CN110223515B

CN110223515B - Vehicle track generation method

Info

Publication number: CN110223515B
Application number: CN201910522534.3A
Authority: CN
Inventors: 沃天宇; 张炳振; 陈凯恒
Original assignee: Beihang University
Current assignee: Beihang University
Priority date: 2019-06-17
Filing date: 2019-06-17
Publication date: 2021-01-01
Anticipated expiration: 2039-06-17
Also published as: CN110223515A

Abstract

The invention provides a vehicle track generation method, which comprises a step 1 of data processing, wherein the data processing comprises the steps of firstly preprocessing track data and map data; step 2, a data generation model stage, wherein the data generation model stage comprises a road section track generation model and a travel track generation model; and 3, a data set generation stage, wherein the data set generation stage is used for obtaining track data by loading the road section track generation model and the travel track generation model.

Description

Vehicle track generation method

Technical Field

The invention relates to a track generation method, in particular to a vehicle track generation method.

Background

With the rapid development of the internet technology and the technological progress of the internet of things, the living standard of people is improved, and meanwhile, a new definition is provided for quality life. The quality life needs to meet the requirements of comfort and convenience, and also needs to be intelligentized and scientific. From the original internet to the now well-established internet of things, there is an increasing incredible but real change in people's lives. The Internet of vehicles is a typical application scene of the technology of the Internet of things in the aspect of realizing intelligent transportation, and aims to construct an intelligent transportation network. The main concerns of the internet of vehicles include data acquisition, data calculation processing at the cloud, offline data mining, development and implementation of a data management system and the like. The data acquisition is the basis of all problem researches, and how to acquire a large amount of real and effective data becomes the first problem faced by researchers.

Since vehicle trajectory data can involve privacy concerns for the driver and passengers, data desensitization is done to the vehicle trajectory data at the time of publication. This may result in some processed data with insufficient details, for example, some public data directly omit the travel path of the vehicle, and only the time and place for getting on or off the vehicle are reserved. This data cannot be used to study problems with streaming data processing systems, spatio-temporal data indexing systems, and trajectory compression. On the one hand, user privacy is protected, and on the other hand, sufficient data characteristics are required for research. One method for solving the contradiction is to invent a vehicle track data generation method. The vehicle track can be freely generated, and the generated data can have similar characteristics with the real data to a certain extent.

In the prior art, computer simulation software and a specific vehicle driving model are mainly used for setting some traffic state parameters to realize traffic simulation. The traffic simulation software takes a vehicle as a description unit, and can more truly describe the following, overtaking, lane changing and other behaviors of the vehicle on the road. The important parameters are the speed and position of each description unit, and furthermore the vehicle type, size, average speed, acceleration etc. can be configured. The main functions of the traffic simulation software are to perform researches on route selection, signal lamp algorithm, simulated vehicle communication, traffic management strategy and the like. But has not been suitable for research in the internet of vehicles scenario. On one hand, traditional traffic simulation software hardly utilizes real historical data to design a vehicle model; on the other hand, the simulated scene of the vehicle trajectory data includes not only the traveling condition of one vehicle, but there may be a large number of vehicles traveling on the road network at the same time. It is therefore desirable to invent a new vehicle trajectory data generation system.

The prior art trajectory generation technique has the following disadvantages:

1. the detailed information of the track data is lost, and some important attributes are hidden

2. Historical real data is not utilized. Cannot reflect the real traffic characteristics.

3. The track mode generated by the prior art is single, only the running condition of a single vehicle is considered, and the overall characteristics of track data of a plurality of vehicles are not considered.

4. The characteristic engineering method for extracting the data inclusion characteristics wastes time and labor, has many subjective factors, and cannot fully reflect the data inclusion characteristics.

