application, the local road network with detectors in the center section is defined as the local road network of the detector section, and the local road network without detectors in the center section is defined as the local road network of the detectorless section.

[0012] As a preferred embodiment, the specific steps of S2 include:

[0013] S21, establishing a city local road network similarity evaluation system: using multi-source city traffic data to construct a road network topology structure index, a road network functional area attribute index, a road network distance index and a road network node attribute index, four indexes comprehensively reflecting the influence of the road network on the road traffic flow from multiple angles, then using objective assignment method to determine the weight of each index, thereby constructing a city local road network similarity evaluation system for weighting and summing the similarity of each evaluation index between local road networks to obtain the similarity score between local road networks;

[0014] S22, target road network road network similarity evaluation: taking the local road network corresponding to the detectorless road segment to be estimated as the target local road network, taking the local road network corresponding to other detector-equipped road segments in the city road network map as the candidate source field set, using the city local road network similarity evaluation system constructed in S21 to comprehensively evaluate the similarity between each detector-equipped road segment local road network in the candidate set and the target local road network, finding the detector-equipped road segment local road network in the candidate source field set with the highest similarity score with the target local road network, i.e. the most similar road segment flow distribution, as the source local road network.

[0015] As a preferred, the specific steps of S3 include:

[0016] S31, taking the local road network of the detectorless road segment to be estimated as the transfer learning target field taking the detector-equipped road segment local road network found in S2 with the most similar flow distribution to the detectorless road segment local road network as the transfer learning source field

[0017] S32, using GE-GAN based on graph embedding and adversarial generation network as the flow estimation model, inputting the graph structure relationship corresponding to the local road network of the detector-equipped road segment in the source field and the historical flow data of all nodes into the GE-GAN model for model training;

[0018] S33, using the fine-tuning method in transfer learning, using the few-sample artificial observation data of the detectorless road segment local road network in the target field to perform secondary training on the GE-GAN model trained in S32, so that the model parameters are more adapted to the target field, thereby realizing real-time flow estimation of the detectorless road segment local road network.

[0019] As a preferred, the road network topology structure index includes two indexes of global topology structure and center node topology structure, and the calculation method of the global topology structure similarity and the center node topology structure similarity between two local road networks is:

[0020] ​S31, statistics of each local road network total node number and total edge number two sub-index, then with two sub-index weighted sum as global topology structure index value, and two sub-index weighted weight is obtained by using objective weight method, weight sum is 1;

[0021] S32, statistics of each local road network center node degree centrality, betweenness centrality, closeness centrality, PageRank centrality and eigenvector centrality five sub-index, then with five sub-index weighted sum as center node topology structure index value, and five sub-index weighted weight is obtained by using objective weight method, weight sum is 1;

[0022] S33, for two local road network G jt And G ia , calculate the global topology structure similarity between them, the calculation formula is:

[0023] S f =1-|K′ fjt -K′ fia |

[0024] In the formula: K′ fjt Is the normalization result of global topology structure index value of local road network G jt , K′ fia Is the normalization result of global topology structure index value of local road network G ia ;

[0025] For two local road network G jt And G ia , calculate the center node topology structure similarity between them, the calculation formula is:

[0026] S n =1-|K njt -K nia |

[0027] In the formula: K njt Is the center node topology structure index value of local road network G jt , K nia Is the center node topology structure index value of local road network G ia .

[0028] As preferred, the road network functional area attribute index is characterized by the distribution of POI near the road network center section, the calculation method of road network functional area attribute index similarity between two local road network G jt And G ia :

[0029]

[0030] In the formula, n is the total number of functional area types; Kpjtm and K piam are respectively G jt The functional area attribute index vector K pjt and G ia The functional area attribute index vector K pia The mth value of G jt and G ia The weighted number of the mth functional area in G

[0031] For any local road network r, the weighted number of the dth functional area is calculated by the following formula:

[0032] K pjtm = TF_IDF rd *n dr

[0033] In the formula, n dr represents the number of POIs in the dth functional area in the local road network r; TF_IDF rd represents the corresponding weighted weight of n dr , and the calculation formula is:

[0034]

[0035] In the formula, ∑n r is the total number of functional areas in the local road network r, ∑n p is the total number of local road networks, and ∑n pd is the total number of local road networks containing the dth functional area.

[0036] As a preferred embodiment, the functional area types include teaching areas, residential areas, tourist areas, commercial areas, medical areas, and working areas, and the corresponding POI types are: middle schools and primary schools for teaching areas, residential areas for residential areas, scenic spots for tourist areas, comprehensive shopping malls and commercial pedestrian streets for commercial areas, and industrial parks and business buildings for working areas.

