A method and system for predicting driving destinations
By using deep learning technology, combined with sequence generation and classification models of long short-term memory networks and convolutional neural networks, the problems of data sparsity and difficulty in determining nodes in graph neural networks for vehicle destination prediction are solved, achieving more efficient vehicle destination prediction and adapting to different cities and vehicle trajectory data.
Patent Information
- Application Number
- CN202310963825.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-08-02
- Publication Date
- 2026-01-06
- Estimated Expiration
- 2043-08-02
AI Technical Summary
Existing graph neural network technology suffers from problems such as data sparsity and difficulty in determining nodes in vehicle destination prediction, resulting in poor prediction performance.
Using deep learning technology, a sequence generation and classification model combining long short-term memory networks and convolutional neural networks is used to predict vehicle destinations using GPS data from operating vehicles. This includes data cleaning, geolocation encoding, sequence generation and classification, and an Encoder-Decoder architecture is used for model training and prediction.
It improves the performance of vehicle destination prediction, has good portability and flexibility, adapts to different cities and vehicle trajectory data, and enhances the accuracy and efficiency of prediction.
Smart Images
Figure CN117171635B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of big data-driven traffic and driving, and in particular to a method and system for predicting driving destinations. Background Technology
[0002] The advent of the automobile over the past century has dramatically changed the way people travel. With the rapid increase in the number of vehicles owned, urban traffic networks have become increasingly complex, significantly impacting people's travel experience. Currently, many ride-hailing and navigation apps offer intelligent route planning functions to help users find routes to their destinations faster. However, these applications all rely on users inputting their destinations, and there are still many issues with destination prediction functions. Predicting a user's destination could significantly improve the user's travel experience. For example, finding gas stations / charging stations along the user's route; planning nearby parking locations before the user arrives at their destination; and suggesting possible destinations of interest to the user.
[0003] In existing destination prediction technologies, trajectory location information is often too sparse. Traditional methods, based on trajectory discretization strategies, employ road grid matching to tightly integrate vehicle trajectories with road information. This abstracts the vehicle's origin, destination, and trajectory into a graph network. The topology of this graph can be obtained by constructing a directed graph, and the next possible node is found using the shortest path problem approach. The path planning algorithms used include non-heuristic methods, such as dynamic Dijkstra's algorithm and queue-based Dijkstra's algorithm, as well as heuristic algorithms, such as A* algorithm and genetic algorithms. Although graph neural network technology has made significant progress in recent years, improving vehicle target recognition and classification, using graph networks to map road information for vehicle destination prediction still faces many challenges, such as difficulty in determining nodes and insufficient data. Summary of the Invention
[0004] This invention addresses the problems existing in the prior art by providing a method and system for predicting travel destinations. It solves the problems of data sparsity and difficulty in determining nodes in traditional graph neural network technology in travel destination prediction scenarios, and has good transferability and flexibility, significantly improving prediction performance.
[0005] This invention discloses a method for predicting a driving destination, comprising:
[0006] Step 1.1: Obtain and clean the GPS data of operating vehicles from the open platform;
[0007] Step 1.2: Preprocess the GPS data of the operating vehicles. This preprocessing includes at least geolocation encoding to generate the corresponding location information string.
[0008] Step 1.3: Input the preprocessed data into the sequence generation algorithm for model training to obtain the sequence generation model; in the case that the sequence generation algorithm is a long short-term memory network, the encoder inputs a position information string of arbitrary length into a vector c, the decoder obtains the vector c, decodes the information, and outputs it as a sequence;
[0009] Step 1.4: Input the preprocessed data into a sequence classification algorithm for model training to obtain a sequence classification model; in the case of Test-CNN, the Embedding layer represents each word as a vector; each GeoHash tag is regarded as a word, and each position tag in the GeoHash sequence needs to be converted into a vector form; then it enters the convolutional layer, inputting a sequence containing s tags, setting the Embedding vector of each tag to d dimensions, then the input is s×d. Treat this input as an image, and use convolution to extract features of adjacent tags. One-dimensional convolution is used, and the width of the convolution kernel is the dimension d of the word vector, so the convolution kernel size is w×d; then the data is passed into the pooling layer to further extract features and compress the matrix size; the sequence classification model is finally connected to a fully connected layer and a Softmax layer to output the final classification.
