CN115206099B - Self-adaptive path inference method for GPS track of vehicle - Google Patents
Self-adaptive path inference method for GPS track of vehicle Download PDFInfo
- Publication number
- CN115206099B CN115206099B CN202210831380.8A CN202210831380A CN115206099B CN 115206099 B CN115206099 B CN 115206099B CN 202210831380 A CN202210831380 A CN 202210831380A CN 115206099 B CN115206099 B CN 115206099B
- Authority
- CN
- China
- Prior art keywords
- candidate
- road
- path
- features
- track
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/01—Detecting movement of traffic to be counted or controlled
- G08G1/0104—Measuring and analyzing of parameters relative to traffic conditions
- G08G1/0125—Traffic data processing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
- G06Q10/047—Optimisation of routes or paths, e.g. travelling salesman problem
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/01—Detecting movement of traffic to be counted or controlled
- G08G1/052—Detecting movement of traffic to be counted or controlled with provision for determining speed or overspeed
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/09—Arrangements for giving variable traffic instructions
- G08G1/0962—Arrangements for giving variable traffic instructions having an indicator mounted inside the vehicle, e.g. giving voice messages
- G08G1/0968—Systems involving transmission of navigation instructions to the vehicle
- G08G1/096805—Systems involving transmission of navigation instructions to the vehicle where the transmitted instructions are used to compute a route
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Human Resources & Organizations (AREA)
- Theoretical Computer Science (AREA)
- Economics (AREA)
- Strategic Management (AREA)
- Radar, Positioning & Navigation (AREA)
- Biomedical Technology (AREA)
- Quality & Reliability (AREA)
- Tourism & Hospitality (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Entrepreneurship & Innovation (AREA)
- Game Theory and Decision Science (AREA)
- Development Economics (AREA)
- Remote Sensing (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Operations Research (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Chemical & Material Sciences (AREA)
- Analytical Chemistry (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a self-Adaptive (API) path inference method for sparse tracks, which comprises (1) constructing a road network model to obtain a training set and a road section characterization vector, and initializing a model parameter set; (2) obtaining a candidate point set and a candidate path set Γ; (3) Calculating single-point features, namely distance features, of the track points, and front and back features, namely space features, time features, angle features and traffic correlation features, of two continuous track points; and calculates a path feature matrix. (4) encoding the features of the candidate trajectories; (5) Inputting the hidden layer characteristics into an attention layer, and obtaining the weight of each characteristic through a full connection layer; (6) And constructing a conditional random field model with self-adaptive weights by using the single-point features, the front-back features and the corresponding feature weights, and solving an optimal inferred path by using a dynamic programming algorithm. (7) the backward propagation algorithm updates the model parameter set. (8) training to converge the model and predict the path. The invention introduces a attention mechanism to adaptively adjust the weights of the characteristics of the track points under road network environments with different complexity, thereby achieving better matching effect under the complex road network environment.
Description
Technical Field
The invention belongs to the field of navigation and intelligent transportation, and particularly relates to a method for GPS track self-adaptive path inference.
Background
With the development of satellite navigation and positioning technology and the increasing reliance of human travel activities on navigation technology, the trajectory data of various human activities is increasingly important. In the background of the gradual popularization of various navigation positioning devices, a great amount of GPS track data is continuously generated at any time. However, there is a large error in the positioning result in the case where GPS signal propagation is disturbed due to the limitation of the GPS positioning itself and a partial scene. The use of non-post-processed GPS trajectory data can cause significant errors. Therefore, the data are often matched with the high-precision urban road network before being processed, so how to efficiently and accurately match the data is also an important loop in the data processing process.
From the process of matching track data, the simplest method for matching the GPS track points to the road network is to match the track points to the road sections closest to the track points. However, in the actual matching process, many factors can occur to influence the accuracy of the matching result. For example, the longer the GPS signal sampling rate has a great influence on the matching, the longer the sampling time interval, the greater the matching difficulty, and the complexity of the road increases the matching difficulty.
