Method for extracting motor vehicle trip chains based on fixed detector data

Abstract

The application discloses a kind of motor vehicle trip chain extraction method based on fixed detector data, comprising the following steps: S1, according to license plate information search all fixed detector observation records of the same vehicle, obtain the motor vehicle motion trajectory monitored by fixed detector;S2, search all reasonable travel and activity mode between every two adjacent fixed detector observation records, generate trip chain candidate set;S3, integrate observation probability based on time analysis and transition probability based on space analysis, obtain the matching probability of fixed detector observation trajectory and all candidate trip chains;S4, select the sub-path with the maximum matching probability as the identification result between corresponding fixed detector observation pairs, finally obtain a complete trip chain.The application uses the candidate trip chain generation algorithm for sparse AVI trajectory, fuses sparse AVI data, a large number of GNSS data and GIS data, and can realize the complete output of motor vehicle driver daily trip chain.

Description

Technical Field

[0001] This invention relates to a method for extracting motor vehicle travel chains, and more particularly to a method for extracting motor vehicle travel chains based on fixed detector data. Background Technology

[0002] A trip chain refers to the continuous spatial and temporal movement of a driver at the individual level, i.e., where they participate in activities, how long they spend on those activities, and which routes they choose between them. Trip chain information can serve as a foundation and key input for traffic demand models, including the classic four-stage model and activity-based models. On the one hand, by aggregating individual trip chain information, a dynamic OD matrix can be extracted for four-stage analysis. On the other hand, the spatiotemporal attributes of trip chains and activity chains can be directly extracted from driver trip chain information to build activity-based models.

[0003] Traditional methods for collecting information on motor vehicle travel chains involve manually surveying individuals' participation in activities. However, this method is costly, lacks scalability, and the amount of data available is very limited.

[0004] With the development of information technology and artificial intelligence, the recognition efficiency and accuracy of fixed detector (AVI) systems have been significantly improved. Data observed by fixed detectors installed at intersections, road sections, and parking lot entrances and exits can accurately and directly capture key moments in drivers' travel and activity decisions. On the one hand, activity attribute information, including activity location, start time, duration, and end time, can be accurately recorded by fixed detectors in parking lots. On the other hand, vehicle routes can also be indirectly inferred from passing vehicle records captured by roadside detectors during the journey.

[0005] Extracting driver travel chains from fixed detector data presents several challenges: First, urban fixed detector systems are only installed in some parking lots, meaning only a small portion of drivers' travel activities are recorded, making it impossible to infer attribute information from unobserved activities. Second, the spatial sparsity of fixed detector data increases uncertainty in vehicle operating states between detectors, making it extremely difficult to generate a reasonable candidate set for each detection pair and find the optimal solution from all available combinations. Summary of the Invention

[0006] Purpose of the invention: The purpose of this invention is to provide a method for extracting motor vehicle travel chains based on fixed detector data that can solve the problem of vehicle path recognition under sparse AVI data and achieve complete extraction of daily travel activity chain information of motor vehicle drivers in near full sample data.

[0007] Technical solution: The method for extracting the travel chain of a motor vehicle according to the present invention includes the following steps:

[0008] S1. Based on the license plate information, retrieve all fixed detector observation records for the same vehicle and arrange these records in chronological order to obtain the vehicle movement trajectory monitored by the fixed detectors;

[0009] S2, search for all reasonable travel and activity patterns between every two adjacent fixed detector observation records to generate a candidate set of travel chains;

[0010] S3. Based on the candidate travel chain set, integrate the observation probability based on time analysis and the transition probability based on spatial analysis to obtain the matching probability between the fixed detector observation trajectory and all candidate travel chains.

[0011] S4. Select the sub-path with the highest matching probability as the recognition result between the corresponding fixed detector observation pairs, thereby connecting a complete travel chain for the vehicle.

[0012] Furthermore, in step S1, the license plate number of the same vehicle is used as the vehicle ID. The information set O recorded when the vehicle passes the fixed detector is retrieved and arranged in chronological order to generate the fixed detector observation sequence [O1, ..., O1] for the same vehicle. i ,···,O I ,], where O i Let I be the i-th observation point, and let I be the length of the observation sequence.

[0013] Furthermore, the specific implementation steps of step S2 are as follows:

[0014] S21, let the previous fixed detector observe O i-1 As the root node, the latest updated fixed detector observation O i This is the terminating node; The intermediate node represents all activities in the sub-itinerary; k represents the level of the tree, which represents the number of activities included; each edge of the tree represents a trip in the sub-itinerary; each path in the tree represents a sub-itinerary.

[0015] S22, Set the subpath candidate set Φ i =φ; feasible set Θ1 = φ; maximum number of activities MK; observation probability threshold;

[0016] S23, based on the time constraints of the observation pairs, determine the possible activity locations that the vehicle may reach, and judge O. i .xin Is it valid?