[0010] Step 2: Obtain vehicle GPS data during the driving process;
[0011] Step 3: Preprocess the vehicle's GPS data, including at least geolocation encoding, to generate the corresponding location information string;
[0012] Step 4: Input the preprocessed data from Step 3 into the sequence generation model and the sequence classification model, and output the inter-journey prediction results and the destination prediction results, respectively;
[0013] Step 5: Display the inter-trip prediction results and the destination prediction results through the results output module.
[0014] Furthermore, in step 1.1, the GPS data of the operating vehicle includes the license plate number, map longitude, map latitude, GPS time, GPS speed, and altitude.
[0015] Furthermore, in step 1.2, the geolocation encoding method is the GeoHash method.
[0016] Furthermore, in step 1.3, the sequence generation algorithm is a long short-term memory network, a recurrent neural network, or a Transformer.
[0017] Furthermore, in step 1.3, the vector c is the hidden state h of the last neuron in the sequence information input to the encoder. N express;
[0018] Alternatively, vector c can be obtained by performing a transformation q on the hidden state of the last neuron in the encoder's input sequence information, i.e., c = q(h) N );
[0019] Alternatively, vector c can be the hidden states h1, h2, ..., hn of all neurons in the sequence information input to the encoder. N It is obtained through a certain transformation q, that is, expressed as c = q(h1, h2, ..., h N ).
[0020] Generally, the transformation q refers to any user-defined function. Preferably, the transformation q is a nonlinear transformation, such as the tan function.
[0021] Furthermore, in step 2, the vehicle driving data includes license plate number, map longitude, map latitude, GPS time, GPS speed, and altitude.
[0022] Furthermore, in step 3, the geolocation encoding method is the GeoHash method.
[0023] Furthermore, in step 3, the vehicle driving data is preprocessed, including removing completely duplicate records, removing invalid noise data, segmenting the trip, and encoding the geographic location.
[0024] This invention also discloses a destination prediction system, which applies the destination prediction method described above, and the system includes:
[0025] The data acquisition module is used to acquire GPS data of the vehicle during the driving process;
[0026] The data preprocessing module is used to preprocess the vehicle's GPS data. This preprocessing includes at least geolocation encoding to generate corresponding location information strings.
[0027] The model prediction module includes a sequence generation model and a sequence classification model, which are used to output the inter-journey prediction results and the destination prediction results respectively after the preprocessed data is input into the sequence generation model and the sequence classification model.
[0028] The results output module is used to obtain and display the prediction results during the journey and the prediction results at the destination.
[0029] The present invention also discloses a storage medium storing a computer program, wherein the computer program is configured to execute the above-described destination prediction method when running on a computer or processor.
[0030] The present invention has at least the following beneficial effects:
[0031] This invention uses deep learning technology to predict vehicle destinations based on vehicle trajectory data, solving the problems of data sparsity and difficulty in determining nodes in traditional graph neural network technology in the context of destination prediction. It has good transferability and flexibility, significantly improving prediction performance and providing a foundation for the currently popular autonomous driving and ITS (Intelligent Traffic System).
[0032] Other beneficial effects of the present invention will be described in detail in the Detailed Description of the Embodiments section. Attached Figure Description
[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0034] Figure 1 This is a schematic diagram of the model input and output principle disclosed in a preferred embodiment of the present invention.
[0035] Figure 2 This is a schematic diagram of the sequence generation model disclosed in a preferred embodiment of the present invention.
[0036] Figure 3 This is a schematic diagram of the sequence classification model disclosed in a preferred embodiment of the present invention.