Disclosure of Invention

In order to solve the problems, the invention provides a vehicle track generation method based on GAN, which comprises a step 1, a data processing stage, a map data processing stage and a track data processing stage, wherein the data processing stage is to pre-process track data and map data; step 2, a data generation model stage, wherein the data generation model stage comprises a road section track generation model and a travel track generation model; and 3, a data set generation stage, wherein the data set generation stage is used for obtaining track data by loading the road section track generation model and the travel track generation model.

Different from the traditional traffic model which is manually defined and based on the physical law, the method realizes the generation of the vehicle track by a data-driven method. The invention takes historical track data and map data as input, and can finally generate a large number of complete tracks through data preprocessing, map matching, road section simulation and travel simulation. The input to this device is a number of vehicles, a number of times, a whole city-wide trajectory. By utilizing the historical track data and the strong learning and fitting capability of the neural network, the finally generated data also has certain relevance. Such data can represent traffic characteristic information of cities, and further meet related research of Internet of vehicles. The generation of the track which can meet the research requirement is realized.

Drawings

FIG. 1 is an overall flow chart of the present invention for vehicle trajectory generation;

FIG. 2 is a diagram of a generator and a judger structure of a road segment trajectory model according to the present invention.

Fig. 3 shows the structure of the travel path model according to the present invention.

FIG. 4 is a diagram of a generator and determiner structure of a travel trajectory model according to the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

The invention provides a vehicle track generation method, which comprises a step 1 and a data processing stage as shown in figure 1, wherein the data processing stage is to firstly preprocess track data and map data; step 2, a data generation model stage, wherein the data generation model stage comprises a road section track generation model and a travel track generation model; and 3, a data set generation stage, wherein the data set generation stage is used for obtaining track data by loading the road section track generation model and the travel track generation model.

The track point is defined as a set of values denoted as P (project point) in terms of time stamp, longitude, and latitude. A track consists of a number of track points ordered by timestamp denoted TR (Tracjectory). The trajectory is therefore represented as:

P＝<t,lng,lat>

TR ═ P1, P2, …, Pn >, P1.t < P2.t < P3.t < … < pn.t, where pi.t denotes the timestamp t of the ith trace point Pi.

The road can be divided into segments according to the map road network data, and the segments are denoted as RS (route segment). A complete trajectory can therefore be seen as a collection of trajectories on road segments connected to each other.

The preprocessing comprises the steps of normalizing the track data, carrying out map matching by using the map data and the track data, then carrying out road network graph embedding by using the map road network data, and obtaining a vector sequence of an input travel track generation model according to the results of the map matching and the road network graph trapping.

The road network Map data uses Open source Street Map (OSM) and contains longitude and latitude information of important roads in cities, and each road is uniquely represented by an ID number. And taking the GPS track data and the map data as input, and outputting the road ID corresponding to each GPS point by a map matching algorithm.

The road network Graph is a Graph structure (Graph) abstracted from roads and intersections. Intersections are represented as nodes (nodes) in the graph, and irregular road relationships such as roads representing edges (edges) connecting the nodes can be represented by a road network graph. Then, using Graph Embedding (Graph Embedding) algorithm, the map data is used as input, and each node in the road network can be represented as a 64-dimensional dense vector through output. In map matching algorithms, original track points have been matched to road segments of the road network. Thus each track point has a unique road ID corresponding to it. A mapping between the link ID to a 64-dimensional vector is obtained.

In order to make the neural network training more stable, the track longitude and latitude needs to be normalized. The method comprises the steps of taking longitude and latitude lng1, lat1, lng2 and lat2 of two nodes at the head and the tail of each road section, and taking a maximum value and a minimum value lng _ max, lat _ max, lng _ min and lat _ min. The longitude and latitude of a track point on a road section are set as lng _ x and lat _ x, and are normalized into

lng_norm＝(lng_x–lng_min)/(lng_max-lng_min)

lat_norm＝(lat_x–lat_min)/(lat_max-lat_min)

Therefore, the longitude and latitude on the road section are converted into decimal between 0 and 1. And selecting all data containing the road section in the historical track data to form a plurality of input samples of the GAN network.