[0017] If true, traverse all possible paths, using the observed probability as a threshold to filter out all reasonable sub-routes in the search tree at level 0, and update the candidate set Φ.i , so that Φ i .appebd(M(K i )=0)∈Φ i ; where Φ i .append(M(K i ) = 0) indicates that there are no active candidates;

[0018] If not, proceed to step S24;

[0019] S24, Let k=1, determine the feasible activity duration d based on the observer's time constraint. * traverse all d * Complete the relevant attributes of the event and update the candidate set Φ i , so that Φ i .append(M(K i )=1)∈Φ i ; where Φ i .append(M(K i ) = 1) indicates that there are active candidates;

[0020] S25, determine whether k≤MK is true;

[0021] If not, return to step S23.

[0022] Otherwise, the candidate options for every two observation pairs are connected in chronological order to form a candidate travel chain.

[0023] Furthermore, the time constraint in step S23 is expressed as follows:

[0024]

[0025] Among them, O i-1 .x is the position of the previous fixed detector; O i .x represents the current detector position; The location of the event is indicated by the traffic area; O i .t is the timestamp when the vehicle passes the detector; O i-1 .t is the timestamp of the vehicle passing the previous detector; SPT() is the shortest time function.

[0026] Furthermore, the time constraint in step S24 is expressed as follows:

[0027]

[0028] in, It is the start time of the first layer with activity candidates; d * This is a feasible activity duration.

[0029] Furthermore, the specific implementation steps of step S3 are as follows:

[0030] S31, Calculate the observation probability: Observation probability Pr(O i .t|M(K i ),O i-1 ) refers to a given hidden motion state h i =M(K i ) and the previous detector observed O i-1 Generate observation O i The probability of [something] is calculated using the following formula:

[0031]

[0032] in, This is the end time of the last activity. Given the motion state M(K) i ) and the previous detector observed O i-1 The probability of the end time of the last activity; It is the distribution function of the path travel time;

[0033] S32, Calculate the transition probability: Transition probability Pr(h) i =M(K i )|h i-1 =M(K i-1 ),O i-1 ) refers to a given AVI observation O i-1 The probability of a vehicle transitioning between adjacent hidden motion states is detected in real time, based on detector observations. i-1 The type is calculated in two cases:

[0034] 321)O i-1 For parking observation

[0035] a1)K i =0

[0036] When the vehicle is in O i-1 and O i When the sub-trip timetable is in continuous driving mode The transition probability at this point is defined as:

[0037]

[0038] in, A binary variable representing whether a vehicle passes through an AVI-monitored section: If When passing through the road section monitored by the AVI system otherwise For connecting O i-1 .x and O i.x A set of candidate paths between two locations; For the route selection model, give the two endpoints of the trip (O) i-1 .x,O i When .x), the driver selects a route. The probability of;

[0039] a2)K i ≥1

[0040] Vehicles in O i-1 and O i There must be at least one activity between them. The transition probability at this point is expressed as:

[0041]

[0042] In the formula, A binary variable representing whether a vehicle passes through the monitored road segment: if When passing through the road section monitored by the detector system otherwise To connect the starting location of the activity and the location of the activity The set of candidate paths between; For the path selection model, given the known starting location of the activity and destination of activities Driver selects route The probability of; Choose a behavior model for the destination, representing the behavior at a given departure point. Choose destination under the conditions The probability of; This is a path selection behavior model connecting intermediate observation points and the first activity;

[0043] 322)O i-1 For roadside observation

[0044] b1)K i ≥1,K i-1 ≥1

[0045] M(K i-1 ) and M(K i Each of these activities contains more than one activity, and the transition probability is expressed as:

[0046]

[0047] in, This gives the probability that the driver will select the sub-trip in the previous itinerary when considering the last trip and activity. The calculation formula is as follows:

[0048]

[0049] Based on intermediate observation O i-1 The spatial constraints on vehicle movement, at observation point O along the known route of the monitored vehicle. i-1 The probability of moving from the last activity of the previous sub-itinerary to the first activity of the current sub-itinerary under condition .x is calculated using the following formula:

[0050]

[0051] Where a and b are parameters in the log-normal distribution; Δd1 is the similarity index between the vehicle's travel distance and the straight-line distance between the two activity locations when the driver moves between two adjacent activities, and the calculation formula is as follows:

[0052]

[0053] in, It is the straight-line distance between two locations. It is the actual distance the driver traveled;

[0054] b2)K i =1,K i-1 ≥0

[0055] Sub-trip schedule M(K) i-1 ) is in continuous driving state, and the trip meter M(K) i It contains at least one activity; its transition probability is expressed as:

[0056]

[0057] in, This gives the probability that the driver will choose the sub-trip when the first trip is on the previous itinerary. Based on space observation O i-1 Restrictions on vehicle activities;

[0058] The probability of a driver selecting a sub-trip can be further decomposed into the product of the route selection model and the destination selection model according to the chain rule:

[0059]

[0060] The restriction term follows a log-normal distribution, as shown below:

[0061]

[0062] The definition of Δd2 is as follows:

[0063]

[0064] b3)K i =0,K i-1 ≥1

[0065] When sub-scheduling M(K) i-1 ) is in continuous driving state and sub-trip table M(K) i When at least one activity is included, its transition probability is expressed as:

[0066]

[0067] in, Is the vehicle in O i and O i-1 The probability of driving continuously without being detected by the detector system is expressed as follows:

[0068]

[0069] The limiting term follows a log-normal distribution and is expressed as follows:

[0070]

[0071] Wherein, Δd3 is defined as follows:

[0072]

[0073] S33, based on the observation probability and transition probability defined in steps S31 and S32, calculate the matching likelihood value between the observed trajectory and the candidate travel chain. The calculation formula is as follows:

[0074]

[0075] Where ∝ represents the proportional relationship; Pr(h1=M(K1)) represents the initial state probability, and Pr(O1|h1=M(K1)) represents the first observation probability. Assuming the driver departs from home and simultaneously generates the first observation O1, M(K1) and O1 are subsequently determined. Therefore, Pr(h1=M(K1))=1 and Pr(O1|h1=M(K1))=1; Pr(O i |h i =M(K i ),O i-1 ) is the observation probability; Pr(h) i =M(K i )|h i-1 =M(K i-1 ),O i-1 ) is the transition probability.

[0076] Furthermore, the probability of the end time in step S31 is expressed as follows:

[0077]

[0078] in, This refers to the duration of the first activity. This represents the probability of activity duration estimated based on a similarity graph model; It is a path The distribution function of travel time;

[0079] The probability of the end time of the kth activity is derived based on the chain rule as follows:

[0080]

[0081] Furthermore, the specific implementation process of step S4 is as follows:

[0082] S41, the matching probability between any daily travel-activity chain and the observation trajectory of the fixed detector is calculated from the length of a path on the candidate map, as follows:

[0083] L(M(K1),…,M(K I ))=logPr(M(K1),…,M(K I ),O1,O I+1 )

[0084] Where, L(M(K1),...,M(K) I )) represents the path length corresponding to a trip-activity chain on the candidate graph; Pr(M(K1),...,M(K I ),O1,...,O I+1 () is the joint distribution of candidate travel-activity chains and fixed detector observation trajectories;

[0085] S42, derive the travel chain that best matches the given fixed detector trajectory conditions. The derivation process includes forward inference and backward tracking:

[0086] S421, Forward Inference

[0087] Defined under the assumption that the driver has already selected the preceding hidden motion state h with the highest probability. 1:i-1 Under the condition, it is in motion state h i =M(K i End and generate the observed AVI trajectory O 1:i The probability is:

[0088]

[0089] Among them, the first item It is the normalization constant, and its calculation is as follows:

[0090]

[0091] The second term Pr(O) i |h i =M(K i ),O i-1 ) represents the observation probability;

[0092] The third item is further broken down as follows:

[0093]

[0094] Among them, Pr(h i =M(K i )|h i-1 =M(K i-1 ),O i-1 ) is the detector observing O i-1 From hidden state h under the space constraints of .x i-1 =M(K i-1 ) Transfer to h i =M(K i The probability of ).

[0095] Therefore, F(M(K) i The calculation is as follows:

[0096]

[0097] S422, backward tracking

[0098] The requirement is to record the previous hidden state during each iteration:

[0099]

[0100] Among them, G(M(K) i For each feasible hidden motion state M(K) i Return the previous motion state with the highest probability.

[0101] Repeat the iterative process until the final state is obtained.

[0102]

[0103] Ultimately, the trajectory O of the same vehicle and the detector was obtained. 1:I A complete matching travel chain:

[0104] Compared with the prior art, the significant advantages of this invention are as follows:

[0105] 1. This invention adopts a candidate travel chain generation algorithm for sparse AVI trajectories, which integrates sparse AVI data, a large amount of GNSS data and GIS data. It can effectively identify the activity locations of motor vehicle drivers, splice the candidate travel activity items between two observation pairs in chronological order, construct a motor vehicle travel chain inference model, and propose an optimal travel chain solution algorithm to achieve the complete output of the daily travel chain of motor vehicle drivers.

[0106] 2. This invention uses semi-supervised learning techniques to solve the problem of vehicle path recognition under sparse AVI data, and realizes the complete extraction of daily travel activity chain information of motor vehicle drivers in near full sample, providing richer basic data for traffic condition monitoring applications, and also providing a new solution for analyzing the causes of congestion at the traffic demand level. Attached Figure Description

[0107] Figure 1 This is a flowchart of the present invention;

[0108] Figure 2 A schematic diagram of the search tree constructed in this invention;

[0109] Figure 3 This is a candidate graph for searching the optimal travel chain in this invention;

[0110] Figure 4(a) is a schematic diagram of the estimation error distribution of the activity location;

[0111] Figure 4(b) is a schematic diagram of the error distribution in the estimation of activity duration;

[0112] Figure 4(c) is a schematic diagram of the estimation error distribution of the driving path;

[0113] Figure 5(a) is a schematic diagram of the estimation error of the rule-based method;

[0114] Figure 5(b) is a schematic diagram of the estimation error of the method based on the HMM model;

[0115] Figure 5(c) is a schematic diagram of the estimation error of the ST-II method;

[0116] Figure 6 This is a diagram showing the estimation accuracy of the present invention under different traffic zone sizes;

[0117] Figure 7 This is a diagram showing the estimation accuracy of the present invention under different AVI observation intervals;

[0118] Figure 8 For Δd 1,2,3 Histogram and log-normal distribution fitting curve. Detailed Implementation

[0119] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0120] The flowchart of the motor vehicle travel chain extraction method of the present invention is as follows: Figure 1 As shown, the specific steps include the following:

[0121] Step S1, Vehicle trajectory extraction: Using the license plate number of the same vehicle as the vehicle ID, retrieve the information set O recorded when the vehicle passes the fixed detector, arrange them in chronological order, and generate the fixed detector observation sequence [O1, ..., O1] for the same vehicle. i , ···, O I ], where O i Let I be the i-th observation point, and let I be the length of the observation sequence.

