CN112948715A - Vehicle classification method based on short-time GPS track data - Google Patents

Vehicle classification method based on short-time GPS track data Download PDF

Info

Publication number
CN112948715A
CN112948715A CN202110228346.7A CN202110228346A CN112948715A CN 112948715 A CN112948715 A CN 112948715A CN 202110228346 A CN202110228346 A CN 202110228346A CN 112948715 A CN112948715 A CN 112948715A
Authority
CN
China
Prior art keywords
data
vehicle
gps
short
time
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.)
Pending
Application number
CN202110228346.7A
Other languages
Chinese (zh)
Inventor
蒋鹏
姚知涵
刘俊
胡华
许欢
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hangzhou Dianzi University
Original Assignee
Hangzhou Dianzi University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Hangzhou Dianzi University filed Critical Hangzhou Dianzi University
Priority to CN202110228346.7A priority Critical patent/CN112948715A/en
Publication of CN112948715A publication Critical patent/CN112948715A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9537—Spatial or temporal dependent retrieval, e.g. spatiotemporal queries
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Databases & Information Systems (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Traffic Control Systems (AREA)

Abstract

The invention relates to a vehicle classification method based on short-time GPS track data. The invention firstly preprocesses original GPS data to obtain GPS data representation for network input, secondly adopts a resampling technology to rebalance a sample space of an unbalanced data set to solve the situation of vehicle type distribution imbalance existing in a real road network, and finally develops a deep learning network model combining multi-level discrete wavelet decomposition and a bidirectional LSTM network for vehicle classification aiming at the time sequence and short-time high-frequency characteristics of a GPS data source, and can complete classification tasks after hyper-parameter selection and model training. The invention not only fully considers the privacy of the GPS data, but also provides a feasible solution for the imbalance of vehicle type distribution in a real road network, and the invention can effectively capture the depth characteristics contained in the GPS track data, thereby achieving higher classification accuracy.