[0037] Figure 4 This is a system block diagram of the preferred embodiment of the driving destination prediction system disclosed in this invention. Detailed Implementation
[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0039] This embodiment discloses a method for predicting vehicle destinations, which includes the following steps:
[0040] Step 1: Train a sequence generation model and a sequence classification model respectively.
[0041] The training method includes:
[0042] Step 1.1: Obtain and clean GPS data of operating vehicles from a certain open platform. The data includes fields such as license plate number, map longitude, map latitude, GPS time, GPS speed, and altitude.
[0043] Step 1.2 involves preprocessing the data, including removing completely duplicate records, removing invalid noise data, segmenting the trip, and geolocation encoding. Preferably, the geolocation encoding method is the GeoHash method.
[0044] GeoHash is an address encoding method that encodes two-dimensional spatial latitude and longitude data into strings. It approximates the corresponding longitude or latitude values infinitely using a binary search method. After processing the coordinates using the GeoHash algorithm, each coordinate generates a corresponding location information string. The longer the GeoHash encoding length, i.e., the longer the location information string, the higher the location accuracy.
[0045] Step 1.3: Input the encoded strings into the sequence generation algorithm for model training. Step 1.4: Input the encoded strings into the sequence classification algorithm for model training.
[0046] For example, the sequence generation algorithm is a Long Short-Term Memory network.
[0047] like Figure 2 As shown, a string of location information of arbitrary length is stored in a vector c. That is... Figure 2 h1, h2, ..., h N After obtaining the context information vector c, the decoder can decode the information and output it as a sequence. That is... Figure 2 h′1, h′2, ..., h′ M The context vector c can be calculated in several ways. c can be directly calculated using the hidden state h of the last neuron. N It can also be obtained by performing some transformation on the hidden state of the last neuron, and can be represented as q(h). N The q function represents a certain transformation; alternatively, the hidden states h1, h2, ..., h of all neurons can be used. N The context vector c is calculated and then passed to the Decoder. The q function, representing a transformation, is generally an arbitrary user-defined function; preferably, it is a non-linear transformation, such as the tan function.
[0048] The sequence generation algorithm can also be a recurrent neural network, Transformer, etc.
[0049] For example, the sequence classification algorithm is Test-CNN.
[0050] The Embeddinglayer represents each word as a vector. It can use a pre-trained model or a randomly initialized one. In our predictions, each GeoHash token can be considered a word, and we need to convert each position token in the GeoHash sequence into a vector. This is then fed into a convolutional layer. The input is a sequence containing *s* tokens. Assuming the Embedding vector of each token is *d*-dimensional, the input is *s×d*. Treating this input as an image, convolution extracts features from adjacent tokens. A one-dimensional convolution is used, with the kernel width equal to the word vector dimension *d*, resulting in a kernel size of *w×d*. The data is then passed to a pooling layer for further feature extraction and matrix compression. Finally, the model is connected to a fully connected layer and a Softmax layer to output the final classification.
[0051] Each input sequence corresponds to a destination. Through extensive training, the model can predict a destination GeoHash tag for an input GeoHash tag sequence. Therefore, the output of this branch is an independent GeoHash tag, i.e., the vehicle's destination.
[0052] Step 2: Obtain actual vehicle driving data, including license plate number, map longitude, map latitude, GPS time, GPS speed, altitude, and other fields.
[0053] Step 3 involves preprocessing the data obtained in Step 2, including removing completely duplicate records, removing invalid noise data, segmenting the trip, and geolocation encoding. Preferably, the geolocation encoding method is the GeoHash method.
[0054] GeoHash is an address encoding method that encodes two-dimensional spatial latitude and longitude data into strings. It approximates the corresponding longitude or latitude values infinitely using a binary search method. After processing the coordinates using the GeoHash algorithm, each coordinate generates a corresponding location information string. The longer the GeoHash encoding length, i.e., the longer the location information string, the higher the location accuracy.