[0122] Step S2, generating a candidate travel chain set: Based on the fixed detector data, search for all reasonable travel and activity patterns between every two adjacent fixed detector observation records, generating a sub-path candidate set for each fixed detector observation pair; specifically including the following steps:

[0123] Step S21, construct the search tree, such as Figure 2 As shown, let the previous fixed detector observe O i-1 As the root node, the latest updated fixed detector observation O i For the termination node, The middle node represents all activities in the sub-itinerary; k represents the tree level (i.e., the search level); k represents the number of activities included; the maximum level is set to 5; each edge of the tree represents a trip in the sub-itinerary; each path in the tree represents a sub-itinerary.

[0124] Step S22, Algorithm initialization. Set the sub-path candidate set Φ i =φ; feasible set Θ1 = φ; maximum number of activities MK = 5; observation probability threshold ε = 0.3.

[0125] Step S23: Generate no-activity candidates. Based on the time constraints of the observation pairs (see formula (1)), determine the possible activity locations that the vehicle may reach, and judge O. i .xin Is it valid?

[0126] If true, traverse all possible paths, using the observed probability as a threshold to filter out all reasonable sub-routes in the search tree at level 0, and update the candidate set Φ. i , so that Φ i .append(M(K i )=0)∈Φ i ; where Φ i .append(M(Ki ) = 0) indicates that there are no active candidates.

[0127] Otherwise, proceed to step S24.

[0128] The time constraint is expressed as follows:

[0129]

[0130] Among them, O i-1 .x is the position of the previous fixed detector; O i .x represents the current detector position; The location of the event is indicated by the traffic area; O i .t is the timestamp when the vehicle passes the detector; O i-1 .t is the timestamp of the vehicle passing the previous detector; SPT() is the shortest time function.

[0131] Step S24: Generate active candidates. Let the search layer number k = 1, and determine the feasible activity duration d based on the observer's time constraint (see formula (2)). * traverse all d * Complete the relevant attributes of the event and update the candidate set Φ i , so that Φ i .append(M(K i )=1)∈Φ i ; where Φ i .append(M(K i =1) indicates that there are active candidates.

[0132] The time constraint is expressed as follows:

[0133]

[0134] in, It is the start time of the first layer with activity candidates; d * This is a feasible activity duration.

[0135] Step S25, Loop body: If the number of search layers k≤5, then return to step S23 and update the candidate path set; otherwise, connect the candidates of every two observation pairs in chronological order to form a candidate path chain.

[0136] Step S3, calculate the matching likelihood value between the observed trajectory and the candidate travel chains: For the set of candidate travel chains for the same vehicle, integrate the observation probability based on time analysis and the transition probability based on spatial analysis to obtain the matching probability between the fixed detector's observed trajectory and all candidate travel chains. This specifically includes the following steps:

[0137] Step S31, calculate the observation probability, as follows:

[0138]

[0139] In the formula: This is the end time of the last activity. Given the motion state M(K) i ) and the previous detector observed O i-1 The probability of the end time of the last activity; It is the distribution function of the path travel time.

[0140] The probability of the end time is represented as follows:

[0141]

[0142] In the formula, This refers to the duration of the first activity. This represents the probability of activity duration estimated based on a similarity graph model; It is a path The distribution function of travel time.

[0143] The probability of the end time of the kth activity can be derived based on the chain rule as follows:

[0144]

[0145] Step S32, calculate the transition probability. Based on the observer's observation O i-1 The transition probability is calculated in two cases for each type.

[0146] (1)O i-1 For parking observation

[0147] a)K i =0

[0148] That is, when the vehicle is in O i-1 and O i When the sub-trip timetable is in continuous driving mode The transition probability at this point can be defined as:

[0149]

[0150] In the formula, A binary variable indicating whether a vehicle passes through an AVI-monitored section: if When passing through the road section monitored by the AVI system otherwise For connecting O i-1 .x and O i .x A set of candidate paths between two locations; For the route selection model, give the two endpoints of the trip (O) i-1 .x,O i When .x), the driver selects a route. The probability of.

[0151] b)K i ≥1

[0152] That is, the vehicle is in O i-1 and O i There must be at least one activity between them. The transition probability at this point can be expressed as:

[0153]

[0154] In the formula, A binary variable indicating whether a vehicle passes through the monitored road segment: if When passing through the road section monitored by the detector system otherwise To connect the starting location of the activity and the location of the activity The set of candidate paths between; For the path selection model, given the known starting location of the activity and destination of activities Driver selects route The probability of; The destination selection behavior model describes the behavior of choosing a destination from a given origin. Choose destination under the conditions The probability of; This is a path selection behavior model connecting intermediate observation points and the first activity.