Description

Vehicle classification method based on short-time GPS track data
Technical Field
The invention relates to a vehicle classification method, in particular to a vehicle classification method based on short-time GPS track data.
Background
In recent years, with the continuous acceleration of urbanization, a large amount of mobile pollution sources of motor vehicles with large conservation, fast acceleration and wide flowing range can discharge a large amount of pollutants into urban atmospheric environment, and the health of people and the production and life of cities are seriously influenced. Therefore, emission monitoring of mobile pollution sources has become a key to reducing urban atmospheric pollution levels and improving urban air quality. The method for measuring the concentration of pollutants in the tail gas of the mobile pollution source is a precondition for monitoring the emission of the mobile pollution source, and the method has the strongest feasibility for simulating the emission of pollutants of the motor vehicle due to the cost advantage by using a software model. For the simulation models, the motion trajectory data (such as driving speed, acceleration, driving mileage and the like) of the mobile pollution source of the motor vehicle and the type of the motor vehicle are important input parameters, and the accurate acquisition of the parameters has important significance for improving the accuracy of model output. The vehicle track data is often supported by a Global Positioning System (GPS), but a big existing challenge is lack of vehicle type information of the vehicle recording journey, so that obtaining the vehicle type information from the vehicle GPS track data is of great significance for emission monitoring of mobile pollution sources.
Existing motor vehicle classification methods are typically based on fixed point sensor data or GPS data. However, the fixed point sensor often has the defects of high installation cost, sparse deployment density, normal traffic interference and the like; it is difficult to classify motor vehicles in large-scale urban road networks using data acquired by fixed-point sensors. Compared with fixed point sensor data, the GPS data has the advantages of low acquisition cost, high deployment density, no disturbance to traffic and the like; therefore, it is necessary to develop a vehicle classification method based on GPS data, which classifies vehicles in a road network into three types of motorcycles, light vehicles, and heavy vehicles according to the 13-type vehicle classification standards set by the federal highway administration (FHWA), i.e., FHWA1, FHWA2-4, and FHWA8-13, respectively, and can provide valuable information for vehicle pollution source emission monitoring.
At present, vehicle type classification methods based on GPS trajectory data are few, and mainly include a traditional supervised learning method (SVM) and a deep learning (RNN, CNN) classification method. However, the existing methods have the following drawbacks:
(1) the condition of unbalanced vehicle type distribution in a real road network cannot be fully considered, so that the problem of unbalanced training samples in the training process of the classification model can be generated, and the performance of the model is greatly reduced;
(2) the privacy of GPS data sources cannot be fully considered, data sets collected by the GPS data sources are low-frequency GPS data of a large scale and a long time period, the privacy protection problem is often caused when large-scale mobile data are collected, and an attacker can steal other privacy information of a vehicle by analyzing the large-scale and long-time GPS mobile data. Therefore, the invention focuses on developing a vehicle type identification model based on short-time GPS data. The short time means that the input of the model is a GPS track with a short time length, so that the privacy of a user can be greatly protected, but the reduction of information contained in input data is also meant, and the model needing to be designed can maximally extract effective features in the GPS track data.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a vehicle classification method based on short-time GPS track data.
The invention comprises the following steps:
step 1: preprocessing raw GPS data, wherein the preprocessing process comprises the following parts:
1) removing some of the repetitive and discontinuous GPS trajectory data;
2) converting the original GPS track data (time stamp and position coordinates) into the movement information (movement distance and driving mileage) and the motion state (including speed and acceleration) of the vehicle through a conversion formula;
3) considering the maximum threshold speeds of different vehicle types, discarding unrealistic invalid GPS points;
4) and superposing the motion characteristic vectors converted from the original GPS data to obtain the GPS data representation used for model input.
Step 2: balancing the data set:
adopting a resampling technology to rebalance the sample space of the unbalanced data set, and artificially generating some new track data for the GPS data (i.e. minority samples) of the motorcycles and heavy vehicles by an Adaptive Synthetic Sampling (ADASYN); some data are deleted from the GPS data (namely most types of samples) of the light vehicles by a random undersampling method so as to solve the problem of unbalanced vehicle type distribution in a real road network.
And step 3: constructing a vehicle classification model:
aiming at the time sequence and short-time characteristics of a GPS data source, the invention designs a deep learning network model (MDWD-BLSTMs) combining multi-level discrete wavelet decomposition (MDWD) and a bidirectional long-short term memory (LSTM) network for vehicle classification. The input of the model is preprocessed vehicle short-time GPS track data, and the output is a corresponding vehicle type.
And 4, step 4: model training and determination of optimal hyper-parameters:
the whole data set is divided into a training set and a testing set, the training data is sent to the model for training according to a certain batch, and the performance of the model under the testing data set at the moment is recorded. And continuously adjusting the performance of the hyper-parameter observation model through a large amount of combination training, and finally determining the optimal hyper-parameter.
And 5: vehicle classification:
and (3) preprocessing short-time GPS track data of unknown vehicle types according to the step 1 to obtain input network representation of the data, inputting the input network representation into the trained MDWD-BLSTMs to obtain the vehicle types corresponding to the tracks, and finishing vehicle classification.
The invention has the beneficial effects that: the invention constructs a task model combining wavelet decomposition and a bidirectional LSTM network, introduces wavelet transformation into GPS time series data application for the first time, not only retains the advantages of multi-level discrete wavelet decomposition in frequency learning, but also can acquire data information in the original time domain scope, and enables the model to have multi-view learning capability. Compared with the traditional method and the partial deep learning baseline method, the classification accuracy of the method is greatly improved.
Drawings