A trip is defined as an order for the vehicle (e.g., from place to place). A complete trajectory may represent the trip. One track is formed by combining tracks on a plurality of road sections. Therefore, a map matching algorithm PRESS system is applied to process the historical track data. And inputting road data and track data, and outputting to obtain each track point and the corresponding road ID.

The data generation model stage of the invention generates a countermeasure network, and the countermeasure network stage comprises a road section track generation model and a travel track generation model, which are as follows.

The link trajectory generation model and the travel trajectory model include a Generator (Generator) and a Discriminator (Discriminator). The generator takes random noise as input and learns to generate data approximate to the target sample distribution. The discriminator aims to discriminate whether the sample is real data or generated data, and if it is real data, the discriminator outputs a high score, whereas it outputs a low score. And then guiding two sub parts to learn through back propagation, and adjusting network parameters by the generator according to the result given by the discriminator to enable the next generated data to be closer to the real data. Meanwhile, the discriminator can distinguish the real sample from the generated sample as much as possible. The final generator generates data having characteristics that are close to those of the real data.

In the road section track generation model, track points of a road section comprise a timestamp and longitude and latitude, so that track data on the road section belong to time sequence data. The trajectory generation means selects a long-short memory (LSTM) network as the generator for generating the countermeasure network. The LSTM structure contains three "gate" structures: an input gate, a forgetting gate and an output gate. The gate structure may select the information entered, passing a portion through the structure, and selecting valid information in the data. The input gate selects the input information of the current state; the forgetting gate function is to forget information which is not used before; the output gate is used for selecting information to be reserved according to the states of the first two gates and generating a new state, namely the output at the current moment. LSTM is better performing in speech recognition, text translation, etc. tasks, so the structure is chosen to handle the trajectory data. The structure of the generator and the judger of the road section track model is shown in fig. 2, and the road section track generation model adopts a one-dimensional LSTM network as the generator. The design of the discriminator adopts a standard convolutional neural network CNN and a fully-connected neural network, and the number of the hidden units of the last layer of the fully-connected layer is 1, which is used for expressing the probability that the input data is real data. The function of the discriminator is to give the probability that the data is true data for the input data. And after training is finished, model parameters of the generator are stored, and road section track data can be obtained in batches according to requirements.

The travel track generation model is input as a vector sequence with uncertain length and used for representing a road section sequence passed by the travel. The overall structure of the model is shown in figure 3. While the vector is a 64-dimensional matrix. The output is also a high-dimensional sequence of vectors of varying length. Since the sequence of links traversed by the trip is arranged in chronological order, the input sequence of vectors is also a time-series structure. The travel track generation model generator adopts a double-layer dynamic _ rnn structure. The variable length structure can dynamically calculate the input effective information according to different input lengths. The processing procedure is to find the input sequence with the longest length and calculate the longest length. And other sequences smaller than the maximum length are padded with 0 s. The newly filled 0 does not affect the subsequent computation of the neural network. The discriminator adopts a multi-layer CNN and full connection layer structure. The travel path generation model generator and arbiter structure is shown in fig. 4.

After the model is trained, the generator parameters are saved. And generating a vector sequence A ═ a1, a2 and … an by using a stroke generating program. And then carrying out vector inverse mapping to map the vector sequence to the road ID sequence. In the pre-processed graph embedding step, each road may be represented by a 64-dimensional vector. The n-path vector sequences are successively selected and expressed as Bi ═ Bi1, Bi2, …, bin ] (1 ═ i ≦ m) as candidate runs, and m is a positive integer. The vector inverse mapping is to calculate the spatial distance of the generated vector sequence a to the m candidate run sequences. The distance formula is:

d＝||A-Bi||(1<＝i<＝m)

the vector sequence Bi with the smallest spatial distance is the closest sequence to a, and therefore the vector sequence generated by the run can be considered to be Bi. Bi corresponds to the road ID sequence path [ ID1, ID2, …, idn ], i.e., the route generation model generation road ID sequence is path. The ID sequence represents one complete trip generated.