[0155] (2)O i-1 For roadside observation

[0156] a)K i ≥1,K i-1 ≥1

[0157] That is, M(K) i-1 ) and M(K i Each of these activities contains more than one activity, and the transition probability can be expressed as:

[0158]

[0159] In the formula, This gives the probability that the driver will select the sub-trip in the previous itinerary when considering the last trip and activity. The calculation formula is as follows:

[0160]

[0161] Based on intermediate observation O i-1 The spatial constraint on vehicle movement, i.e., the observation point O along the known route of the monitoring vehicle. i-1 The probability of moving from the last activity of the previous sub-itinerary to the first activity of the current sub-itinerary under condition .x is calculated using the following formula:

[0162]

[0163] In the formula, a and b are unknown parameters in the log-normal distribution, which can be estimated from the observed samples (the estimation results are as follows). Figure 8 Fitting results: The goodness-of-fit ratio is 0.99. Δd1 is a similarity index between the vehicle's travel distance and the straight-line distance between the two activity locations when the driver moves between two adjacent activities. The calculation formula is as follows:

[0164]

[0165] In the formula, It is the straight-line distance between two locations. and It is the actual distance the driver traveled.

[0166] b)K i =1,K i-1 ≥0

[0167] That is, sub-trip table M(K) i-1 ) is in continuous driving state, and the trip meter M(K) i It contains at least one activity. Its transition probability can be expressed as:

[0168]

[0169] In the formula, This gives the probability that the driver will choose the sub-trip when the first trip is on the previous itinerary. Based on space observation O i-1 Restrictions on vehicle activities.

[0170] The probability of a driver choosing a sub-trip can be further decomposed into the product of a route selection model and a destination selection model using the chain rule:

[0171]

[0172] The constraint term follows a log-normal distribution and can be represented as follows:

[0173]

[0174] In the formula, Δd2 is defined as follows:

[0175]

[0176] c)K i =0,K i-1 ≥1

[0177] That is, when the sub-trip table M(K) i-1 ) is in continuous driving state and sub-trip table M(K) i When at least one activity is included, its transition probability can be expressed as:

[0178]

[0179] In the formula, In essence, it is the vehicle in O i and O i-1 The probability of driving continuously without being detected by the detector system is expressed as follows:

[0180]

[0181] The limiting term follows a log-normal distribution and is expressed as follows:

[0182]

[0183] Wherein, Δd3 is defined as follows:

[0184]

[0185] Step S33: Based on the observation probability and transition probability, calculate the matching likelihood value between the observed trajectory and the candidate travel chain. The calculation formula is as follows:

[0186]

[0187] In the formula, ∝ represents a proportional relationship, so the Bayesian rule can be expressed as Pr(A|B)∝Pr(B|A)Pr(A); Pr(h1=M(K1)) represents the initial state probability, and Pr(O1|h1=M(K1)) represents the first observation probability. Assuming the driver starts from home and generates the first observation O1, M(K1) and O1 are determined accordingly, then Pr(h1=M(K1))=1 and Pr(O1|h1=M(K1))=1; Pr(O i |h i =M(K i ),O i-1 ) is the observation probability; Pr(h) i =M(K i )|h i-1 =M(K i-1 ),O i-1) represents the transition probability. Step S4 will utilize an optimal search algorithm to solve for this probability.

[0188] Step S4: Search for the travel chain with the highest matching likelihood value: Based on the matching probability of the observation trajectory of the fixed detector of the same vehicle and the candidate set of sub-paths, select the sub-path with the highest matching probability as the identification result between the corresponding fixed detector observation pairs, thereby connecting a complete travel chain of the vehicle.

[0189] Step S41: Construct candidate graphs. For example... Figure 3 As shown, the candidate graph structure G C (V C E C The graph is a directed acyclic graph. Nodes represent feasible sub-routes between every two adjacent AVI observations (generated in step S2), and edges represent the correlations between nodes of different feasible sub-routes. Only reasonable candidate sub-routes (gray dots) are connected by edges. The weight of each node in the candidate graph is defined by the observation probability of the sub-routes it represents (see step S31), while each edge is defined by the transition probability between two sub-routes (see step S32).

[0190] Based on the candidate graph defined above, the matching probability between any daily travel-activity chain and the observation trajectory of a fixed detector can be calculated from the length of a path on the candidate graph, as follows:

[0191] L(M(k1),…,m(k i ))=logPr(m(k1),…,m(k i ),O1,O I+1 ) (twenty one)

[0192] In the formula, L(M(k1),...,M(K) I )) represents the path length corresponding to a trip-activity chain on the candidate graph; Pr(M(K1),...,M(K I ),O1,...,O I+1 ) is the joint distribution of candidate travel-activity chains and observed trajector trajector trajector trajector trajector . Since the ST-II model assumes that the first sub-trip schedule is known, we have logPr(h1=M(k1))=0.

[0193] Step S42: Solve for the optimal solution. Derive the travel chain that best matches the given fixed detector trajectory conditions. The specific derivation process includes forward inference and backward tracking.