[0037] As a preferred embodiment, when counting the number of functional areas of the local road network, the counting range is a rectangular area near the center road segment, specifically including a rectangular area on the right side of the center road segment and a rectangular area on the right side of the road segment leading to the center road segment. When counting the number of functional areas, the rectangular range search function of the map API is used to determine whether the POI point is located on the right side of the road segment. When determining the longitude and latitude of the diagonal points of the rectangular area, the longitude and latitude (lon1, lat1) of one point A is the starting point of the road segment, and the longitude and latitude (lon3, lat3) of the other point C are calculated by the following formula:

[0038] lat3 = lat2 - dis * sin (Azi * π / 180) / (2π * ARC / 360);

[0039] lon3=lon2+dis*cos(Azi*π / 180) / (2π*ARC*cos(lat3*π / 180));

[0040] In the formula, lat2 and lon2 are the latitude and longitude of the end point B of the road segment, dis is the width of the rectangle, Azi is the azimuth angle of the line connecting the start and end points of the rectangle relative to due north, and ARC is the length of the Earth's equatorial radius.

[0041] Preferably, the road network distance index is the spatial distance between the central road segment of the target domain road network and the central road segment of the candidate source domain road network. The formula for calculating the similarity of the road network distance index between two local road networks is as follows:

[0042] S d =e -2.3x

[0043] In the formula, x is the spatial distance between two local road networks.

[0044] Preferably, the road network node attribute index is the road design capacity of the central road segment of the road network, and the formula for calculating the similarity of road network node attribute indices between two local road networks is:

[0045] S c =1-|W s -W t |

[0046] In the formula, W s and W t These represent two local road networks G. jt and G ia The normalized result of road capacity.

[0047] Preferably, the indicators of the urban local road network similarity evaluation system are determined using the CRITIC weighting method, with a total weight of 1. The formula for calculating the total similarity score between the local road network in the target domain and the local road network in the candidate source domain is as follows:

[0048] S = W f *S f +W n *S n +W p *S p +W d *S d +W c *S c

[0049] In the formula, S f S n S p S d and Sc are the similarity of global topology index, central node topology index, road network functional area attribute index, road network distance index and road network node attribute index respectively, W f , W n , W p , W d and W c are the weights of global topology index, central node topology index, road network functional area attribute index, road network distance index and road network node attribute index respectively.

[0050] Compared with the prior art, the beneficial effects of the present application are as follows:

[0051] The present application quantitatively analyzes and assigns weights to the multi-dimensional road network characteristics that affect road section flow, establishes a similarity evaluation system to find a detector-equipped road network similar to the flow distribution of a non-detector-equipped road network, and then uses the fine-tuning method to perform transfer learning of the flow estimation model with the non-detector-equipped road network as the target learning domain and the detector-equipped road network found by the similarity evaluation system as the source domain, thereby completing real-time flow estimation of the non-detector-equipped road section. The present application fully explores the influence of urban road networks on flow and achieves high-precision flow estimation of non-data road sections with a completely new data-driven approach. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 is the overall flowchart of the method of the present application.

[0053] Figure 2 is an abstracted schematic diagram of the Huzhou road network.

[0054] Figure 3 is a schematic diagram of calculating the latitude and longitude of the diagonal points when selecting the POI rectangular range.

[0055] Figure 4 is the overall framework of the flow estimation model GE-GAN.

[0056] Figure 5 is a schematic diagram of the transfer learning of the flow estimation model. DETAILED DESCRIPTION

[0057] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application are described in detail below. In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the concept of the present application, so the present application is not limited to the specific embodiments disclosed below. The technical features in each embodiment of the present application can be combined accordingly without conflict.

[0058] In a preferred embodiment of the present application, a real-time detector-free road section flow estimation method based on transfer learning for urban road networks is provided, which comprises the following steps:

[0059] 1) Select an urban road network according to the estimation requirements, remove part of the small-flow secondary trunk roads and branch roads according to the detector distribution of the road network and the connectivity level of the urban roads, and obtain the processed urban road network and the road connection relationship.

[0060] 2) According to the distribution of the crossroads in the processed urban road network, divide the roads into a plurality of road sections and label the serial numbers, take the road sections as nodes and the connection direction of the road sections as edge direction, abstract the urban road network into a dual graph, and use mathematical symbol G=(V,E) to represent, G is the abstracted network, V is the abstracted node set of the network individual.

[0061] 3) Define the local road network of a node as the road network composed of all nodes and edge connection relationships within the two-hop range of the node, define the node set with detectors as V i , define the node set without detectors as V j , define the local road network set of the node with detectors as G i , and define the local road network set of the node without detectors as G j .