FIG. 1 is an example diagram of a GPS trajectory.
FIG. 2 is a diagram of new data artificially synthesized by ADASYNN algorithm with less sample data.
FIG. 3 is a diagram of the MDWD-BLSTMs vehicle classification framework.
Fig. 4 is a BLSTMs vehicle classification framework.
FIG. 5 is a graph comparing a training process with a deep learning baseline approach.
FIG. 6 is a graph comparing ROC curves with a conventional supervised learning method.
Detailed Description
The method for classifying the vehicles based on the GPS data has the advantages of incomparable low cost, no interference, high coverage and the like compared with the traditional vehicle classification method. Considering the sensitivity of GPS data and the unbalanced distribution of vehicle types in a real road network, the invention provides a vehicle classification method based on short-time GPS track data, which comprises the following steps:
step 1: the raw GPS data is preprocessed.
For example, a section of the driving track of a vehicle in the road network is intercepted to remove some repeated and discontinuous GPS track data, as shown in FIG. 1, the continuous GPS track data of the vehicle from A to B is composed of a series of collected GPS data points
Figure BDA0002957768720000051
And (4) forming. The movement information (information such as a movement distance and a mileage) and the motion state (including information such as a speed and an acceleration) of the vehicle are calculated by the following formulas:
1) vehicle travel distance d between any two GPS data pointsiCan be calculated by Haverine formula, pi(lati,loni) Indicates the longitude and latitude, p, of the GPS pointi-1(lati-1,loni-1) And pi(lati,loni) A distance d betweeniIs shown as：
Figure BDA0002957768720000052
2)tiTotal odometer distance o at timeiCan be from the first moment to the moment by a distance diThe summation of (a) results in:
Figure BDA0002957768720000053
3) the instantaneous speed v of the vehicle at the i-th instantiAnd instantaneous acceleration aiExpressed as:
Figure BDA0002957768720000054
4) the interval velocity represents an average velocity over a period of time, and may convey different and less noisy motion information. Assuming there are k GPS data points in a certain time period, the interval velocity in this time period can be calculated by the following formula
Figure BDA0002957768720000061
And interval acceleration
Figure BDA0002957768720000062
Figure BDA0002957768720000063
The raw GPS position coordinates can be converted into a vehicle motion signature sequence by these formulas. Then considering the maximum threshold speed of different vehicle types, discarding unrealistic invalid GPS points. And finally, overlapping the motion characteristic sequences of each sample according to the same length to obtain the GPS data representation for network input. Wherein each sample comprises a 6-channel structure, and each channel represents a moving distance diMileage oiInstantaneous speedDegree viInstantaneous acceleration aiSpeed of separation
Figure BDA0002957768720000064
And interval acceleration
Figure BDA0002957768720000065
Step 2: the data sets are balanced.
Since the distribution of vehicle types in the real road network is not uniform, the light vehicle (FHWA2-4) tends to have the highest proportion in the real road network, and therefore the GPS data set collected from the real road network causes a problem of imbalance of samples of each category. For this purpose, a resampling technique is used to rebalance the sample space of the unbalanced data set, new GPS track data is synthesized by using ADASYN algorithm for the GPS track data of the few types of samples, and the data set is balanced by using random undersampling for the GPS track data of the most types of samples. Table 1 is pseudo code for the ADASYN algorithm.
TABLE 1 ADASYNN Algorithm
Figure BDA0002957768720000066
Figure BDA0002957768720000071
Fig. 2 is a schematic diagram of instantaneous speed sequence data (few samples) of 15 motorcycles artificially synthesizing 70 new data by using the ADASYN algorithm, wherein a solid line represents original motorcycle track data, and a dotted line represents new track data synthesized by the algorithm, and it can be seen in the diagram that all new synthesized data are within the boundary range of the original data and are distinguished from the original data, so that the usability of the synthesized data is ensured, a certain overfitting risk is reduced, and the generalization capability of the model is improved.
And step 3: and constructing a vehicle classification model.
Aiming at the time sequence and short-time characteristics of a GPS data source, a novel model (MDWD-BLSTMs) for automatically classifying vehicles by combining Multistage Discrete Wavelet Decomposition (MDWD) with a bidirectional long short-term memory network (BLSTM) for short-time GPS trajectory data is provided, and the framework of the model is shown as figure 3. Decomposing the motion sequence of the vehicle into subsequences with different detail attributes through wavelet decomposition, taking the subsequences as the input of a plurality of independent bidirectional LSTM classification networks, and finally connecting all classifiers with different levels by adopting a residual error learning method to obtain a final classification result.
The MDWD-BLSTMs comprises a plurality of sub-classifiers, wherein the classifier 1 is a main classifier, and the classifiers 2-5 are auxiliary classifiers. The structure of the main classifier 1 is shown in fig. 4, and is a vehicle classification basic model comprising a two-layer BLSTM network structure, which is named as BLSTMs. The input of the network is GPS track points of the vehicle in a period of time
Figure BDA0002957768720000081
Wherein n is the number of GPS track points in the time period, and the GPS track points correspond to a group of input feature vector sequences
Figure BDA0002957768720000082
A Long Short Term Memory (LSTM) unit includes three gate control structures, namely a forgetting gate, an input gate and an output gate. The forgetting gate is controlled by a simple single-layer network, determines the amount of information forgotten from the previous cell state, and is defined as follows:
ft＝σ(Wf[at-1,xt]+bf)
wherein WfIs the weight vector of the forgetting gate, bfIs an offset vector and σ is a logic sigmoid function used to control the output of the forgetting gate. Candidate states within an LSTM cell at the current time
Figure BDA0002957768720000083
Can be calculated from the tanh function, and the formula is as follows:
Figure BDA0002957768720000084
the information that the input gate determines how many candidate states are stored in the cell is defined as follows:
it＝σ(Wi[at-1,xt]+bi)
then, the LSTM current state at time t is defined as ctWhere x represents an element-by-element multiplication.
Figure BDA0002957768720000091
The output gate determines how much information is output from the cell state by:
ot＝σ(Wo[at-1,xt]+bo)
the final LSTM cell output is:
at＝ot×tanh(ct)
the bidirectional LSTM network (BLSTM) has a forward and a backward loop network, both of which are connected to the same output layer to generate output information, the output being defined as follows:
Figure BDA0002957768720000092
Figure BDA0002957768720000093
wherein x istFor the input vector, the forward layer output is defined as
Figure BDA0002957768720000094
The output of the backward layer is defined as