[0194] (1) Forward inference

[0195] Defined under the assumption that the driver has already selected the preceding hidden motion state h with the highest probability. 1:i-1Under the condition (i.e., in multiple preceding hidden motion states h) 1:i-1 In the process, the preceding hidden motion state h with the highest probability is selected. 1:i-1 ), which is in motion state h i =M(K i End and generate the observed AVI trajectory O 1:i The probability is:

[0196]

[0197] In the formula, the first term It is the normalization constant, and its calculation is as follows:

[0198]

[0199] The second term Pr(O) i |h i =M(K i ),O i-1 ) is the observation probability, and its calculation method is shown in step S31;

[0200] The third item can be further broken down as follows:

[0201]

[0202] In the formula, Pr(h) i =M(K i )|h i-1 =M(K i-1 ),O i-1 ) is the detector observing O i-1 From hidden state h under the space constraints of .x i-1 =M(K i-1 ) Transfer to h i =M(K i The probability of ).

[0203] Therefore, F(M(K) i It can be calculated using the following formula:

[0204]

[0205] (2) Backtracking

[0206] The derivation process of the travel-activity chain defined by ST-II is a continuously iterative recursive process. The preceding hidden state h is used to output the derivation. 1:i-1 This requires recording the previous hidden state during each iteration:

[0207]

[0208] In the formula, G(M(K) iFor each feasible hidden motion state M(K) i Return the previous motion state with the highest probability. (Note: The previous state of motion with the highest probability) That is, from multiple previous states of motion Choose the one with the highest probability.

[0209] Repeat the above procedure until the final state is obtained.

[0210]

[0211] Ultimately, the trajectory O of the same vehicle and the detector can be obtained. 1:I A complete matching travel chain:

[0212]

[0213] Example 1:

[0214] (1) Data collection and extraction

[0215] Highway network: The experiment used the highway network of Shenzhen, China, which is 21,985 kilometers long. The city's network map covers a spatial area of ​​40×50 kilometers, containing 237,440 nodes and 215,771 road connections.

[0216] AVI Dataset: An AVI dataset was collected in Shenzhen from September 1st to October 31st, 2016, based on data from 715 fixed sensors. The mean, median, and variance of the time intervals between AVI observation pairs were calculated to be 12.18 min, 5.54 min, and 252.83 min, respectively. 2 The corresponding statistical data for spatial intervals are 16.25 km, 10.23 km, and 292.93 km, respectively. 2 The statistical indicators of AVI data are shown in Table 1. The temporal and spatial intervals of AVI data exceed the scope of sparsity issues considered in existing map matching research.

[0217] Table 1 AVI Data Statistical Indicators

[0218] mean median variance Time (min) 12.18 5.54 252.83 Space (km) 16.25 10.23 292.93

[0219] (2) Experimental Design

[0220] The experiment compares the method proposed in this invention with two other types of methods:

[0221] The first category is a direct and simple rule-based model. Considering that existing methods cannot identify unobserved activities, improvements are needed based on some theoretical methods proposed in this work. First, a candidate set generation algorithm is used to construct feasible sub-trip tables between every two AVI observations. Then, the item with the highest observation probability is selected as the matching result for that observation pair. By concatenating the matching sub-trip tables between all AVI observations, the driver's complete daily travel chain information can be obtained. In this case, the rule-based method is essentially an ST-II model that does not consider observation probabilities.

[0222] The second type of approach is the model-based method, which will be used as the prototype. Given that this method assumes a one-to-one correspondence between data observations and hidden activity states, it also cannot be directly applied to sparse AVI data. To address this issue, sub-trip schedules are defined as hidden node states, and the observation probabilities and transition probabilities connecting each node in the HMM model are redefined. Both this and the ST-II model can be considered dynamic Bayesian models. The difference is that the HMM model assumes that each data observation is independent, while the ST-II model considers not only the correlation between data observations but also the spatiotemporal constraints of data observations on travel behavior.

[0223] (3) Experimental Results

[0224] Using the proposed algorithm for extracting motor vehicle travel chains based on fixed detector data, compared with two other methods, it is demonstrated that the proposed algorithm has higher accuracy in terms of activity location, activity duration, and driving route, as shown in Figures 4(a), (b), and (c).

[0225] When the number of activities in the sub-trip schedules of AVI observation pairs is large, accurately inferring the travel-activity chain information of motor vehicle drivers becomes a very challenging task. Therefore, it is necessary to test the robustness of different methods to the number of activities in the sub-trip schedules. Figures 5(a), (b), and (c) show the estimation errors of three methods for sub-trip schedules with different numbers of activities, respectively. When the actual sub-trip schedule is a continuous travel state (K... i =0) and 1 activity (K) i When the number of activities is 1, all methods show good accuracy. However, as the number of activities continues to increase, the computational errors of rule-based and HMM-based methods increase rapidly, while the computational accuracy of the ST-II method only decreases slightly. In the test set samples, 77.68% of the observation pairs had fewer than 3 activities in their sub-runtime tables. This means that the proposed ST-II method can perform well in most application scenarios. In summary, this method can be considered to have good robustness to the number of activities with missing observations.