[0055] Step 4: Input the preprocessed data from Step 3 into the model prediction module, load the preset sequence generation model and sequence classification model, and output the inter-journey prediction results and the destination prediction results, respectively.
[0056] Step 5: The results output module displays the results from step 4.
[0057] like Figure 4 As shown, the present invention also discloses a destination prediction system, including: a data acquisition module, a data preprocessing module, a model prediction module, and a result output module.
[0058] The data acquisition module acquires at least GPS coordinates and time-series data during vehicle operation. The data preprocessing module receives the GPS coordinates and time-series data from the data acquisition module, reduces and discretizes the data, and then performs geolocation encoding. Geolocation encoding refers to converting the GPS coordinate data into a string representation. The model prediction module pre-sets one or more deep learning-based models to perform inference analysis on the strings and time-series data processed by the data preprocessing module.
[0059] In this embodiment, the deep learning model includes a model for pathpoint prediction (sequence generation model), such as... Figure 1 and Figure 2 As shown, this model is built on a Seq2Seq sequence generation framework and includes an Encoder-Decoder architecture. This invention can implement the Encoder and Decoder parts using LSTM, which generates a new sequence based on the input sequence, representing the vehicle's subsequent trajectory. Its output contains multiple sequences labeled with the spatial index string, where each label can serve as the vehicle's next destination during its journey. The last label in the sequence can be considered the end point of the journey. The Encoder and Decoder in the Seq2Seq structure can be implemented in various ways, employing algorithm frameworks such as LSTM, GRU, or Transformer.
[0060] like Figure 1 and Figure 3 As shown, the deep learning model also includes a destination prediction model (sequence classification model) based on a sequence classification algorithm. Each input sequence corresponds to a destination, and through extensive training, this model can predict a destination string label for the input spatial index label sequence. Therefore, the output of the branch is an independent string index label, i.e., the vehicle's destination.
[0061] The result output module obtains and displays the prediction results from the model prediction module, and the display method can be a mobile APP or a web page platform.
[0062] Regarding the aforementioned destination prediction system, this invention specifically discloses its usage process:
[0063] (1) Obtain latitude and longitude coordinates, time series and other data of vehicle travel;
[0064] (2) Preprocess the data described in step (1), including removing completely duplicate records, removing invalid noise data, and forming segments.
[0065] Preprocessing methods also include converting latitude and longitude coordinates into strings of a certain length using GeoHash location space indexing.
[0066] (3) Input the data described in step (1) into a preset deep learning model and output the prediction results. The deep learning model includes a user journey prediction model based on Seq2Seq sequence generation and a journey destination prediction model based on sequence classification algorithm.
[0067] (4) Output the model prediction results described in step (3) to the result display model and display them on multiple terminals.
[0068] This invention improves and optimizes vehicle destination prediction technology, and has the following advantages and effects:
[0069] (1) Good transferability. Our training data is based on a large number of anonymous operating vehicle trajectories in Shenzhen. When it is necessary to deal with the vehicle destination prediction problem in other cities or regions, only the training data needs to be replaced, and the model does not need to be modified.
[0070] (2) The model is flexible. Our deep learning model is based on an Encoder-Decoder sequence generation model, where the Encoder and Decoder can be implemented using different techniques, such as the currently used CNN / LSTM. As sequence generation technology continues to develop, we can update this structure with more advanced techniques, such as Transformer. This continuously improves the model's predictive ability. In addition, the flexible model structure also allows our vehicle destination prediction scheme to be adjusted for various types of vehicle trajectory data, achieving better prediction performance.
[0071] The present invention also discloses a storage medium storing a computer program, wherein the computer program is configured to execute the driving destination prediction method disclosed in the above embodiments when running on a computer or processor.