Figure BDA0002957768720000095
Figure BDA0002957768720000096
Output vector of BLSTM network
Figure BDA0002957768720000097
Is a combination of these two outputs, and yt∈R2d. Network representation X ═ X { X } with 6 dimensional information can be obtained from raw GPS trajectory data according to step 11,x2,...,xn},
Figure BDA0002957768720000098
The input of the main classifier 1 is processed by two layers of BLSTMs with different neuron numbers, then processed by a shedding layer and finally processed by two layers of full connection layers, wherein the first full connection layer is activated by a Linear rectification function (ReLU), and the output vector of the last layer is directly transferred to a normalization index function (Softmax) so as to generate 3 output values Y { Y ═ between 0 and 11,y2,y3And the probability distributions are used for representing the probability distributions of three different vehicle types, namely a motorcycle (FHWA1), a light vehicle (FHWA2-4) and a heavy vehicle (FHWA 8-13). The structure of the auxiliary classifier is implemented by reducing the number of layers based on the main classifier, and the structure of each classifier is referred to table 2.
TABLE 2 classifier structures in MDWD-BLSTMs model
Figure BDA0002957768720000101
Secondly, extracting the instantaneous speed Xv＝{xv1,xv2,...,xvn},
Figure BDA0002957768720000102
Instantaneous acceleration Xa＝{xa1,xa2,...,xan},
Figure BDA0002957768720000103
Wavelet decomposition is performed separately, and "H" and "L" in fig. 3 represent a high-pass (HP) filter and a low-pass (LP) filter of a multi-level discrete wavelet decomposition, and the sequence decomposition process is defined as follows:
Figure BDA0002957768720000104
Figure BDA0002957768720000105
the input master timing sequence is denoted x [ n ]]Obtained from the continuous signal x (t),
Figure BDA0002957768720000106
and
Figure BDA0002957768720000107
a high pass filter and a low pass filter, respectively, the first wavelet decomposition will generate new low and high subsequences
Figure BDA0002957768720000108
And
Figure BDA0002957768720000109
obtained approximation coefficient
Figure BDA00029577687200001010
The approximation coefficients of the second stage can be obtained again by two filters
Figure BDA00029577687200001011
And detail coefficient
Figure BDA00029577687200001012
The above process is repeated until a specified level is reached.
Each independent classifier 1-5 in the MDWD-BLSTMs model is represented as
Figure BDA00029577687200001013
The output of the ith classifier is:
Figure BDA00029577687200001014
in addition, the MDWD-BLSTMs adopt a residual error learning method to connect the outputs u (i) of all classifiers, and when i is 1, the output predicted value is
Figure BDA0002957768720000111
Where S (x) represents the Softmax function,
Figure BDA0002957768720000116
representing the classification result 1, when i is more than or equal to 2, outputting a predicted value by each layer
Figure BDA0002957768720000112
Comprises the following steps:
Figure BDA0002957768720000113
wherein λiThe weight of the classification decision of the previous layer can be determined according to the importance degree of each level information, and the final classification result in the MDWD-BLSTMs model is
Figure BDA0002957768720000114
It is the result of the co-operation of a main classifier and a number of auxiliary classifiers.
And 4, step 4: model training and determination of optimal hyper-parameters.
The training process of the deep learning network is a long process, and firstly, a Batch Normalization (Batch Normalization) technology is adopted to normalize all data to be between 0 and 1, so that the purposes of simplifying calculation and reducing a numerical range are achieved. The training process is carried out based on a minimum classification cross entropy Loss function (Cross entropy Loss), and finally, an adaptive moment estimation (Adam) optimizer is adopted to update model parameters in a back propagation process. Through a number of training experiments, table 3 reports some of the key parameters used by each classifier in the training phase.
TABLE 3 partial Key parameters of the MDWD-BLSTMs model
Figure BDA0002957768720000115
Classifier 1 in the MDWD-BLSTMs model is a main classifier, the input of which contains all information, so that the decision weight parameter of the first layer is lambda1The input of the auxiliary classifier is partial features with different detail attributes, from which a certain proportion of decision information can be obtained, specifying λ2＝λ3＝λ4＝λ5λ, according to a number of experiments, the network performance is relatively highest when λ is 0.4. Thus selecting λ by default1＝1，λ2＝λ3＝λ4＝λ50.4 is the optimal parameter of the MDWD-BLSTMs model.
And 5: and (5) classifying the vehicles.
And (3) preprocessing short-time GPS track data of unknown vehicle types according to the step 1 to obtain input network representation of the data, and inputting the input network representation into the trained MDWD-BLSTMs to obtain the vehicle types corresponding to the tracks so as to finish the vehicle classification task.
In order to embody the performance of the MDWD-BLSTMs, a series of deep learning models containing LSTM units are designed, wherein the structures of the BLSTMs and the MDWD-BLSTMs are shown in figures 4 and 3, and the LSTMs and the MDWD-LSTMs are obtained by replacing all BLSTM units with LSTM units on the original basis. Fig. 5 shows a comparison of the performance of these models in the same training batch, and it is evident that the MDWD-BLSTMs model is more successful for the vehicle classification task. Furthermore, the network model with multi-level discrete wavelet decomposition performs much better than the base model.
Meanwhile, the invention also compares some existing classical machine learning technologies, which are Random Forest (RF), K-Nearest Neighbor (K-Nearest Neighbor) and Support Vector Machine (SVM), which are supervision algorithms widely used in GPS-based vehicle classification and driving mode detection methods. FIG. 6 shows the Receiver Operating Characteristic (ROC) curve of the MDWD-BLSTMs compared with the traditional machine learning model and the Area (AUC) enclosed by the coordinate axes under the ROC curve, and the result shows that the MDWD-BLSTMs are superior to the traditional model in the level of total AUC and have advantages in most specificity levels, and the AUC value reaches 0.9106.
In conclusion, the vehicle classification method based on the short-time GPS track data not only fully considers the privacy of the GPS data, but also provides a feasible solution for the imbalance of vehicle type distribution in a real road network, makes up for the defects of the prior art, and can effectively capture the depth features contained in the GPS track data, thereby achieving higher classification accuracy.
The above embodiments are merely to illustrate the technical solutions of the present invention and not to limit the present invention, and the present invention has been described in detail with reference to the preferred embodiments. It will be understood by those skilled in the art that various modifications and equivalent arrangements may be made without departing from the spirit and scope of the present invention and it should be understood that the present invention is to be covered by the appended claims.

