CN104408203A - Method for predicting path destination of moving object - Google Patents

Abstract

The invention discloses a method for predicting the path destination of a moving object, and mainly solves the problems of sparse data, a low hit rate and poor real-time performance during the prediction of the path destination in the prior art. The method includes the following steps: off-line calculating the history path data in a training data set to obtain a location set, a prior probability of each location, a single-step transition probability matrix and a comprehensive transition probability matrix; calculating the conditional probability and the posterior probability of each location being the destination based on the off-line calculated data; on-line predicting the destination of the path to be predicted based on the posterior probability. Compared with the existing path prediction method, the method provided by the invention has the advantages that a prediction can be performed even when the data in the training data set is sparse; the calculation process of the prediction is optimized, so that the impact of the non-aftereffect property of Markov chain on the prediction accuracy is reduced; the prediction accuracy is improved; the method can be used for pushing location-related targeted advertisements and deploying criminal arrest plans in advance.

Description

A kind of method predicting mobile object final on trajectory

Technical field

The invention belongs to computer realm, relate to the prediction in data mining and place, a kind of specifically method predicting mobile object final on trajectory, namely by carrying out analysis mining to obtained historical trajectory data, thus new track is carried out to the method for destination prediction, can be used for the directional advertisement sending relevant to place, in advance deployment criminal and arrest plan etc.

Background technology

Along with the development of technology of Internet of things, various sensor or to be embedded in portable mobile terminal, or participate in the daily routines of people to be fixedly mounted on the first-class form of communal facility.As GPS sensor in taxi and smart mobile phone etc.They can catch in real time and change is moved in the position of recording in people's daily life.Analysis mining is carried out to the historical trajectory data of these sensor collection, builds destination forecast model, be conducive to scientifically decision-making, improve the quality of living.As in commercial affairs, businessman can by carrying out directional advertisement sending to the destination prediction of client in advance.In law enforcement, police can by disposing to the trajectory predictions of criminal the plan of arresting, the concepts such as the intelligent city of nowadays rising in addition and Smart Home.

But due to the data often substantial amounts and value density is low that sensor produces, how from these mass datas, to extract interesting knowledge efficiently, as track frequent mode, place transition probability etc., and building and have the destination forecast model of high hit rate, is nowadays urgent problem.Now existing a lot of scholar is to this has been research.

Mainly following three major types can be divided into the forecast model of destination at present by historical trajectory data:

The first kind is the trajectory predictions model based on path matching.It mates with track in historical track storehouse mainly through track to be predicted, utilizes number of matches isometry index to find out the historical track of optimum matching, and then predicts the destination of track to be predicted.

Equations of The Second Kind method has mainly used the bayes predictive model in statistical learning.First historical trajectory data is integrated to the prior probability and conditional probability that obtain each place, then calculated the posterior probability in each place by Bayesian formula.

But, above-mentioned two kinds of methods easily run into the problem of Sparse in actual applications, although the historical trajectory data substantial amounts arrived by GPS sensor collection, the place transfer number of combinations of n intersite is n (n-1), and length is the number of combinations of the track of m is n (n-1) ^m-1, relative to the track combination number of exponential growth like this, although the substantial amounts of historical track, still can Sparse Problem be run into, thus the accuracy rate of impact prediction.

3rd class methods mainly utilize Markov chain model to carry out trajectory predictions, but use in Markov model, and when taking advantage of by the company of one-step transition probability matrix the multistep transition probability matrix asking intersite, the computation complexity of multiplication of matrices is O (n ³), computing cost sharply can rise along with the increase of matrix size, affects real-time performance and the precision of prediction of algorithm.

Also some scholar's join probability graph model has carried out expanding to above three kinds of basic skills and has improved, they introduce other based on probability graph model some affect the factor of destination, as weekend, season and weather etc. factor, to improve the predictablity rate of model.But the accuracy rate that these methods obtain promotes all based on specific application scenarios, and versatility is bad.

Summary of the invention

Technical matters to be solved by this invention is for the above-mentioned existing a kind of method providing prediction mobile object final on trajectory newly, to solve the Sparse Problem using bayes predictive model easily to meet with, reduce the computing time using Markov model simultaneously, improve the accuracy of destination prediction.