In the process of matching the track to the road network, we often need to weigh the front and back trend of the track points and the duty ratio of the characteristics such as the distance from a single point to the road network in different road network environments, when the track is matched to the road network, the weight parameters input to the conditional random field model should be self-adaptive to the weights of the characteristics according to the complexity of the road network, so that the error matching of the complex area of the road network can be greatly reduced.
Disclosure of Invention
First, the technical problem to be solved
In order to solve the problems existing in the existing track matching method under the complex road network, the invention provides a self-adaptive path deducing method of a vehicle GPS track.
(II) technical scheme
The invention comprises the following steps:
an adaptive path inference method for a vehicle GPS track, comprising the steps of:
(1) Constructing a road network model, and obtaining the dependency relationship between road sections and the characterization vector of each road section from the historical track by using a Skip-Gram graph embedding method; the calibration track points correspond to the matching road sections and the matching point data to obtain a training set T trn Initializing a model parameter set { θ }, whichIncluding full connection layer parameters and attention model parameters.
(2) At each locus point p t (x t ,y t ) The side length is D for the center c According to the distance, selecting q road sections with the smallest distance as candidate road sections, and taking the foot corresponding to the minimum vertical distance of the track point in each candidate road section as a candidate point
Locus point p t All candidate points constitute a candidate point set->
For any track tau, a candidate point set C corresponding to all track points is found out from the road network t A set of composed candidate paths Γ.
(3) Calculating characteristics of the track points and the candidate paths; for any track tau, M track points are included, which are divided into u local tracks with length of M, and overlapping degree Ov is set for adjacent tracks i,i+1 Calculating a locus point p t Single point feature of its candidate points
I.e. distance feature->
Two consecutive trace points p t And p t-1 Front and rear characteristics between->
Namely spatial characteristics->
Time characteristics->
Angle feature->
Corresponding matching of two track pointsTraffic-related characteristics of road section->
After solving the characteristics of all the track points, connecting all the characteristics to form a path characteristic matrix LF= [ EF, PF]＝[DF,SF,TF,AF,CF]。
(4) Encoding candidate track features; calculating potential functions of each candidate path in the candidate path set Γ
I.e., a weighted sum of all candidate point features in the candidate path; select->
The largest top q' candidate paths, where the alpha path gamma α Feature matrix +.>
Coding with fully connected layer to obtain hidden layer characteristic of the path
Wherein FC is a fully connected layer; the selected q' candidate path feature matrixes are subjected to full-connection coding to obtain hidden layer features corresponding to each candidate path +.>
Wherein h= [ h ] DF ,h SF ,h TF ,h AF ,h CF ]The method comprises the steps of carrying out a first treatment on the surface of the The full-connection layer encodes the single-point features and the front and back features of all the candidate points to obtain hidden layer features;
(5) Inputting the hidden layer characteristics into an attention layer to obtain each characteristic weight; the processed hidden layer features are respectively used as a query term Q, a key term K and a value term V of the attention model, and the hidden layer features are formed by Att (Q, K) =softmax (QK) T ) Calculating the tensile dot product, taking the tensile dot product as a relation matrix of the relation strength between the characteristics of all hidden layers, and finally passing through the full-connection layer W a =fc (Att (Q, K)), where W a ＝[W DF ,W SF ,W TF ,W AF ,W CF ]I.e., the size of each feature weight;
(6) Constructing a conditional random field model API-CRF of self-adaptive weight; taking the distance characteristic DF as the emission probability, PF= [ SF, TF, AF, CF ]]As transition probabilities, weights W of the respective features are determined a Constructing a conditional random field CRF as the weight of each probability, and taking the average value of the weights of the two local tracks for the weight of the overlapping part of the two local tracks; and finally, according to the characteristics and the weights of the track points and the candidate points, solving an optimal inferred path by using a dynamic programming algorithm, namely, based on the weighted accumulation of the characteristics of the track points of one track and the candidate points of the track as non-normalized probability, finding the maximum non-normalized probability and backtracking to find the corresponding candidate track as the optimal inferred path.