[0062] 4) Construct a city local road network similarity evaluation system, which includes four indexes reflecting the influence of road network on road section flow, i.e. road network topology, road network functional area attribute, road network distance and road network node attribute. The index values are set in the range of [0, 1] or (0, 1] through normalization, the index weights are calculated through CRITIC weight method, and then the similarity between two road network indexes is calculated. Here, G jt and G ia are the evaluation objects, and the specific steps are as follows:

[0063] 4-1) Construct a road network topology index to reflect the similarity of the influence of road network on road section flow in connectivity and centrality, which is respectively represented as global topology Kf and center node topology V n Two sub-indices, global topology K f including total node number K node , total edge number K edge Two sub-indices, center node topology including local road network center node degree centrality K dc , betweenness centrality K bc , closeness centrality K cc , PageRank centrality K pr and eigenvector centrality K ec Five sub-indices, global topology index K f The calculation formula is:

[0064] K f = W node *K node + W edge *K edge

[0065] W node + W edge = 1

[0066] In the formula, W node and W edge are the weights of the corresponding sub-indices, the weight sum is 1, and the total node number K node , total edge number K edge of all road networks are substituted into the objective weight method to obtain;

[0067] The calculation formula of K n is:

[0068] K n = W dc *K dc + W bc *K bc + W cc *K cc + W pr *K pr + W ec *K ec

[0069] W dc + W bc + W cc + W pr + W ec = 1

[0070] In the formula, W dc , W bc , W cc , W pr and W ecrespectively, the weights and are 1, and are obtained by substituting the centrality values of all road networks into the objective weight method.

[0071] 4-2), calculating G jt and G ia The similarity on the road network topology structure index, the similarity calculation formula of the global topology structure index is:

[0072]

[0073] S f = 1 - |K fjt -K fia |

[0074] In the formula, K f ' is the normalized global topology structure value, min(K f ) and max(K f ) are the minimum value and maximum value of the global topology structure in all road networks, S f is the similarity of G jt and G ia on the global topology structure index, K' fjt is the global topology structure value of G jt after normalization, and K' fia is the global topology structure value of G ia after normalization.

[0075] It should be noted that the two types of sub-indices for the global topology structure index described above need to be normalized before the similarity score is obtained. The normalization method can adopt the maximum and minimum value normalization, and the specific calculation formula is as follows:

[0076]

[0077] In the formula, x is the value of the index, min(x) is the minimum value of the index in the total road network, and max(x) is the maximum value of the index in the total road network.

[0078] The similarity calculation formula of the central node topology structure index is:

[0079] S n = 1 - |K njt -K nia |

[0080] In the formula, S n is the similarity of G jt and G ia on the central node topology structure index, K njt is the central node topology structure value of g jt , and K niais G ia the center node topology value of G

[0081] 4-3), build road network functional area attribute index, first count the POI type and quantity corresponding to the functional area in the road network to clarify the road network functional area attribute, then use TF-IDF algorithm to determine the weight of each functional area in each road network, and then use cosine similarity to calculate G jt and G ia the similarity of the functional area attribute between the road networks.

[0082] The above functional area categories are teaching area, residential area, tourist area, commercial area, medical area and working area, and the POI types corresponding to the functional areas are as follows: middle school and primary school for teaching area, residential area for residential area, scenic spot for tourist area, comprehensive mall and commercial pedestrian street for commercial area, and industrial park and commercial building for working area.

[0083] The POI selection range of the local road network functional area is defined as the rectangular area near the center road segment, which specifically includes the rectangular area on the right side of the center road segment and the rectangular area on the right side of the road segment leading to the center road segment (i.e. the road segment downstream in the direction of the center road segment), and the latitude and longitude calculation formula of the diagonal node C of the path endpoint B of the rectangular area with A as the starting point is

[0084] lat3=lat2-dis*sin(Azi*π / 180) / (2π*ARC / 360);

[0085] lon3=lon2+dis*cos(Azi*π / 180) / (2Azi*cos(lat3*π / 180));

[0086] wherein lat2 and lon2 are the latitude and longitude of the road segment endpoint B, dis is the width of the rectangular range, Azi is the azimuth angle of the rectangular starting point and endpoint, the azimuth angle of the rectangular starting point and endpoint, and ARC is the length of the Earth equatorial radius.

[0087] The formula for determining the weight by TF-IDF algorithm is:

[0088] TF_IDF rd =TF rd *IDF d

[0089]

[0090] wherein TF_IDF rd is the weight value of the dth functional area in any local road network r, TF rd measures the frequency proportion of the dth functional area in the local road network r, n drLet ∑n be the number of functional zones of type d in the local road network r (based on the number of POIs belonging to that functional zone type). r IDF is the sum of functional zones in the local road network r. d The frequency of the d-th functional zone in all road networks is measured by ∑n. p For the total number of roads, ∑n pd This represents the sum of all road networks that include the d-th functional area.

[0091] For any local road network r, the general formula for calculating the weighted quantity of the d-th functional area is:

[0092] K pjtm =TF_IDF rd *n dr

[0093] The weighted quantity K of all n functional areas mentioned above pjtm Arranged in order, they can form a vector of functional area attribute indicators with a dimension of n.

[0094] 4-4) Calculate road network G jt and G ia The similarity of road network functional area attribute indicators is calculated using the following formula:

[0095]

[0096] In the formula, S p For road network G jt and G ia The similarity in attribute indicators of road network functional areas, with a value range of [0,1], K pjt and K pia G jt and G ia The functional area attribute index vector, where n is the sum of functional area types, which is 6 in this case, and m is the m-th type of functional area.