Claims (5)

1. The vehicle classification method based on the short-time GPS track data is characterized by comprising the following steps:
step 1: preprocessing raw GPS data, wherein the preprocessing process comprises the following parts:
1) removing some of the repetitive and discontinuous GPS trajectory data;
2) converting original GPS track data into movement information and a motion state of a vehicle through a conversion formula;
3) considering the maximum threshold speeds of different vehicle types, discarding unrealistic invalid GPS points;
4) superposing the motion characteristic vectors converted from the original GPS data to obtain GPS data representation used for deep learning network model input;
step 2: employing a resampling technique for rebalancing the sample space of the unbalanced data set;
and step 3: constructing a vehicle classification model: aiming at the time sequence and short-time characteristics of a GPS data source, a deep learning network model is designed for vehicle classification; the input of the model is the vehicle short-time GPS track data after the step 2, and the output is the corresponding vehicle type; decomposing a motion sequence of a vehicle into subsequences with different detail attributes through wavelet decomposition, taking the subsequences as input of a plurality of independent bidirectional long-short term memory networks, and connecting all classifiers at different levels by adopting a residual learning method to obtain a final classification result;
and 4, step 4: deep learning network model training and optimal hyper-parameter determination: dividing the whole data set into a training set and a testing set, sending training data into the deep learning network model according to a certain batch for training, and recording the performance of the deep learning network model under the testing data set at the moment; through a large amount of combination training, continuously adjusting the performance of the hyper-parameter observation deep learning network model, and determining the optimal hyper-parameter;
and 5: vehicle classification: and (3) preprocessing short-time GPS track data of unknown vehicle types according to the step 1 to obtain input network representation of the data, inputting the input network representation into the trained deep learning network model to obtain vehicle types corresponding to the tracks, and finishing vehicle classification.
2. The short-time GPS track data-based vehicle classification method of claim 1, wherein: in the step 2:
and generating new track data for the GPS data of the motorcycles and the heavy vehicles by an adaptive synthesis sampling method.
3. The short-time GPS track data-based vehicle classification method of claim 1, wherein: in the step 2:
and deleting some data of the GPS data of the light vehicle by a random undersampling method.
4. The short-time GPS track data-based vehicle classification method of claim 1, wherein: in the step 3:
the deep learning network model comprises a main classifier and four auxiliary classifiers.
5. The short-time GPS trajectory data-based vehicle classification method of claim 4, characterized in that: the main classifier is a vehicle classification basic model containing two layers of long and short term memory networks; the method comprises the steps that a network representation with a plurality of dimensional information is used as input of a main classifier, the network representation passes through two layers of long and short term memory networks with different neuron numbers, then passes through a falling layer and finally is provided with two layers of full connection layers, wherein the first full connection layer is activated by a linear rectification function, and the output vector of the last layer is directly transmitted to a normalization exponential function, so that 3 output values ranging from 0 to 1 are generated.
CN202110228346.7A 2021-03-02 2021-03-02 Vehicle classification method based on short-time GPS track data Pending CN112948715A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110228346.7A CN112948715A (en) 2021-03-02 2021-03-02 Vehicle classification method based on short-time GPS track data