[0226] (4) Sensitivity analysis

[0227] ① Traffic area dimensions

[0228] The study area was divided into 100×100m 2 200×200m 2 500×500m 2 1000×1000m 2 and 2000×2000m 2 Five types of TAZ meshes were used, and experiments were conducted on the test datasets respectively. Figure 6 Three error metrics (i.e., η) are presented. location ,η duration ,andη path A line graph showing the variation with traffic cell size. It can be observed that the proposed ST-II method is quite sensitive to traffic cell size. When the resolution of traffic cell division is lower than 500×500m... 2 The estimation error limit decreases as the traffic cell size increases. Intuitively, higher traffic cell resolution would increase the difficulty of identifying activity locations, thus leading to a higher estimation error. However, surprisingly, the estimation error limit decreases when the traffic cell size exceeds 500×500m. 2 At that time, the estimation error increases in the same direction as the traffic cell size. This may be because some short-distance trips are covered by excessively large traffic cells, limiting ST-II's ability to identify travel routes. This can also be used to explain why η path As the size of the traffic cell increases, it is proportional to η location and η duration The growth rate is faster. The above analysis suggests that for travel-activity chain inference, there may be an optimal value for the traffic cell size that can achieve the highest accuracy (e.g., 500×500m in this experiment). 2 ).

[0229] ②AVI observation spacing

[0230] Because the distribution of AVI detectors on urban road networks is very uneven and sparse, it is necessary to test the impact of AVI observation spacing on the accuracy of estimation results, so as to provide reference experience for the layout scheme of AVI detectors.

[0231] Based on the spatial interval of the AVI observation pairs, the AVI observations of the test dataset were divided into four levels: [0-2km, 2-5km, 5-10km, 10-15km], and the distribution of their estimation errors was statistically analyzed for each level. Figure 7 The distribution of the three error indices was displayed using a box plot. Figure 8 For Δd 1,2,3The histogram and log-normal distribution fitting curves were obtained. Unsurprisingly, the estimation error increased with the increase of the spatial distance between AVI observation pairs. Furthermore, it was observed that the dispersion of the error distribution for each index also increased with their spacing, indicating that the sparser the AVI observation points, the greater the uncertainty of the travel-activity chain estimation results. The above analysis shows that when pursuing higher travel-activity chain estimation accuracy, the AVI system requires a denser detector deployment. This will consume a significant amount of AVI system construction costs and computational storage resources. Therefore, when the proposed ST-II is applied to travel-activity chain inference, a balance needs to be struck between accuracy and efficiency. Figure 7 The experimental results presented will serve as important reference data for the practical application of ST-II. It should be understood that the above description of the preferred embodiments is quite detailed and should not be construed as limiting the scope of protection of this invention. Those skilled in the art, guided by this invention, can make substitutions or modifications without departing from the scope of the claims, all of which fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for extracting motor vehicle travel chains based on fixed detector data, characterized in that, include The following steps are required: S1. Based on the license plate information, retrieve all fixed detector observation records for the same vehicle and arrange these records in chronological order to obtain the vehicle movement trajectory monitored by the fixed detectors; S2, search for all reasonable travel and activity patterns between every two adjacent fixed detector observation records to generate a candidate set of travel chains; S3. Based on the candidate travel chain set, integrate the observation probability based on time analysis and the transition probability based on spatial analysis to obtain the matching probability between the fixed detector observation trajectory and all candidate travel chains. S4. Select the sub-path with the highest matching probability as the identification result between the corresponding fixed detector observation pairs, and obtain a complete travel chain based on the identification result. In step S1, the license plate number of the same vehicle is used as the vehicle ID. The information set O recorded when the vehicle passes the fixed detector is retrieved and arranged in chronological order to generate the fixed detector observation sequence of the same vehicle. ,in For the first i One observation point, I The length of the observation sequence; The specific implementation steps of step S3 are as follows: S31, Calculate the observation probability: Observation probability This refers to a given hidden motion state Observations from the previous detector Generate observations The probability of [something] is calculated using the following formula: in, This is the end time of the last activity. Given motion state Observations from the previous detector The probability of the end time of the last activity; It is the distribution function of the path travel time; S32, Calculate the transition probability: Transition probability This refers to a given AVI observation The probability of a vehicle transitioning between adjacent hidden motion states is detected in real time, based on detector observations. The type is calculated in two cases: 321) For parking observation a1) When the vehicle is and When the sub-trip timetable is in continuous driving mode The transition probability at this point is defined as: in, A binary variable representing whether a vehicle passes through an AVI-monitored section: If When passing through the road section monitored by the AVI system ,otherwise ; For connection and A set of candidate routes between two locations; For the route selection model, give the two endpoints of the trip. At that time, the driver selects a route The probability of; a2） The vehicle is and There must be at least one activity between them. The transition probability at this point is expressed as: In the formula, A binary variable representing whether a vehicle passes through the monitored road segment: if When passing through the road section monitored by the detector system ,otherwise ; To connect the starting location of the activity and the location of the activity The set of candidate paths between; For the path selection model, given the known starting location of the activity and destination of activities Driver selects route The probability of; Choose a behavior model for the destination, representing the behavior at a given departure point. Choose destination under the conditions The probability of; This is a path selection behavior model connecting intermediate observation points and the first activity; 322 For roadside observation b1) and Each activity contains more than one activity, and the transition probability is expressed as: in, This gives the probability that the driver will select the sub-trip in the previous itinerary when considering the last trip and activity. The calculation formula is as follows: Based on intermediate observations The spatial constraints on vehicle movement, at known observation points along the monitoring vehicle's route. The probability of moving from the last activity of the previous sub-itinerary to the first activity of the current sub-itinerary under certain conditions is calculated using the following formula: in, and These are the parameters in the log-normal distribution; This is a similarity index between the vehicle's travel distance and the straight-line distance between the two activity locations when the driver moves between two adjacent activities. The calculation formula is as follows: in, , It is the actual distance the driver traveled; b2） Sub-itinerary It is in continuous driving mode, and the trip meter It contains at least one activity; its transition probability is expressed as: in, This gives the probability that the driver will choose the sub-trip when the first trip is on the previous itinerary. Based on space observation Restrictions on vehicle activities; The probability of a driver selecting a sub-trip can be further decomposed into the product of the route selection model and the destination selection model according to the chain rule: The restriction term follows a log-normal distribution, as shown below: The definition is as follows: ； b3） When the schedule For continuous driving and sub-trip table When at least one activity is included, the transition probability is expressed as: in, Is the vehicle in and The probability of driving continuously without being detected by the detector system is expressed as follows: The limiting term follows a log-normal distribution and is expressed as follows: in, The definition is as follows: ； S33, based on the observation probability and transition probability defined in steps S31 and S32, calculate the matching likelihood value between the observed trajectory and the candidate travel chain. The calculation formula is as follows: in, Represents proportional relationships; This represents the initial state probability. This represents the probability of the first observation, assuming the driver departs from home, thus generating the first observation. Then there is = 1 and = 1; Pr It is the probability of observation; It is the transition probability.