[0097] 4-5) Construct a road network distance index, which is the road network distance index G. jt and G ia The spatial distance between central nodes, rather than the actual travel distance, is used. The spatial correlation of road network traffic decreases exponentially with distance. The correlation between roads with a distance of 1 km decreases to 0.1. Therefore, the formula for calculating the similarity of road network distance indicators is:

[0098] S d =e -2.3x

[0099] In the formula, S dThe distance index is the similarity between road networks, with a value range of (0,1]. x is the distance between the center nodes of the road network (in meters), and the coordinates of the center node of the road are taken as the midpoint of the coordinates of the start and end points of the road.

[0100] 4-6) Construct road network node attribute indicators, specifically reflecting the road design capacity of the central road segment. This reflects the upper limit of traffic flow that the road network node can handle. In the calculation of the design capacity of a road segment, the design capacity of one lane is generally used as the basis. The specific calculation formula is as follows:

[0101] N a =N o *γ*η*C*n′

[0102] In the formula, N a N represents the design capacity of the road (pcu / h). o Let γ be the design capacity of a lane (pcu / h), η be the correction factor for bicycle travel, η be the correction factor for lane width, C be the correction factor for intersection, and n′ be the correction factor for the number of lanes. In this invention, only the design capacity and number of lanes of a single lane are considered, and the other correction factors are not considered for the time being and are assumed to be 1.

[0103] Normalize the index using the formula in 4-2), and the road network G jt and G ia The formula for calculating the similarity of road network node attributes is:

[0104] S c =1-|K′ cjt -K′ cia |

[0105] In the formula, S c For the similarity of road network node attributes, K′ cjt The local road network G of the section to be estimated jt Normalized design capacity of the central node road, K′ cia Local road network G with detectors ia Normalized design traffic capacity of central node roads.

[0106] 4-7) Use the objective weighting method to determine the weights of the above-mentioned construction indicators. The sum of the weights is 1. Road network G jt and G ia The formula for calculating the total similarity is:

[0107] S = W f *S f +W n *S n +W p *S p +W d *Sd +W c *S c

[0108] In the formula, S is the total similarity score of the road network, W f , W n , W p , W d and W c are the weights of the road network topology index (global topology index, central node topology index), road network functional area attribute index, road network distance index and road network node attribute index, respectively, S f , S n , S p , S d and S c are the similarities of the indexes.

[0109] 5) Find the most similar local road network with detector road sections to the local road network of the to-be-estimated road section according to the local road network similarity evaluation system constructed in 4), train the traffic estimation model in the local road network with detector road sections, use the knowledge of transfer learning to fine-tune the trained model, and then apply it to the local road network of the to-be-estimated road section to obtain real-time traffic estimation data of the to-be-estimated road section. The specific steps are as follows:

[0110] 5-1) Take the local road network of the to-be-estimated road section without detectors as the target field of transfer learning Take the local road network of the road section with detectors as the candidate set Use the local road network similarity evaluation system constructed in S4 to score the similarity of all local road networks in to the target field road network, and select the highest scoring road network as the source field of transfer learning

[0111] 5-2) Introduce a GE-GAN model based on graph embedding and adversarial generative network as the traffic estimation model. The graph embedding step makes the models trained for road networks with different numbers of nodes have transferability. The graph structure relationship and all node historical traffic data of the source field road network are input into the model for training.

[0112] 5-3) Fine-tune the model trained in 5-2) using the few-sample artificial observation data of the detectorless road section in the target field (which can be collected by artificial observation and only a small amount of samples are needed for fine-tuning), so that the model parameters are more adapted to the target field, thereby realizing real-time traffic estimation of the detectorless road section.

[0113] The estimation method described in the above embodiments will be applied to a real road network scenario. Taking the urban road network of Huzhou as the research object, the present invention will be explained in more detail in combination with the real road network scenario, and the advantages and implementation details of the present invention will become clearer.

[0114] Example

[0115] like Figure 1 The diagram shown illustrates the overall process of this invention. The real-time traffic estimation method for detector-free road sections in urban road networks based on transfer learning specifically includes the following steps:

[0116] Step 1: Select the Huzhou urban road detector dataset and perform data preprocessing on the urban road traffic dataset, including anomaly detection and missing data filling for both short-term and long-term anomalies.

[0117] Step 2: Based on the detector distribution and urban road connectivity levels given in the dataset, remove some low-traffic secondary arterial roads and branch roads from the real road network to obtain the processed urban road network and road connectivity.

[0118] Step 3: Divide the roads into several segments using the intersections in the city road network map and label them with serial numbers. Using road segments as nodes and the connection directions between road segments as edge directions, abstract the city road network into a city road network map. For example... Figure 2 As shown, the urban road network map is represented by the mathematical symbol G=(V,E), where G is the abstracted network, V is the set of nodes after abstracting individual network members, denoted as V={V1,V2,…,Vn}, and the number of network nodes n=|V|. E represents the set of edges after abstracting the interactions between individual network members, denoted as E={e1,e2,…,em}, and the number of network edges m=|E|.