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110228346.7A CN112948715A (en) 2021-03-02 2021-03-02 Vehicle classification method based on short-time GPS track data

Publications (1)

Publication Number Publication Date
CN112948715A true CN112948715A (en) 2021-06-11

Family

ID=76247026

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110228346.7A Pending CN112948715A (en) 2021-03-02 2021-03-02 Vehicle classification method based on short-time GPS track data

Country Status (1)

Country Link
CN (1) CN112948715A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116759014A (en) * 2023-08-21 2023-09-15 启思半导体(杭州)有限责任公司 Random forest-based gas type and concentration prediction method, system and device
CN117473398A (en) * 2023-12-26 2024-01-30 四川国蓝中天环境科技集团有限公司 Urban dust pollution source classification method based on slag transport vehicle activity

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109285348A (en) * 2018-10-26 2019-01-29 深圳大学 A kind of vehicle behavior recognition methods and system based on two-way length memory network in short-term
CN110533007A (en) * 2019-09-13 2019-12-03 东南大学 A kind of vehicle-mounted strain of bridge influences the identification of line feature intelligent and extracting method
CN110865625A (en) * 2018-08-28 2020-03-06 中国科学院沈阳自动化研究所 Process data anomaly detection method based on time series
CN111053549A (en) * 2019-12-23 2020-04-24 威海北洋电气集团股份有限公司 Intelligent biological signal abnormality detection method and system
CN112257847A (en) * 2020-10-16 2021-01-22 昆明理工大学 Method for predicting geomagnetic Kp index based on CNN and LSTM