(7) According to training set T in step (1) trn And (3) calculating an error value by the path formed by the matching points and the optimal inferred path y obtained in the step (6), and updating a model parameter set by a backward propagation algorithm.
(8) Repeating the steps (2) to (7) until the model converges, and inputting track points to be predicted into the model to predict, so as to obtain a predicted path.
(III) beneficial effects
The invention has the advantages that:
the self-adaptive path inference model is provided, and a conditional random field model with self-adaptive adjustment weight is designed aiming at matching logic of GPS track points under road networks with different complexity degrees so as to infer the real running path of the GPS track points. The method uses a graph embedding method to extract traffic correlation between road segments, and uses an attention model to capture the relationship between the track and surrounding road segments, so as to adaptively weight different matching features under different complexity degrees in the process of deducing an optimal driving path. The invention selects the true matching data sets with road networks with different complexity degrees, is used for testing the proposed method under different road environments, adaptively adjusts the characteristic weights according to surrounding road information, and can accurately infer the running path corresponding to the GPS track.
Drawings
FIG. 1 is a flow chart of steps performed in the present invention.
FIG. 2 is a block diagram of a Skip-Gram embedded layer network.
FIG. 3 is a diagram of an adaptive weight conditional random field network model according to the present invention.
Fig. 4 is a graph comparing the matching results of the API-CRF of the present invention and the fixed weight C-CRF model, STD-based matching method, and HMM-based matching method on the same test sample.
Detailed Description
For the purposes of clarity, content, and advantages of the present invention, the following detailed description of the specific embodiments thereof will be described in connection with the accompanying drawings and examples:
referring to fig. 1, the specific implementation steps of the present invention are:
1. initializing model parameter set { θ }, data preprocessing
1.1 initializing a model parameter set { θ }
{ θ } represents training parameters, including full connection layer parameters
And attention model parameters (W K ,W Q ,W V ) The method comprises the steps of carrying out a first treatment on the surface of the All parameters in the model parameter set { θ } are given random initial values.
1.2 building a road network model
The road network is defined as a directed graph R (N, E), where N is the connecting nodes of the road network and E is the road segments between the nodes. Each road segment comprises an initial node and a terminal node id of a road, longitude and latitude coordinates of each point in a road network, the length of the road and the traffic time of the road, the middle node comprises longitude and latitude and id information of the node, the dependence relationship among road segments and the characterization vector of each road segment are obtained from a historical track by using a Skip-Gram graph embedding method, and fig. 2 is a schematic diagram of obtaining the road segment characterization vector by using Skip-Gram.
1.3 vehicle track sequence division: for a track tau with M track points, it can be divided into u local tracks with M track points, where adjacent tracks are provided with oneConstant overlap, overlap of i and i+1 local tracks Ov i,i+1 The calculation formula is as follows:
the local track formed by m track points is taken as a batch. The information of the track points in each training batch includes id, longitude, latitude, matching road network id, vehicle speed and track point recording time. Each training batch is combined to form a training data set T trn . The whole training data is then divided into batches and only one batch is processed at a time, each batch comprising a list of track points, the size of which is determined by this parameter of batch size.
1.4 selecting candidate road segments and candidate points
At each locus point p t (x t ,y t ) The side length is D for the center c Selecting q road sections with the smallest distance as candidate road sections according to the distance, and forming a candidate road section set Γ; foot drop corresponding to minimum drop of track point on each candidate road section is used as candidate point
The q value is the minimum value of the candidate point numbers corresponding to the track points, and the maximum value of the candidate point numbers corresponding to the track points.