[0119] Step 4: Based on the presence or absence of detectors in road segments recorded in the urban road network dataset, classify road network nodes into two categories: nodes with detectors and nodes without detectors. The set of nodes with detectors is denoted as V. i The set of nodes without detectors is denoted as V. j .

[0120] Step 5: Define a local road network as a sub-network consisting of all nodes and edge connections within a node's two-hop range, centered on that node. The set of local road networks centered on nodes with detectors is denoted as G. i The set of local road networks centered on nodes without detectors is denoted as G. j .

[0121] Step 6: Construct a local urban road network similarity evaluation system to evaluate the similarity of nodes V to be estimated. jt The local road network G centered on jt With G iThe similarity of all local road networks is calculated, and the local road network G jt The local road network G is The process of constructing the local road network similarity evaluation system includes four links of index construction, index weight determination, and similarity calculation, and the specific sub-steps of the construction process are steps 61 to 67, which are described as follows:

[0122] Step 61, construct the road network topology index, which includes two sub-indices of global topology K f and center node topology V n The global topology K f includes two sub-indices of total node number K node and total edge number K edge The center node topology includes five sub-indices of degree centrality K dc , betweenness centrality K bc , closeness centrality K cc , PageRank centrality K pr , and eigenvector centrality K ec The calculation formula of the global topology K f is as follows:

[0123] K f = W node *K node +W edge *K edge

[0124] W node +W edge = 1

[0125] In the formula, W node and W edge are the weights of the corresponding sub-indices, and the sum of the weights is 1, which is calculated by the CRITIC weight method in the application, and other reasonable weight calculation methods can also be used;

[0126] The calculation formula of the center node topology V n is as follows:

[0127] V n = W dc *K dc +W bc *K bc +W cc *K cc +W pr *K pr +W ec *K ec

[0128] Wdc +W bc +W cc +W pr +W ec = 1

[0129] wherein W dc , W bc , W cc , W pr and W ec are weights corresponding to the centrality, the weights and are 1, and are calculated by CRITIC weight method in the present application, and other reasonable weight calculation methods can also be used;

[0130] Step 62, after the road network topology structure index is constructed in step 61, the similarity of the road network to be compared in the road network topology structure index needs to be calculated, wherein the similarity calculation formula of the global topology structure index is:

[0131]

[0132] S f = 1 - |K fjt - K fia |

[0133] wherein K f ' is the normalized global topology structure value, K f is the global topology structure value of the road network, min(K f ) is the minimum value of the global topology structure of all road networks, max(K f ) is the maximum value of the global topology structure of all road networks, S f is the similarity of the road network G jt to be estimated and any road network G ia with detector road section in the global topology structure index, the value range is [0, 1], K' fjt is the global topology structure value of G jt after normalization, and K' fia is the global topology structure value of G ia after normalization;

[0134] The similarity calculation formula of the central node topology structure index is:

[0135] S n = 1 - |K njt - K nia |

[0136] wherein S n is the similarity of G jt and G ia in the central node topology structure index, K njt is the global topology structure value of Gjt the center node topology value of K nia is G ia the center node topology value of K

[0137] Step 63, build the road network functional area attribute index, first clear the required statistical functional area categories are teaching area, residential area, tourist area, commercial area, medical area and work area, corresponding POI types are: teaching area corresponds to middle school and primary school, residential area corresponds to residential area, tourist area corresponds to scenic spot, commercial area corresponds to comprehensive mall and commercial pedestrian street, work area corresponds to industrial park and business building;

[0138] The POI selection range of the local road network functional area is defined as the rectangular area near the center road segment, which specifically includes the rectangular area on the right side of the center road segment and the rectangular area on the right side of the road segment leading to the center road segment;

[0139] The diagonal point coordinates of the rectangular area required for statistics are input to the map API to perform POI statistics in the rectangular area, as shown in Figure 3 The latitude and longitude of one diagonal point is the starting point A of the road segment, and the latitude and longitude of the other point C is calculated by the following formula:

[0140] lat3 = lat2-dis*sin(Azi*π / 180) / (2π*ARC / 360);

[0141] lon3 = lon2+dis*cos(Azi*π / 180) / (2π*ARC*cos(lat3*π / 180));

[0142] In the formula, lat2 and lon2 are the latitude and longitude of the end point B of the road segment, dis is the width of the rectangular range, Azi is the azimuth angle of the rectangular starting point and end point, which can be calculated by the Python geodesic library, and ARC is the length of the Earth's equatorial radius. In addition, the starting point coordinates A of the road segment are selected as a diagonal point, and the diagonal starting point coordinates are substituted into the formula to obtain the coordinates of the other diagonal point;

[0143] After obtaining the data of each road network functional area, the TF-IDF algorithm is used to calculate the weight of each type of functional area in the road network, and the formula is:

[0144] TF_IDF rd = TF rd *IDF d

[0145]

[0146]

[0147] In the formula, TF_IDFrd TF is the weight value of the dth functional area in the local road network r, IDF rd n measures the frequency proportion of the dth functional area in the local road network r, ∑n dr ∑n is the number of the dth functional area in the local road network r, ∑n r IDF is the sum of functional areas in the local road network r, IDF d ∑n measures the frequency of the dth functional area in all road networks, ∑n p ∑n is the sum of all road networks, ∑n pd ∑n is the sum of all road networks containing the dth functional area;

[0148] For any local road network r, the weighted number of the dth functional area is calculated by the general formula:

[0149] K pjtm = TF IDF rd *n dr

[0150] The above weighted number K of all n types of functional areas pjtm After sorting, the functional area attribute index vector is constructed, and the dimension is n.