And generating a road ID sequence obtained by the model according to the track points on each road obtained by the road section track generation model and the travel track. Finally, a complete travel track sequence is obtained.

The invention can generate a large amount of vehicle track data which can reflect city characteristics similar to historical data. And after the real track data are preprocessed, inputting the preprocessed real track data into a generation countermeasure network. And generating a confrontation network, training and adjusting parameters, and storing the model. And then loading the model and calling a generator for generating the countermeasure network to obtain the generated track data in batch. After the model is trained once, the saved model can be used for multiple times, and the track can be quickly obtained. And the characteristics contained in the track data do not need to be extracted manually. The mode of combining the microcosmic (single road section) and the macroscopic (whole travel) can furthest reserve the effective information in the track and reduce the complexity of model learning.

Finally, it should be noted that: the above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.

Claims

1. A vehicle track generation method is characterized by comprising a step 1 of data preprocessing, wherein the data preprocessing is to preprocess track data and map data; step 2, a data generation model stage, wherein the data generation model stage comprises a road section track generation model and a travel track generation model; step 3, a data set generation stage, wherein the data set generation stage is used for obtaining track data by loading the road section track generation model and the travel track generation model; the preprocessing comprises the steps of normalizing track data, carrying out map matching by using the map data and the track data, then carrying out road network graph embedding by using the map road network data, and obtaining a vector sequence of an input travel track generation model according to the results of the map matching and the road network graph embedding; the normalization processing mode is as follows: the latitude is normalized as follows: lng _ norm (lng _ x-lng _ min)/(lng _ max-lng _ min), and longitude is normalized to: lat _ norm is (lat _ x-lat _ min)/(lat _ max-lat _ min), wherein long _ x and lat _ x are longitude and latitude of track points on the road sections, lat _ max and lat _ min are the maximum value and the minimum value of the longitude of the head node and the tail node of each road section, and long _ max and long _ min are the maximum value and the minimum value of the latitude; the road section track generation model comprises a road section track generation model generator and a road section track model discriminator; the road section track generation model generator adopts a one-dimensional LSTM network, the LSTM structure comprises an input gate, a forgetting gate and an output gate, the input gate selects input information of the current state, the forgetting gate removes useless information before, the output gate selects information to be kept and simultaneously generates new state, namely output at the current moment according to the states of the previous two gates, the road section track model discriminator uses a standard Convolutional Neural Network (CNN) and a fully-connected neural network, the number of the hidden units of the last layer of the fully-connected layer is 1, the hidden units are used for representing the probability that the input data are real data, and road section track data are obtained in batches according to needs; the stroke track generation model comprises a stroke track generation model generator and a stroke track model discriminator, wherein the stroke track model generator adopts a double-layer dynamic _ rnn structure, the double-layer dynamic _ rnn structure dynamically calculates input effective information according to input different lengths, the dynamic calculation process comprises the steps of firstly finding an input sequence with the longest length, calculating the maximum length, then filling other sequences with 0 which is smaller than the maximum length, the newly filled 0 does not influence the subsequent calculation of the neural network, and the stroke track model discriminator adopts a multi-layer CNN and a full connection layer structure.

2. The method according to claim 1, wherein the data set is generated in a specific manner by generating a vector sequence a by using a travel generation program, then performing vector inverse mapping to map the vector sequence to a road ID sequence, and then obtaining a track sequence of a complete travel according to a track point on each road obtained by a road section track generation model and the road ID sequence obtained by the travel track generation model, wherein the vector inverse mapping is performed in a manner of calculating spatial distances between the travel generation model and the vector sequence a to m candidate travel sequences, and the vector sequence with the smallest spatial distance corresponds to the road ID sequence to generate the road ID sequence for the travel generation model.