The present invention solves the problems of the technologies described above adopted technical scheme: a kind of method predicting mobile object final on trajectory, is characterized in that: comprise the steps:

Step 1, calculated off-line step:

1a), input training dataset H, this training dataset H are the set of the moving target initial trace represented with longitude and latitude sequence form collected in advance;

1b), input place grain size parameter n, region to be predicted is on average divided into the grid of n × n according to longitude and latitude, each grid represents the three unities, obtains place set L={l _i| i=1,2,3 ..., n ², L is gathered in place and stores in the form of a file;

1c), by the initial trace represented with longitude and latitude sequence form in training dataset H convert the track represented with place sequence form to, obtain historical trajectory data collection wherein represent i-th place in the track m in historical track set, k is the length of track;

1d), traversal history track data collection T, for place set L in each place l _i, statistical history track data is concentrated with l respectively _ifor the track data collection T of terminal _ithe quantity of middle track | T _i|, calculate each place l _ibecome the prior probability of final on trajectory and by this prior probability P (l _i) store in the form of a file;

1e), the track in historical trajectory data collection T is split into place pair obtain place to S set;

Gather all places in L 1f), over the ground, add up it from i-th place to the place in a jth place to <l _i, l _jthe number of times t that > occurs in place is to S set _i,j, calculate place l _ito place l _jone-step transition probability obtain the one-step transition probability matrix M of intersite, wherein t _{i, *}represent and gather the place of anywhere L to the number of times occurred in place is to S set from i-th place to place, one-step transition probability matrix M is stored in the form of a file;

1g), set the maximum scale-up factor c>1 that detours, utilize and calculate aided solving matrix array TM based on Markov chain method;

1h), according to aided solving matrix array TM, calculate comprehensive transition probability matrix M ^final, and store in the form of a file;

Step 2, on-line prediction step:

2a), place set L, prior probability P (l that calculated off-line step obtains is loaded into _i), one-step transition probability matrix M and comprehensive transition probability matrix M ^final;

2b), track to be predicted is inputted wherein represent i-th place in track tra to be predicted, d is the length of track tra to be predicted, d>=1, and d is integer;

2c), calculate with the place l in place set L _ifor in the track of terminal, the conditional probability containing track tra to be predicted: $P\left(\mathrm{tra}\right|l_i)=\frac{\underset{j=1}{\overset{d-1}{\mathrm{\&Pi;}}}{M}_{{\mathrm{tra}}_j,{\mathrm{tra}}_{j+1}}\&CenterDot;M_{{\mathrm{tra}}_d,{\mathrm{tra}}_i}^{\mathrm{final}}}{M_{{\mathrm{tra}}_1,{\mathrm{tra}}_i}^{\mathrm{final}}};$

In above formula, represent be in one-step transition probability matrix M under be designated as tra _jand tra _j+1element; that represent is comprehensive transition probability matrix M ^finalin under be designated as tra _dand tra _ielement; that represent is comprehensive transition probability matrix M ^finalin under be designated as tra ₁and tra _ielement;

2d), each place l in place set L is calculated according to Bayesian formula _ibecome the posterior probability of the terminal of track tra to be predicted and the posterior probability P (l that it becomes the terminal of track tra to be predicted is pressed on the ground gathered in L by place _i| tra) size descending sort;

2e), in step 2d) obtain getting the destination as prediction, a place coming above in the place set L after sorting, wherein a needs the ground predicted to count out.

As improvement, described step 1g) in, utilize and carry out as follows based on Markov chain method calculating aided solving matrix array TM:

1g.1), r matrix T M [r]=M in aided solving matrix array TM is established ^r, it represents that the r of intersite walks transition probability matrix, and r is integer, and its span is:

1g.2), p=2 is made;

1g.3), TM [p]=TM [p-1]+TM [p] is calculated;

1g.4), the value of p is made to increase by 1, if then return step 1g.3); Otherwise, perform step 1h).

Improve again, described step 1h) in, calculate comprehensive transition probability matrix M ^finalcarry out as follows:

1h.1), create an empty effective place to list List, empty place to a list mapping set Map and empty interim place to list pList;

1h.2), check single step transition probability matrix M, if all elements of all elements of this one-step transition probability matrix M i-th row and jth row is 0, then by place to <l _i, l _j> joins effective place in list List, wherein i≤n, j≤n;

1h.3), by effective place to all places in list List to the manhatton distance of place centering two intersite for key assignments, join place to the place in list mapping set Map corresponding to this key assignments in list;

1h.4), step=1 is made;

1h.5), interim place to list pList be place to key assignments in list mapping set Map be the place of step to list, if this list be sky, calculate interim aided solving matrix make interim place to the place in list pList to <l _i, l _jcomprehensive transition probability corresponding to > be that the place queue table of step is left out from place list mapping set Map by key assignments;

1h.6), judge whether place is empty to list mapping set Map: if for empty, terminate to calculate, otherwise, then make the value of step increase by 1, return step 1h.5).