2 obtaining a track matching feature matrix LF
2.1 calculating distance feature DF, i.e. single point feature
For each locus point p t All have candidate points
Correspondingly, the distance +_can then be determined by the formula of the distance between the two points>
Distance feature determination by distance
2.2 calculation of spatial signature SF
Calculating two adjacent track points by using Di Jie Style algorithm to obtain candidate points
Is the shortest path length of (a)
Two adjacent track points p t-1 And p t Spatial characterization (transfer characterization) thereof>
Can be expressed as:
wherein dis (p) t-1 ,p t ) Is the linear distance of the two track points in the projection coordinate system.
2.3 calculation of time characteristics TF
For the locus point p t-1 And p t Their instantaneous average speed can be approximately regarded as the running average speed of two track points, the candidate point pair thereof
Time characteristics of->
Can be approximated as:
time true ＝t t -t t-1
wherein t is t And t t-1 P is respectively t And p t-1 At the moment v t And v t-1 P is respectively t And p t-1 Is used for the instantaneous speed of the vehicle.
2.4 calculation of angular feature AF
The angle characteristic is the included angle between the track point connecting line and the candidate point connecting line, which can be expressed as:
2.5 calculation of traffic correlation characteristics CF
The road context features obtain the dependency relationship among road segments through the historical tracks, and are improved on the basis of a traditional Graph casting model, and the road relationship model is expressed as G (v, epsilon), wherein v (i) is the road segments among road network nodes N, epsilon (i) is the road segments among road network nodes N, and epsilon (i) is the topology and traffic correlation among the road segments E. Obtaining the shortest path set S between all road segments in a certain length in the road network s In order to express the topological correlation of the two road segments; simultaneous training set T trn Vehicle driving path set S for matching high-frequency track points to road network d The characteristics of driving habits of drivers, road traffic conditions and the like are hidden; path set S after map matching recovery in use d Shortest path set S s Training, and training to obtain a characterization vector emb of each road section E in the road network by using a Skip-gram method.
For the locus point p t-1 And p t Two candidate road segments corresponding to them
The traffic correlation characteristics between can be calculated using the following formula:
wherein the method comprises the steps of
Is p t-1 The j candidate point of the point is located on the road section, < ->
For point p t The j' th candidate point of (c) is located on the road section,
is->
Is characterized by vector>
Is->
Is described.
2.6 all features are connected to form a Path feature matrix LF
LF＝[EF,PF]＝[DF,SF,TF,AF,CF]
Wherein EF is a single point feature, namely a distance feature DF, PF is a front-back feature, and the front-back feature comprises a spatial feature SF, a temporal feature TF, an angle feature AF and an intersection context feature CF.
3 encoding the features
For the trajectory τ, a set of candidate paths may be found from the road network for the α -th candidate path γ of the candidate paths Γ α Including the locus point p t To candidate points
Single point feature between->
Candidate dot->
And candidate point->
Front and rear characteristics between->
t=1, 2, …, M being the number of track points in the track τ.
All features are encoded through the full connection layer to obtain hidden layer features.
Wherein LF represents path characteristics, FC is a fully connected layer, M is the number of tracks, when k=1
A single-point feature representing a t-th track point in an alpha-th candidate path in the track tau; when k is>1, the->
Representing the front-to-back characteristics between the t and t +1 track points in the alpha candidate path in track tau.
Recalculating potential functions of each candidate path in the candidate path set Γ
Selecting the front q' candidate paths with the maximum potential function, and obtaining hidden layer characteristics corresponding to each candidate path after full-connection coding of all the selected candidate paths>
h＝[h 1 ,h 2 ,…,h n ]N is the number of features and h corresponds to the features of each trace point.
4 hidden layer feature input attention layer
Will hide layer feature h k After post-processing, respectively using the post-processing query term Q, key term K and value term V as attention models, and then calculating a stretching dot product by using the following formula to obtain a relation matrix Att (Q, K) of each hidden layer characteristic:
Q＝W Q *h
K＝W K *h
V＝LF
wherein the method comprises the steps of
Mapping the relation mapping strength between the features into the weight of each feature through the full connection layer:
W a ＝FC(Att(Q,K))
wherein W is a The vector has a dimension of 1*n, where n is the number of features, W at this time a Meaning LF feature vector weight, W a Corresponding to LF one by one, and at this time W a Is a common parameter, W for all candidate paths a Are all consistent.