[0151] Step 64, after the construction of the road network functional area attribute index in step 63, the similarity of the road network to be compared in the road network functional area attribute index needs to be calculated, and the similarity calculation formula is:

[0152]

[0153] In the formula, S p is the similarity of the road network G jt and G ia in the road network functional area attribute index, the value range is [0, 1], K pjt and K pia are the functional area attribute index vectors of G jt and G ia , n is the sum of the types of functional areas, which is 6 here, and m is the mth functional area.

[0154] Step 65, construct the road network distance index, which is the spatial distance between the central nodes of the road network G jt and G ia . The central node of the road segment starting point is substituted into the Haversine formula to calculate the spatial distance between the central nodes of the road network G d and G -2.3x . The spatial correlation of road network flow decreases exponentially with distance, and the correlation between roads 1KM apart is reduced to 0.1. Therefore, the similarity calculation formula of the road network distance index is:

[0155] S d = e -2.3x

[0156] In the formula, S d is the distance index similarity between road networks, the value range is (0, 1], x is the distance between road center nodes (meters), and the road center node coordinates are the midpoint of the road start and end point coordinates;

[0157] Step 66, constructing road network node attribute index, specifically embodied in road design traffic capacity of road network central road section, road traffic capacity is a comprehensive reflection of road grade and lane number, which reflects the upper limit of vehicle flow that road network node can carry, in the design traffic capacity calculation of road section, generally based on the design traffic capacity of a lane, the specific calculation formula is:

[0158] N a = N o *γ*η*C*n′

[0159] In the formula, N a is the design traffic capacity of road (pcu / h), N o is the design traffic capacity of a lane (pcu / h), γ is the bicycle driving correction coefficient, η is the lane width correction coefficient, C is the intersection correction coefficient, and n′ is the lane number correction coefficient, in the present application, only the design traffic capacity of a lane and the influence of lane number are considered, and the remaining correction coefficients are not considered and are defaulted to 1;

[0160] In China, the design value of the design traffic capacity N o of a lane in the city is related to the design traffic speed of road, and the specific corresponding relationship is recorded in the “Urban Road Engineering Design Specification”, as shown in the following table:

[0161] Table 1 Design traffic capacity of a lane

[0162]

[0163] The design traffic speed of road is related to the road grade, and the design speed interval of each grade road is recorded in the “Urban Road Engineering Design Specification”, in the present application, the design speed of each road grade is taken as the highest value in the interval, so as to obtain the corresponding result of table 1, and the lane number correction coefficient n′ is obtained based on the parameters used in “Traffic Engineering”, specifically:

[0164] Table 2 Lane number correction coefficient

[0165]

[0166] After the road design traffic capacity of each road section is calculated, the road network node attribute similarity calculation formula of the target field road network and the selected source field road network is carried out by using the normalization formula in step 62:

[0167] S c =1-|K′ cjt -K′ cia |

[0168] where S c is the similarity of road network node attributes, K′ cjt is the normalized design capacity of the central node road of the local road network G jt of the to-be-estimated road section, and K′ cia is the normalized design capacity of the central node road of the local road network G ia of the road section with detectors.

[0169] Step 67, the indexes of the urban local road network similarity evaluation system are determined by using the CRITIC weight method, the total weight is 1, and the calculation formula of the total similarity of the local road network of the to-be-estimated road section and the local road network of the road section with detectors is as follows:

[0170] S=W f *S f +W n *S n +W p *S p +W d *S d +W c *S c

[0171] where S is the total similarity score of the road network, W f , W n , W p , W d and W c are the weights of the road network topology index (global topology index, central node topology index), road network functional area attribute index, road network distance index and road network node attribute index respectively, S f , S n , S p , S d and S c are the similarities of the indexes.

[0172] Step 7, the local road network G jt of the to-be-estimated road section is defined as the target field road network, the index information of G jt and the index information of all local road networks in G i are input into the urban local road network similarity evaluation system, and the road network G jt with the highest similarity to G is is obtained as the source field road network.

[0173] Step 8, according to G jt , Gis The connection relationship of the middle node generates two adjlist files, according to G jt , G is The historical traffic data of all nodes contained in the middle node generates two traffic csv files, according to G is The adjlist file and the csv file of the road network are input into the traffic estimation model GE-GAN, which is proposed in the paper "GE-GAN: A novel deep learning framework for road traffic state estimation", the model architecture is as shown in Figure 4 , which realizes the information representation of the road network based on graph embedding, and then generates real-time traffic data based on the representation information using the GAN network, thereby obtaining the real-time estimated traffic of the source field road network G is It should be noted that other traffic estimation models based on graph information can also be used.