Compared with prior art, the invention has the advantages that:

1, the present invention carries out trajectory predictions in conjunction with Markov chain and bayesian theory, solves on the one hand and only adopts bayesian theory to carry out the Sparse Problem run into when destination is predicted, makes it more can prediction in the insufficient situation of adaptation training data; On the other hand, optimize the computing time of multistep transition probability matrix, and calculate in conjunction with Bayesian model the probability that place becomes destination, decrease the impact of markovian markov property on predictablity rate, compared with the method using traditional Markov chain to carry out predicting, it is more accurate to predict;

2, the present invention is by being divided into calculated off-line and on-line prediction two parts, the operations such as matrix multiplication large for expense computing time is placed on line lower part and carries out, improve the real-time response ability of prediction.

Accompanying drawing explanation

Fig. 1 be the method predicting mobile object final on trajectory in the embodiment of the present invention realize general flow chart;

Fig. 2 is calculated off-line sub-process figure in the embodiment of the present invention;

Fig. 3 is on-line prediction sub-process figure in the embodiment of the present invention.

Embodiment

Below in conjunction with accompanying drawing embodiment, the present invention is described in further detail.

The invention provides a kind of method predicting mobile object final on trajectory, be described in detail below for Beijing's taxi track data collection, this taxi track data collection is from the Geolife project of Microsoft Research, Asia: its project home page network address is: http:// research.microsoft.com/en-us/projects/geolife/default.as px, 182 taxis of this data set record Beijing from July, 2007 in August, 2012 by a series of GPS initial trace information represented with the place that longitude and latitude represents.Whole data set comprises 17621 tracks altogether, and track length reaches 1292951 kms.

Above-mentioned Beijing taxi track data collection is divided into training dataset H and test data set TH by this example at random, and wherein in training dataset H and test data set TH, the ratio of tracking quantity is 9:1.Using the true destination of the terminal of the track of each in test data set TH as this track, then remove the terminal of this track, build track collection PH to be predicted with the track obtained, the track treated in prediction locus collection PH carries out destination prediction.

With reference to Fig. 1, performing step of the present invention is as follows:

Step 1: input training dataset H;

Step 2: carry out calculated off-line to the historical trajectory data in the training dataset H of input, obtains place set L, each place prior probability P (l _i), one-step transition probability matrix M and comprehensive transition probability matrix M ^final, wherein l _ifor i-th place in place set L;

With reference to Fig. 2, being implemented as follows of this step:

2a), input place grain size parameter is n=30, and Beijing's map partitioning is 900 places by the grid with 30 × 30, obtains place and gathers L={l _i| i=1,2,3 ..., n ², and store with document form;

2b), by the initial trace represented with longitude and latitude sequence form in training dataset H convert the track represented with place sequence form to, obtain the historical trajectory data collection containing 18091 tracks wherein represent i-th place in the track m in historical trajectory data collection T, k is the length of track;

2c), traversal history track data collection T, for place set L in each place l _i, statistical history track is concentrated with l respectively _ifor the track data collection T of terminal _ithe quantity of middle track | T _i|, calculate each place l _ibecome the prior probability of final on trajectory than and by this prior probability P (l _i) store in the form of a file;

2d), the track in historical trajectory data collection T is split into place pair obtain place to S set;

Gather all places in L 2e), over the ground, add up it from i-th place to the place in a jth place to <l _i, l _jthe number of times t that > occurs in place is to S set _i,j, calculate place l _ito place l _jone-step transition probability obtain the one-step transition probability matrix M of intersite, wherein t _{i, *}represent that the place of anywhere from i-th place to L is to the number of times occurred in place is to S set, stores in the form of a file by one-step transition probability matrix M;

2f) set the maximum scale-up factor c>1 that detours, utilize and calculate aided solving matrix array TM based on Markov chain method:

2f.1), r matrix: TM [r]=M in aided solving matrix array TM is established ^r, it represents that the r of intersite walks transition probability matrix, and r is integer, and its span is:

2f.2), p=2 is made;

2f.3), TM [p]=TM [p-1]+TM [p] is calculated;

2f.4), the value of p is made to increase by 1, if then return step 2f.3); Otherwise, perform step 2g);

2g), according to aided solving matrix array TM, calculate comprehensive transition probability matrix M ^final, and store in the form of a file:

2g.1), create an empty effective place to list List, empty place to a list mapping set Map and empty interim place to list pList;