5 adaptive weight conditional random field model API-CRF
Taking the distance characteristic DF as the emission probability, PF= [ SF, TF, AF, CF ]]As transition probability, the weight obtained in the step 4 is used as the weight value of each probability to construct CRF; CRF represents a given track sequence τ (p 1 ,p 2 ,…,p m )，Deducing the path sequence gamma (y 1 ,y 2 ,…,y m ) Is a probability of (2). This probability can be expressed as:
wherein Z (τ) is a normalization factor, defined as:
for the construction of CRF model, potential function is calculated
And (3) decomposing:
wherein lambda is k For characteristic weights, f k Calculating a function lambda for the feature k′ Weighting, mu, single-point features k″ Is the weight of the front and back characteristics, f k′ As a single point characteristic function, t k″ Is a front-back characteristic function.
Representing the weight of a single point feature as ew i The weights of the front and rear features are denoted as ow i ,EF i For single point feature, PF i-1,i Is a front-to-back feature. The potential function can thus be expressed as:
wherein the method comprises the steps of
And when the attention model obtains a weight value, M is the number of track points, and n is the number of features.
Dividing a track into local tracks according to the step 1.3, and thenCalculating each feature of the local track according to the step 2, and obtaining a weight value W corresponding to each feature of the local track through the step 4 a And taking the average value of the two local tracks by the characteristic weight value of the overlapped part between the two local tracks.
6 deducing the optimal Path
Finally, according to the characteristics of the track points and the candidate points thereof and the weights thereof, solving the optimal inferred path gamma by using a Viterbi algorithm * 。
The Viterbi algorithm is an optimal choice problem for a multi-step multi-choice model, which preserves the minimum total cost from the previous step to the current step and the choice of the previous step in the case of the current cost for each choice; based on the characteristics of each candidate trajectory point and the weight thereof, the viterbi recursion of the non-normalized probability is defined as:
non-normalized probability of initial position
Wherein q represents the number of candidate points corresponding to the first GPS point,/>
Path characteristics corresponding weight indicating i-th sub-track,/->
And the kth feature of the jth candidate point corresponding to the kth GPS point is represented.
After calculating the delta variable, calculating a maximum non-normalized probability value by iteration:
then backtracking the optimal path by using the following method
Finally, the optimal inferred path is obtained
6 backward propagation { θ } and updating parameter set
6.1 calculation of losses
The model adopts a maximum likelihood method to estimate { theta }, and the training set is formed by inputting a sequence tau= { p 1 ,p 2 ,…,p M Sum tag sequence γ= { y 1 ,y 2 ,…,y M Likelihood function is:
by replacing the above feature function, it is possible to obtain:
z is the normalization operation of the step 5; taking the negative value of the above, namely the loss function:
Loss＝-L(w)
6.2 backward propagation and updating { θ }
The training requirement is that the real label should correspond to the probability value of the maximum inferred path, i.e., minimize Loss, and the parameters are solved using an optimization method of random gradient descent, with the back propagation update { θ }.
7 repeating the steps 2 to 6 until the loss function reaches the convergence condition, stopping iteration, otherwise, further updating { theta }; and inputting the track points to be predicted into the model for prediction after training is completed, and obtaining a predicted path.
The above process is illustrated in the accompanying drawings
Experimental results:
in the embodiment, 1804 tracks are randomly selected from two areas, namely 766442 track points and corresponding matching point sets, wherein the area A comprises 1042 tracks, 457895 GPS points, the area B comprises 726 tracks and 308547 GPS points, and the sampling rate is 3s; carrying out downsampling treatment on each track and a matching point set thereof to obtain data sets with sampling intervals of 6s, 12s, 18s, 24s, 30s, 36s, 42s, 48s, 54s, 60s, 90s and 120s, and randomly extracting 50% of the data sets from the data sets as training sets and 50% of the data sets as test sets; and (3) selecting four methods of an STD (Standard test device) -based matching method, an HMM (hidden Markov model) -based matching method, a fixed weight CRF model C-CRF and a self-adaptive weight CRF model API to conduct comparison experiments, predicting a test set, and taking a matching accuracy PCM (pulse code modulation) as a judgment index, wherein the comparison results of the experiments are shown in figure 4.