[0174] Step 9, the GE-GAN model trained in the source field road network G jt is applied to the target field road network G jt after fine-tuning, as shown in Figure 5 , the specific steps are: input the adjlist file of the road network G jt into the model to obtain the information representation of the road network G jt , and then use a small amount of manually collected traffic data of the central node V jt of the road network G jt , the data duration is not more than one day, as Ground Truth, to perform secondary training on the basis of the GE-GAN model parameters trained in the source field road network G jt , and the real-time estimated traffic of the target field road network G jt can be generated using the model after secondary training.

[0175] The above-described embodiments are only a preferred scheme of the present application, and are not intended to limit the present application. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present application. Therefore, any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present application.

Claims

1. A method for real-time traffic flow estimation on detector-free road sections in urban road networks based on transfer learning, characterized in that, include: S1. For the target city road network, divide the road into several road segments with intersections as the dividing points and mark them with serial numbers. Use road segments as nodes and the connection direction between road segments as the edge direction to abstract the city road network into a city road network graph. Then, based on the city road network graph, construct the local road network corresponding to each road segment node in the graph. The local road network of each road segment node is a road network composed of all nodes and edge connections within the two-hop range of the central node, with the road segment node as the central node. S2. Using road network topology indicators, road network functional area attribute indicators, road network distance indicators, and road network node attribute indicators as multi-dimensional evaluation indicators to comprehensively reflect the impact of the road network on road traffic flow, a local urban road network similarity evaluation system is established by quantitative analysis and weight assignment of the multi-dimensional evaluation indicators. Then, the similarity between the local road network without detectors to be estimated and the local road networks with detectors is calculated using the local urban road network similarity evaluation system, thereby finding the local road network with detectors that is most similar to the traffic flow distribution of the local road network without detectors to be estimated. S3. Using the local road network of the undetected road segment to be estimated as the target domain of transfer learning, and the local road network of the detected road segment found by the urban local road network similarity evaluation system as the source domain of transfer learning, the fine-tuning method is used to transfer learn the traffic estimation model that has been trained in the source domain in advance, so as to use the traffic estimation model obtained by transfer learning to complete the real-time traffic estimation of the undetected road segment.

2. The method for real-time traffic flow estimation of detector-free road segments in urban road networks based on transfer learning according to claim 1, characterized in that, The specific steps of S2 include: S21. Using multi-source urban traffic data, four indicators are constructed to comprehensively reflect the impact of the road network on road traffic flow from multiple perspectives: road network topology, road network functional area attributes, road network distance, and road network node attributes. Then, the weight of each indicator is determined by the objective assignment method, thereby constructing an urban local road network similarity evaluation system. This system is used to obtain the similarity score between local road networks by weighted summation of the similarity of each evaluation indicator between local road networks. S22. Using the local road network corresponding to the detectorless road segment to be estimated as the target domain local road network, and the local road networks corresponding to other detector-equipped road segments in the urban road network map as the candidate source domain set, the urban local road network similarity evaluation system constructed in S21 is used to comprehensively evaluate the similarity between each detector-equipped road segment local road network in the candidate set and the target domain local road network. The detector-equipped road segment local road network with the highest similarity score to the target domain local road network in the candidate source domain set, i.e., the road segment with the most similar traffic distribution, is selected as the source domain local road network.

3. The method for real-time traffic flow estimation of detector-free road segments in urban road networks based on transfer learning according to claim 2, characterized in that, The specific steps of S3 include: S31. The local road network of the undetector-free road segment to be estimated is taken as the target domain for transfer learning. The local road network with detectors, found in S2, that is most similar to the local road network traffic distribution of the road segment without detectors, is used as the source domain for transfer learning. S32. Using GE-GAN based on graph embedding and generative adversarial networks as the traffic estimation model, the source domain... The graph structure relationship of the local road network corresponding to the road segment with detectors and the historical traffic data recorded by all nodes are input into the GE-GAN model for model training; S33. Using the fine-tuning method in transfer learning, the GE-GAN model trained in S32 is retrained using a small number of manually observed data from the local road network of detectorless road segments in the target domain. This makes the model parameters more suitable for the target domain, thereby achieving real-time traffic estimation for the local road network of detectorless road segments.