2g.2), check single step transition probability matrix M, if all elements of all elements of this one-step transition probability matrix i-th row and jth row is 0, then by place to <l _i, l _j> joins effective place in list List, wherein i≤n, j≤n;

2g.3), by effective place to all places in list List to the manhatton distance of place centering two intersite for key assignments, join place to the place in list mapping set Map corresponding to this key assignments in list;

2g.4), step=1 is made;

2g.5), make interim place to list pList be place to key assignments in list mapping set Map be the place of step to list, if this list is not empty, calculate interim aided solving matrix make interim place to the place in list pList to <l _i, l _jcomprehensive transition probability corresponding to > be that the place queue table of step is left out from place list mapping set Map by key assignments;

2g.6), judge whether place is empty to list mapping set Map: if for empty, terminate to calculate, otherwise, then make the value of step increase by 1, return step 2g.5);

Step 3: input track collection PH to be predicted;

Step 4: on-line prediction is carried out to the track collection PH to be predicted of input, obtains the prediction destination of the track in track collection PH to be predicted:

With reference to Fig. 3, being implemented as follows of this step:

4a), the prior probability P (l in L, each place is gathered in the place that importing calculated off-line part obtains _i), one-step transition probability matrix M and comprehensive transition probability matrix M ^final;

4b), input track collection PH to be predicted, convert the initial trace represented with longitude and latitude sequence form in data set PH to be predicted to represent with place sequence form track, obtain containing 2025 track collection PH to be predicted;

4c), a track to be predicted in track collection PH to be predicted is got wherein represent i-th place in the track tra in track set PH to be predicted, d is the length of track, d>=1, and d is integer;

4d), calculate with the place l in place set L _ifor in the track of terminal, containing with track tra to be predicted conditional probability: $P\left(\mathrm{tra}\right|l_i)=\frac{\underset{j=1}{\overset{d-1}{\mathrm{\&Pi;}}}{M}_{{\mathrm{tra}}_d,{\mathrm{tra}}_{d+1}}\&CenterDot;M_{{\mathrm{tra}}_k,{\mathrm{tra}}_i}^{\mathrm{final}}}{M_{{\mathrm{tra}}_1,{\mathrm{tra}}_i}^{\mathrm{final}}};$

In above formula, represent be in one-step transition probability matrix M under be designated as tra _jand tra _j+1element; that represent is comprehensive transition probability matrix M ^finalin under be designated as tra _dand tra _ielement; that represent is comprehensive transition probability matrix M ^finalin under be designated as tra ₁and tra _ielement;

4e), each place l in place set L is calculated according to Bayesian formula _ibecome the posterior probability of the terminal of track tra to be predicted $P\left(l_i\right|\mathrm{tra})=\frac{P\left(\mathrm{tra}\right|l_i)\&CenterDot;P\left(l_i\right)}{\underset{l_j\&Element;L}{\mathrm{\&Sigma;}}P\left(\mathrm{tra}\right|l_j)\&CenterDot;P\left(l_j\right)};$

4f), by place, the posterior probability P (l that it becomes the terminal of track tra to be predicted is pressed on the ground gathered in L _i| tra) size descending sort, count out a on the ground predicted as required, gets front a place in the set of place as the prediction destination of track tra to be predicted, and record the destination of this trajectory predictions tra to be measured;

4g), track tra to be predicted is deleted from track collection PH to be predicted;

4h), judge whether track collection PH to be predicted is empty: if track collection PH to be predicted is not for empty, then return step 4c); Otherwise, terminate the on-line prediction treating prediction locus collection PH, perform step 5;

Step 5: the prediction destination exporting the track each to be predicted of record in step 4, prediction terminates.

By above step, obtain the prediction destination of track in track collection PH to be predicted, finally the prediction destination of track in track collection PH to be predicted and true destination are compared, add up the prediction hit rate of this forecast model, learn that this forecast model can dope the destination of the track to be predicted concentrating extraction from Beijing's taxi track data exactly.

Claims (3)

1. predict a method for mobile object final on trajectory, it is characterized in that: comprise the steps:

Step 1, calculated off-line step:

1a), input training dataset H, this training dataset H are the set of the moving target initial trace represented with longitude and latitude sequence form collected in advance;

1b), input place grain size parameter n, region to be predicted is on average divided into the grid of n × n according to longitude and latitude, each grid represents the three unities, obtains place set L={l _i| i=1,2,3 ..., n ², L is gathered in place and stores in the form of a file;