2. The method for extracting motor vehicle travel chains based on fixed detector data according to claim 1, characterized in that, The specific implementation steps of step S2 are as follows: S21, let the previous fixed detector observe As the root node, the latest updated fixed detector observations This is the terminating node; , ··· The intermediate node represents all activities in the sub-itinerary; k represents the level of the tree, which represents the number of activities included; each edge of the tree represents a trip in the sub-itinerary; each path in the tree represents a sub-itinerary. S22, Set the subpath candidate set ; Feasible set Determine the maximum number of activities (MK) and the observation probability threshold; S23, based on the time constraints of the observation pairs, determine the possible activity locations that the vehicle may reach, and judge... in Is it valid? If true, traverse all possible paths, filter out all reasonable sub-routes in the search tree at level 0 based on the observed probability threshold, and update the candidate set. ,make ;in, This indicates that there are no active candidates. If not, proceed to step S24; S24, Let k=1, determine the feasible activity duration based on the observer's time constraints. traverse all Complete the relevant attributes of the event and update the candidate set. ,make ;in, This indicates that there are activity candidates; S25, Judgment Is it valid? If not, return to step S23; If true, connect the candidate options of every two observation pairs in chronological order to form a candidate travel chain.

3. The method for extracting motor vehicle travel chains based on fixed detector data according to claim 2, characterized in that, The time constraint in step S23 is expressed as follows: in, It is the position of the previous fixed detector; This is the current detector position; The location of the event is indicated by the traffic area; It is the timestamp when the vehicle passes the detector; It is the timestamp of the vehicle passing the previous detector; ( ) is the shortest time function.

4. The method for extracting motor vehicle travel chains based on fixed detector data according to claim 3, characterized in that, The time constraint in step S24 is expressed as follows: in, It is the start time of the activity candidates in the first layer; This is a feasible activity duration.

5. The method for extracting motor vehicle travel chains based on fixed detector data according to claim 1, characterized in that, The probability of the end time in step S31 is expressed as follows: in, This refers to the duration of the first activity. This represents the probability of activity duration estimated based on a similarity graph model; It is a path The distribution function of travel time; No. The probability of the end time of an activity is derived based on the chain rule as follows: 。 6. The method for extracting motor vehicle travel chains based on fixed detector data according to claim 1, characterized in that, The specific implementation process of step S4 is as follows: S41, the matching probability between any daily travel-activity chain and the observation trajectory of the fixed detector is calculated from the length of a path on the candidate map, as follows: in, This indicates the path length corresponding to a trip-activity chain on the candidate graph; It is the joint distribution of candidate travel-activity chains and observed trajector ... S42, derive the travel chain that best matches the given fixed detector trajectory conditions. The derivation process includes forward inference and backward tracking: S421, Forward Inference Defined under the assumption that the driver has already selected the preceding hidden motion state with the highest probability. Under certain conditions, it is in motion. End and generate the observed AVI trajectory The probability is: Among them, the first item It is the normalization constant, calculated as follows: Second item It is the probability of observation; The third item is further broken down as follows: ( , ) = ( ) F( ) Among them, Pr( ) is the detector observation From hidden state under spatial constraints Transferred to The probability of; therefore, The calculation is as follows: ； S422, backward tracking The requirement is to record the previous hidden state during each iteration: in, For each feasible hidden motion state Return to the previous motion state with the highest probability. ; Repeat the iteration until the final state is obtained. : Ultimately, the trajectory of the same vehicle and the detector was obtained. A complete matching travel chain: .

Images