4. The method for real-time traffic flow estimation of detector-free road segments in urban road networks based on transfer learning as described in claim 2, characterized in that, The road network topology index includes two indicators: global topology and central node topology. The calculation method for the similarity of global topology and central node topology between two local road networks is as follows: S31. Calculate the total number of nodes and the total number of edges in each local road network. Then, use the weighted sum of the two sub-indicators as the global topology index value. The weights of the two sub-indicators are obtained by using the objective weighting method, and the weight sum is 1. S32. Calculate the degree centrality, betweenness centrality, proximity centrality, PageRank centrality and eigenvector centrality of each local road network center node. Then, use the weighted sum of the five sub-indicators as the topological structure index value of the center node. The weights of the five sub-indicators are obtained by using the objective weighting method, and the weight sum is 1. S33, for two local road networks G jt With G ia The global topological similarity between the two is calculated using the following formula: S j =1-|K′ fjt -K′ fia | In the formula: K′ fjt It is a local road network G jt The normalized result of the global topology index value, K′ fia It is a local road network G ia The normalized result of the global topology index value; For two local road networks G jt With G ia The formula for calculating the similarity of the central node topology between the two is as follows: S n =1-|K njt -K nia | Where: K njt It is a local road network G jt The central node topology index value, K nia It is a local road network G ia The topology index value of the central node.

5. The method for real-time traffic flow estimation of detector-free road segments in urban road networks based on transfer learning as described in claim 2, characterized in that, The road network functional area attribute index is characterized by the POI distribution near the central road segment of the road network, and the two local road networks G jt and G ia The method for calculating the similarity of attribute indicators of road network functional areas between them is as follows: In the formula, n is the total number of functional area types; K pjtm and K piam G jt Functional area attribute index vector K pjt and G ia Functional area attribute index vector K pia The m-th value represents G. jt and G ia The weighted number of functional areas of type m in the middle; For any local road network r, the general formula for calculating the weighted quantity of the d-th functional area is: K pjtm =TF_IDF rd *n dr Where: n dr TF_IDF represents the number of POIs belonging to the d-th functional area in the local road network r. rd Represents n dr The corresponding weighting weights are calculated as follows: Where: ∑n r ∑n represents the sum of functional zones in the local road network r. p Let ∑n be the total number of local road networks. pd This represents the total number of local road networks that include the d-th functional area.

6. The method for real-time traffic flow estimation of detector-free road segments in urban road networks based on transfer learning as described in claim 5, characterized in that, The functional zones include teaching zones, residential zones, tourist zones, commercial zones, medical zones, and work zones. The corresponding POI types for each functional zone are as follows: teaching zones correspond to middle schools and primary schools, residential zones correspond to residential communities, tourist zones correspond to scenic spots and historical sites, commercial zones correspond to shopping malls and pedestrian streets, and work zones correspond to industrial parks and office buildings.

7. The method for real-time traffic flow estimation of detector-free road segments in urban road networks based on transfer learning as described in claim 6, characterized in that, When counting the number of functional zones in a local road network, the statistical scope is the rectangular area near the central road segment, specifically including the rectangular area to the right of the central road segment and the rectangular area to the right of the road segment leading from the central road segment. When counting the number of functional zones, the rectangular range search function of the map API is used to determine whether a POI (Point of Interest) is located on the right side of a road segment. When determining the latitude and longitude of the diagonal points within the rectangular area, the latitude and longitude (lon1, lat1) of one point is the starting point A of the road segment, and the latitude and longitude (lon3, lat3) of another point C are calculated using the following formula: lat3=lat2-dis*sin(Azi*π / 180) / (2π*ARC / 360); lon3=lon2+dis*cos(Azi*π / 180) / (2π*ARC*cos(lat3*π / 180)); In the formula, lat2 and lon2 are the latitude and longitude of the end point B of the road segment, dis is the width of the rectangle, Azi is the azimuth angle of the line connecting the start and end points of the rectangle relative to due north, and ARC is the length of the Earth's equatorial radius.

8. The method for real-time traffic flow estimation of detector-free road segments in urban road networks based on transfer learning as described in claim 2, characterized in that, The road network distance index is the spatial distance between the central road segment of the target domain road network and the central road segment of the candidate source domain road network. The formula for calculating the similarity of the road network distance index between two local road networks is: S d =e -2.3x In the formula, x is the spatial distance between two local road networks.

9. The method for real-time traffic flow estimation of detector-free road segments in urban road networks based on transfer learning as described in claim 2, characterized in that, The road network node attribute index is the road design capacity of the central road segment of the road network. The formula for calculating the similarity of road network node attribute indices between two local road networks is as follows: S c =1-|W s -W t | In the formula, W s and W t These represent two local road networks G. jt and G ia The normalized result of road capacity.

10. The method for real-time traffic estimation of detector-free road segments in urban road networks based on transfer learning as described in claim 4, characterized in that, The indicators of the urban local road network similarity evaluation system are determined using the CRITIC weighting method, with a total weight of 1. The formula for calculating the total similarity score between the target domain local road network and the candidate source domain local road networks is as follows: S=W f *S f +W n *S n +W p *S p +W d *S d +W c *S c In the formula, S f S n S p S d and S c The similarity of W represents the similarity of global topology indicators, central node topology indicators, road network functional area attribute indicators, road network distance indicators, and road network node attribute indicators. f W n W p W d and W c These are the weights of the global topology structure index, the central node topology structure index, the road network functional area attribute index, the road network distance index, and the road network node attribute index, respectively.

Images