1c), by the initial trace represented with longitude and latitude sequence form in training dataset H convert the track represented with place sequence form to, obtain historical trajectory data collection $T=\{<l_{m_1},l_{m_2},l_{m_3},...,l_{m_i},...,l_{m_k}>|l_{m_i}\&Element;L,k\&GreaterEqual;2\},$ Wherein represent i-th place in the track m in historical track set, k is the length of track;

1d), traversal history track data collection T, for place set L in each place l _i, statistical history track data is concentrated with l respectively _ifor the track data collection T of terminal _ithe quantity of middle track | T _i|, calculate each place l _ibecome the prior probability of final on trajectory and by this prior probability P (l _i) store in the form of a file;

1e), the track in historical trajectory data collection T is split into place pair obtain place to S set;

Gather all places in L 1f), over the ground, add up it from i-th place to the place in a jth place to <l _i, l _jthe number of times t that > occurs in place is to S set _i,j, calculate place l _ito place l _jone-step transition probability obtain the one-step transition probability matrix M of intersite, wherein t _{i, *}represent and gather the place of anywhere L to the number of times occurred in place is to S set from i-th place to place, one-step transition probability matrix M is stored in the form of a file;

1g), set the maximum scale-up factor c>1 that detours, utilize and calculate aided solving matrix array TM based on Markov chain method;

1h), according to aided solving matrix array TM, calculate comprehensive transition probability matrix M ^final, and store in the form of a file;

Step 2, on-line prediction step:

2a), place set L, prior probability P (l that calculated off-line step obtains is loaded into _i), one-step transition probability matrix M and comprehensive transition probability matrix M ^final;

2b), track to be predicted is inputted wherein represent i-th place in track tra to be predicted, d is the length of track tra to be predicted, d>=1, and d is integer;

2c), calculate with the place l in place set L _ifor in the track of terminal, the conditional probability containing track tra to be predicted: $P\left(\mathrm{tra}\right|l_i)=\frac{\underset{j=1}{\overset{d-1}{\mathrm{\&Pi;}}}{M}_{\mathrm{tr}{a}_j,{\mathrm{tra}}_{j+1}}\&CenterDot;M_{{\mathrm{tra}}_d,{\mathrm{tra}}_i}^{\mathrm{final}}}{M_{{\mathrm{tra}}_1,{\mathrm{tra}}_i}^{\mathrm{final}}};$

In above formula, represent be in one-step transition probability matrix M under be designated as tra _jand tra _j+1element; that represent is comprehensive transition probability matrix M ^finalin under be designated as tra _dand tra _ielement; that represent is comprehensive transition probability matrix M ^finalin under be designated as tra ₁and tra _ielement;

2d), each place l in place set L is calculated according to Bayesian formula _ibecome the posterior probability of the terminal of track tra to be predicted and the posterior probability P (l that it becomes the terminal of track tra to be predicted is pressed on the ground gathered in L by place _i| tra) size descending sort;

2e), in step 2d) obtain getting the destination as prediction, a place coming above in the place set L after sorting, wherein a needs the ground predicted to count out.

2. the method for prediction mobile object final on trajectory according to claim 1, is characterized in that: described step 1g) in, utilize and carry out as follows based on Markov chain method calculating aided solving matrix array TM:

1g.1), r matrix T M [r]=M in aided solving matrix array TM is established ^r, it represents that the r of intersite walks transition probability matrix, and r is integer, and its span is:

1g.2), p=2 is made;

1g.3), TM [p]=TM [p-1]+TM [p] is calculated;

1g.4), the value of p is made to increase by 1, if then return step 1g.3); Otherwise, perform step 1h).

3. the method for prediction mobile object final on trajectory according to claim 1, is characterized in that: described step 1h) in, calculate comprehensive transition probability matrix M ^finalcarry out as follows:

1h.1), create an empty effective place to list List, empty place to a list mapping set Map and empty interim place to list pList;

1h.2), check single step transition probability matrix M, if all elements of all elements of this one-step transition probability matrix M i-th row and jth row is 0, then by place to <l _i, l _j> joins effective place in list List, wherein i≤n, j≤n;

1h.3), by effective place to all places in list List to the manhatton distance of place centering two intersite for key assignments, join place to the place in list mapping set Map corresponding to this key assignments in list;

1h.4), step=1 is made;

1h.5), interim place to list pList be place to key assignments in list mapping set Map be the place of step to list, if this list be sky, calculate interim aided solving matrix make interim place to the place in list pList to <l _i, l _jcomprehensive transition probability corresponding to > be that the place queue table of step is left out from place list mapping set Map by key assignments;

1h.6), judge whether place is empty to list mapping set Map: if for empty, terminate to calculate, otherwise, then make the value of step increase by 1, return step 1h.5).