As can be seen from experimental results, the PCM of the self-adaptive path inference method provided by the invention under each sampling rate is higher than the map matching result of the C-CRF, HMM, STD method, so that the track points can be accurately matched to the road network.
As described above, the specific implementation steps of the present invention make the present invention more clear. Any modifications and changes made to the present invention fall within the spirit of the invention and the scope of the appended claims.
Claims (2)
1. An adaptive path inference method for a vehicle GPS track, comprising the steps of:
(1) Constructing a road network, defining the road network as a directed graph R (N, E) formed by a node N and a road segment E between the nodes, and obtaining a dependency relationship between the road segments and a characterization vector of each road segment from a historical track by using a graph embedding method; the calibration track points correspond to the matching road sections and the matching point data to obtain a training set T trn ；
(2) At each locus point p t (x t ,y t ) The side length is D for the center c According to the distance, selecting q road sections with the smallest distance as candidate road sections, and taking the foot corresponding to the minimum vertical distance of the track point in each candidate road section as a candidate point
Locus point p t All candidate points constitute a candidate point set->
Further finding out candidate point set C corresponding to all track points in road network t A set of candidate paths Γ;
(3) Calculating characteristics of the track points and the candidate paths; for any track tau, M track points are included, which are divided into u local tracks with length of M, and overlapping degree Ov is set for adjacent tracks i,i+1 For the alpha-th candidate path gamma in the candidate paths gamma α Including the locus point p t Candidate points
Distance feature of>
I.e. single point feature->
And two consecutive trace points p t-1 ,p t And candidate point +.>
Front and rear characteristics between->
Namely spatial characteristics->
Time characteristics->
Angular feature
Calculating traffic correlation characteristics of matched road sections corresponding to two track points according to the characterization vector of each road section>
Wherein t=1, 2, …, M; all features are connected to form a candidate path feature matrix LF= [ EF, PF]＝[DF,SF,TF,AF,CF]；
(4) Encoding candidate track features; calculating potential functions of each candidate path in the candidate path set Γ
I.e., a weighted sum of all candidate point features in the candidate path; select->
The largest top q' candidate paths, where the alpha path gamma α Feature matrix +.>
Coding with fully connected layer to obtain hidden layer feature of the path +.>
Wherein FC is a fully connected layer; the selected q' candidate path feature matrixes are subjected to full-connection coding to obtain hidden layer features corresponding to each candidate path +.>
Where h=[h DF ,h SF ,h TF ,h AF ,h CF ]The method comprises the steps of carrying out a first treatment on the surface of the The full-connection layer encodes the single-point features and the front and back features of all the candidate points to obtain hidden layer features;
(5) Inputting the hidden layer characteristics into an attention layer to obtain each characteristic weight; the processed hidden layer features are respectively used as a query term Q, a key term K and a value term V of the attention model, and the hidden layer features are formed by Att (Q, K) =softmax (QK) T ) Calculating the tensile dot product, taking the tensile dot product as a relation matrix of the relation strength between the characteristics of all hidden layers, and finally passing through the full-connection layer W a =fc (Att (Q, K)), where W a ＝[W DF ,W SF ,W TF ,W AF ,W CF ]I.e., the size of each feature weight;
(6) Constructing a conditional random field model API-CRF of self-adaptive weight; taking the distance characteristic DF as the emission probability, PF= [ SF, TF, AF, CF ]]As transition probabilities, weights W of the respective features are determined a Constructing a conditional random field CRF as the weight of each probability, and taking the average value of the weights of the two local tracks for the weight of the overlapping part of the two local tracks in the step (3); CRF is expressed as a sequence of tracks represented by a given sequence of tracks τ (p 1 ,p 2 ,…,p m ) Deducing the path sequence gamma (y 1 ,y 2 ,…,y m ) Probability of (2)
Wherein Z (τ) is a normalization factor, +.>
Can be written as +.>
Wherein->
For characteristic weights, LF t,k A kth feature represented as a kth trace point; finally, according to the characteristics between the track points and the candidate pointsThe method comprises the steps of obtaining features and weights of the features, solving an optimal inferred path by using a dynamic programming algorithm, namely, based on the weighted accumulation of the features of the track points of one track and the candidate points of the track as non-normalized probability, finding the maximum non-normalized probability and backtracking to find the corresponding candidate track as the optimal inferred path;
(7) Calculating a Loss function, backward propagating and updating a parameter set { theta }, stopping iteration until the Loss function reaches convergence, and otherwise, further updating { theta }; and inputting the track points to be predicted into the model for prediction after training is completed, and obtaining a predicted path.