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110865625A (en) * 2018-08-28 2020-03-06 中国科学院沈阳自动化研究所 Process data anomaly detection method based on time series
CN109285348A (en) * 2018-10-26 2019-01-29 深圳大学 A kind of vehicle behavior recognition methods and system based on two-way length memory network in short-term
CN110533007A (en) * 2019-09-13 2019-12-03 东南大学 A kind of vehicle-mounted strain of bridge influences the identification of line feature intelligent and extracting method
CN111053549A (en) * 2019-12-23 2020-04-24 威海北洋电气集团股份有限公司 Intelligent biological signal abnormality detection method and system
CN112257847A (en) * 2020-10-16 2021-01-22 昆明理工大学 Method for predicting geomagnetic Kp index based on CNN and LSTM

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
MATTEO ET AL.,: "Vehicle classifification from low-frequency GPS data with recurrent neural networks", 《TRANSPORTATION RESEARCH》, 13 April 2018 (2018-04-13), pages 32 - 33 *
王习昇等: "动态小波变换网络的短时交通流量预测", 《单片机与嵌入式系统应用》, 1 November 2020 (2020-11-01), pages 179 - 188 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116759014A (en) * 2023-08-21 2023-09-15 启思半导体(杭州)有限责任公司 Random forest-based gas type and concentration prediction method, system and device
CN116759014B (en) * 2023-08-21 2023-11-03 启思半导体(杭州)有限责任公司 Random forest-based gas type and concentration prediction method, system and device
CN117473398A (en) * 2023-12-26 2024-01-30 四川国蓝中天环境科技集团有限公司 Urban dust pollution source classification method based on slag transport vehicle activity
CN117473398B (en) * 2023-12-26 2024-03-19 四川国蓝中天环境科技集团有限公司 Urban dust pollution source classification method based on slag transport vehicle activity

Similar Documents

Publication Publication Date Title
CN107492251B (en) Driver identity recognition and driving state monitoring method based on machine learning and deep learning
CN112948715A (en) Vehicle classification method based on short-time GPS track data
CN110738855B (en) Road traffic flow condition prediction method in data sparse time period
CN116010885A (en) Method and system for detecting abnormal space-time data of vehicle under long-sequence condition
CN112508056A (en) Urban air quality monitoring method based on mobile multi-source perception
CN110930693B (en) Online short-term traffic flow prediction method for road section
CN111753667A (en) Intelligent automobile single-target tracking method based on twin network
Menegazzo et al. Multi-contextual and multi-aspect analysis for road surface type classification through inertial sensors and deep learning
CN112612820A (en) Data processing method and device, computer readable storage medium and processor
CN113610188A (en) Bow net contact force non-section abnormity identification method and device
CN112884014A (en) Traffic speed short-time prediction method based on road section topological structure classification
CN113222385A (en) Method for constructing and evaluating driving condition of electric automobile
Li et al. Vehicle classification and speed estimation based on a single magnetic sensor
Wang et al. Convolutional neural network-based moving ground target classification using raw seismic waveforms as input
Zhao et al. FMCNN: A factorization machine combined neural network for driving safety prediction in vehicular communication
Lee et al. Road type classification using deep learning for Tire-Pavement interaction noise data in autonomous driving vehicle
Marciniuk et al. Machine learning applied to acoustic-based road traffic monitoring
CN118070105A (en) Intelligent self-adaptive pavement detection and maintenance method and system
Alazeb et al. Intelligent Transportation Activity Recognition Using Deep Belief Network
Zhang et al. Structural Damage Identification System Suitable for Old Arch Bridge in Rural Regions: Random Forest Approach.
Sun et al. Vehicle acoustic and seismic synchronization signal classification using long-term features
Yang et al. Research on evaluation model for vehicle interior sound quality based on an optimized BiLSTM using genetic algorithm
Wang et al. Contrastive GNN-based traffic anomaly analysis against imbalanced dataset in IoT-based its
CN116092037A (en) Vehicle type identification method integrating track space-semantic features
Chen et al. Road roughness level identification based on bigru network

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
RJ01 Rejection of invention patent application after publication

Application publication date: 20210611

RJ01 Rejection of invention patent application after publication