[0072] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A travel destination prediction method characterized by comprising: The method comprises the following steps: Step 1.1: Obtain the GPS data of the operating vehicles of the open platform and clean it up; Step 1.2: Preprocess the GPS data of the operating vehicles, which at least includes geocoding to generate corresponding position information strings; Step 1.3: Input the preprocessed data into a sequence generation algorithm to train a sequence generation model; in the case that the sequence generation algorithm is a long short-term memory network, an encoder inputs an arbitrary length of position information strings into a vector c, and a decoder decodes the information after obtaining the vector c and outputs it as a sequence; Step 1.4: Input the preprocessed data into a sequence classification algorithm to train a sequence classification model; in the case that the sequence classification algorithm is Test-CNN, an Embedding layer represents each word as a vector; each GeoHash marker is regarded as a word, and each position marker in the GeoHash sequence needs to be converted into a vector form; then the data is input into a convolution layer, an input sequence containing s markers is set, the Embedding vector of each marker is d-dimensional, the input is sxd, the input is regarded as an image, the convolution extracts the features of adjacent markers, one-dimensional convolution is adopted, the width of the convolution kernel is the dimension d of the word vector, and the size of the convolution kernel is wxd; then the data is transmitted to a pooling layer to further extract features and compress the matrix size; the sequence classification model is connected with a fully connected layer and a Softmax layer at the end to output the final classification; Step 2: Obtain the GPS data of the vehicle during the driving process; Step 3: Preprocess the GPS data of the vehicle, which at least includes geocoding to generate corresponding position information strings; Step 4: Input the preprocessed data of step 3 into the sequence generation model and the sequence classification model to output the intermediate travel prediction result and the end prediction result respectively; Step 5: Display the intermediate travel prediction result and the end prediction result through a result output module.
2. The travel destination prediction method according to claim 1, characterized by, In step 1.1, the GPS data of the operating vehicles includes license plate number, map longitude, map latitude, GPS time, GPS speed, and altitude.
3. The travel destination prediction method according to claim 1, characterized by, In step 1.2, the geocoding method is the GeoHash method.
4. The travel destination prediction method according to claim 1, characterized by, In step 1.3, the sequence generation algorithm is a long short-term memory network or a recurrent neural network or a Transformer.
5. The travel destination prediction method according to claim 1, characterized by, In step 1.3, the vector c takes the hidden state h of the last neuron in the sequence information input by the encoder N denotes; Alternatively, the vector c is obtained by applying some transformation q to the hidden state of the last neuron in the sequence information input to the encoder, i.e. c = q(h N ). Alternatively, the vector c takes the hidden states hi, h2,..., h N through some transformation q, i.e. denoted as c = q(hi, h2,..., h N ) ; The certain transformation q refers to an arbitrary user-defined function; or the certain transformation q is a nonlinear transformation.
6. The travel destination prediction method according to claim 1, characterized by, In step 2, the vehicle driving data includes license plate number, map longitude, map latitude, GPS time, GPS speed, and altitude.
7. The travel destination prediction method according to claim 1, characterized by, In step 3, the geocoding method is the GeoHash method.
8. The travel destination prediction method according to claim 1, characterized by, In step 3, the preprocessing of the vehicle driving data includes removing completely repeated records, removing invalid noise data, segmenting the journey, and geocoding.
9. A travel destination prediction system characterized by comprising: The system applies the driving destination prediction method of any one of claims 1 to 8, and the system comprises: a data acquisition module configured to acquire vehicle driving GPS data during driving; a data preprocessing module configured to preprocess the vehicle driving GPS data, the preprocessing at least including geographical position coding to generate a corresponding position information string; a model prediction module including a sequence generation model and a sequence classification model, configured to output an inter-journey prediction result and a terminal prediction result respectively after inputting the preprocessed data into the sequence generation model and the sequence classification model; a result output module configured to acquire and display the inter-journey prediction result and the terminal prediction result.
10. A storage medium, characterized by a computer program is stored, wherein the computer program is configured to execute the driving destination prediction method of any one of claims 1 to 8 when running on a computer or a processor.