2. The method for adaptive path inference of vehicle GPS tracks according to claim 1, wherein the step (3) calculates traffic correlation characteristics CF of road segments with matching two track points t,t-1 The method is specifically as follows:
(1) Representing the road relation model as G (v, epsilon), wherein v is the road segments between road network nodes N, epsilon is the topology and traffic correlation between the road segments E; the road context characteristics can obtain the dependence among road sections through the history track, and the shortest path set S among all road sections in a certain length in the road network is obtained s In order to express a topological correlation between the two road segments; at the same time from training set T trn Intercepting a result of matching the track points of the same road section to the road network, namely a path set S of vehicle running d Path set S d Implicit driving habits of drivers and road traffic conditions; s is S s And S is d Can represent the intensity of the traffic correlation between road sections, and S is as follows s And S is d The Skip-gram is used as a training set to train, the trained Skip-gram model can predict the characterization vectors of the front road section and the rear road section by using the characterization vector of the current road section, and the Skip-gram method is used for predicting each road section to obtain the characterization vector emb of each road section E in the road network;
(2) For the locus point p t-1 And p t Two candidate road segments corresponding to them
The traffic correlation characteristics between can be calculated using the following formula:
wherein the method comprises the steps of
Is p t-1 The j candidate point of the point is located on the road section, < ->
For point p t The j' th candidate point of (c) is located on the road section,
is->
Is characterized by vector>
Is->
Is described.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210831380.8A CN115206099B (en) | 2022-07-15 | 2022-07-15 | Self-adaptive path inference method for GPS track of vehicle |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210831380.8A CN115206099B (en) | 2022-07-15 | 2022-07-15 | Self-adaptive path inference method for GPS track of vehicle |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115206099A CN115206099A (en) | 2022-10-18 |
| CN115206099B true CN115206099B (en) | 2023-06-20 |
Family
ID=83581586
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210831380.8A Active CN115206099B (en) | 2022-07-15 | 2022-07-15 | Self-adaptive path inference method for GPS track of vehicle |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115206099B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115759386B (en) * | 2022-11-11 | 2023-07-07 | 中国民航科学技术研究院 | Method and device for predicting flight execution result of civil aviation flight and electronic equipment |
| CN116523433B (en) * | 2023-07-03 | 2023-09-01 | 常州唯实智能物联创新中心有限公司 | Four-way vehicle scheduling method and system based on bidirectional dynamic side weight |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110637213A (en) * | 2017-05-16 | 2019-12-31 | 北京嘀嘀无限科技发展有限公司 | System and method for digital path planning |
| CN111858817A (en) * | 2020-07-23 | 2020-10-30 | 中国石油大学(华东) | BilSTM-CRF path inference method for sparse track |
| WO2021183137A1 (en) * | 2020-03-13 | 2021-09-16 | Zenuity Ab | Methods and systems for vehicle path planning |
| WO2021206793A1 (en) * | 2020-04-06 | 2021-10-14 | B&H Licensing Inc. | Method and system for detecting jaywalking of vulnerable road users |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP2095148B8 (en) * | 2006-11-06 | 2012-12-12 | TomTom Global Content B.V. | Arrangement for and method of two dimensional and three dimensional precision location and orientation determination |
| US10663305B2 (en) * | 2018-07-16 | 2020-05-26 | Here Global B.V. | Map matched aggregation for K-anonymity in trajectory data |
-
2022
- 2022-07-15 CN CN202210831380.8A patent/CN115206099B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110637213A (en) * | 2017-05-16 | 2019-12-31 | 北京嘀嘀无限科技发展有限公司 | System and method for digital path planning |
| WO2021183137A1 (en) * | 2020-03-13 | 2021-09-16 | Zenuity Ab | Methods and systems for vehicle path planning |
| WO2021206793A1 (en) * | 2020-04-06 | 2021-10-14 | B&H Licensing Inc. | Method and system for detecting jaywalking of vulnerable road users |
| CN111858817A (en) * | 2020-07-23 | 2020-10-30 | 中国石油大学(华东) | BilSTM-CRF path inference method for sparse track |
Non-Patent Citations (1)
| Title |
|---|
| 于娟 等.高级地图匹配算法：研究现状和趋势.电子学报.2021,第49卷(第9期),全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115206099A (en) | 2022-10-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115206099B (en) | Self-adaptive path inference method for GPS track of vehicle | |
| CN110097755A (en) | Freeway traffic flow amount state identification method based on deep neural network | |
| CN112037506B (en) | Vehicle track prediction model construction method, vehicle track prediction method and system | |
| KR102313119B1 (en) | Learning method and learning device for attention-driven image segmentation by using at least one adaptive loss weight map to be used for updating hd maps required to satisfy level 4 of autonomous vehicles and testing method and testing device using the same | |
| CN114202120A (en) | Urban traffic travel time prediction method aiming at multi-source heterogeneous data | |
| CN113139446B (en) | End-to-end automatic driving behavior decision method, system and terminal equipment | |
| CN110047096B (en) | A kind of multi-object tracking method and system based on depth conditions random field models | |
| CN110444022A (en) | The construction method and device of traffic flow data analysis model | |
| CN112651374B (en) | Future trajectory prediction method based on social information and automatic driving system | |
| CN112632202B (en) | Dynamic map matching method based on high-order hidden Markov model | |
| CN115049009A (en) | Track prediction method based on semantic fusion representation | |
| CN112991750B (en) | Local traffic optimization method based on reinforcement learning and generation type countermeasure network | |
| US20230038673A1 (en) | Sequential pedestrian trajectory prediction using step attention for collision avoidance | |
| CN116186358B (en) | Depth track clustering method, system and storage medium | |
| CN114090718B (en) | Bi-LSTM prediction and fuzzy analysis based interrupted track correlation method | |
| CN114368387B (en) | Attention mechanism-based driver intention recognition and vehicle track prediction method | |
| Hecker et al. | Learning accurate and human-like driving using semantic maps and attention | |
| CN107179085A (en) | A kind of condition random field map-matching method towards sparse floating car data | |
| CN111695737A (en) | Group target advancing trend prediction method based on LSTM neural network | |
| CN113947928A (en) | Traffic signal lamp timing method based on combination of deep reinforcement learning and extended Kalman filtering | |
| CN117474168A (en) | Ship track prediction method and device of CNN-BiGRU based on attention mechanism | |
| CN114970058A (en) | Large-scale network signal control optimization method based on belief domain Bayes | |
| CN113537580A (en) | Public transport passenger flow prediction method and system based on adaptive graph learning | |
| CN114419884B (en) | Self-adaptive signal control method and system based on reinforcement learning and phase competition | |
| CN114219078A (en) | Neural network model interactive training method and device and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |