Map matching method and device based on sparse cellular signaling data and storage medium
By segmenting and spatially encoding the base station trajectory sequence of cellular signaling data, and combining it with a deep learning model for data augmentation, the problem of low matching accuracy caused by low cellular signaling data quality was solved, achieving high-precision map matching results while reducing energy consumption and costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CENT SOUTH UNIV
- Filing Date
- 2022-12-21
- Publication Date
- 2026-06-02
AI Technical Summary
Existing map matching technologies suffer from low matching accuracy when using cellular signaling data due to low data quality, failing to meet the requirements for high precision.
By acquiring base station trajectory sequences from cellular signaling data, the data is divided into sub-base station trajectory sequences, and spatial encoding and data augmentation are performed. Data augmentation is then carried out using a deep learning model, including a GRU network and an Attention layer. Finally, a map matching model is used to obtain accurate matching results.
It improves the map matching accuracy of cellular signaling data, reduces energy consumption, and is safer due to the passive collection method of cellular signaling data, almost completely covering the user's activity range without additional cost.
Smart Images
Figure CN115797586B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of map matching technology, and in particular relates to a map matching method, device and storage medium based on sparse cellular signaling data. Background Technology
[0002] As one of the most important driving forces for the construction of smart cities, location-based services play an indispensable role in people's daily lives and are widely used in navigation software, social networks, and other fields. Location-based services refer to services that rely on the location information of mobile users, which can be achieved through mobile operators' radio communication networks (such as 4G networks) or external positioning methods (such as GPS).
[0003] However, due to the influence of harsh environments, both raw GPS and cellular data are drift points, with positioning errors ranging from a few meters to hundreds of meters. This results in inaccurate and uncertain positioning, which may prevent some trajectory-based applications from meeting their requirements, such as Google Maps, Amap, and route optimization. Therefore, map matching emerged to match raw, inaccurate GPS or cellular trajectory points to road networks to identify the actual movement paths and locations of moving objects. With the emergence of various location-based service applications such as navigation, positioning, trajectory analysis, and prediction, map matching technology has been widely used in various software, advancing the era of intelligent transportation.
[0004] Most existing map matching technologies are based on GPS tracks. Because GPS data is of relatively high quality, the matching accuracy is satisfactory, especially for GPS tracks with high sampling rates. However, GPS track data also has some drawbacks, such as intermittent use, limited user coverage, additional data collection costs, and a high risk of privacy breaches.
[0005] With the continuous advancement of mobile communication technology, mobile communication base stations have expanded globally. Compared to GPS tracing, cellular tracing, especially cellular signaling tracing, offers significant advantages through its passive data collection method, including continuous availability, complete coverage, and no additional cost. However, achieving satisfactory map matching performance using cellular data presents a considerable challenge.
[0006] First, in order to achieve method generalization, only basic cellular signaling data can be used, and the quality of user cellular signal trajectories may be very low, which may lead to map matching errors.
[0007] Secondly, traditional map matching models based on GPS trajectories are not suitable for cellular signal data with much lower data quality, resulting in low matching accuracy. Summary of the Invention
[0008] The purpose of this invention is to provide a map matching method, device, and storage medium based on sparse cellular signaling data, so as to solve the problem of map matching errors and low matching accuracy caused by the low quality of cellular signaling data when using cellular signaling data for map matching.
[0009] This invention solves the above-mentioned technical problems through the following technical solution: a map matching method based on sparse cellular signaling data, comprising the following steps:
[0010] The base station trajectory sequence for acquiring cellular signaling data, wherein each base station trajectory sequence is composed of the IDs of each base station connected to by a passenger during a certain mode of transportation and the corresponding connection time;
[0011] The base station trajectory sequence is divided into trajectory segments to obtain the trajectory sequences of each sub-base station;
[0012] Each of the sub-base station trajectory sequences is spatially encoded to obtain the corresponding spatially encoded sequence;
[0013] Data augmentation is performed on each of the spatial coding sequences to obtain a dense representation sequence of sparse cellular signaling data in the corresponding sub-base station trajectory sequence;
[0014] Each of the dense representation sequences is input into the map matching model to obtain the map matching result of the corresponding sub-base station trajectory sequence, and then the map matching result of the base station trajectory sequence is obtained.
[0015] Furthermore, the trajectory sequence of the base station is divided using a sliding window, the specific implementation process of which is as follows:
[0016] Using the connection time of the first base station in the base station trajectory sequence as the start time of the sliding window, and setting the width of the sliding window to T, the first sub-base station trajectory sequence is obtained.
[0017] Slide the sliding window forward by H base stations to obtain the trajectory sequence of the next sub-base station;
[0018] By analogy, the trajectory sequences of each sub-base station of the base station trajectory sequence are obtained.
[0019] This invention uses cellular signaling data for map matching. Compared to GPS data, cellular signaling data almost completely covers the user's activity range, is continuously available, has no additional cost, lower energy consumption requirements, and is more secure. By dividing the base station trajectory sequence into sub-base station trajectory sequences, the original data volume is expanded, achieving data augmentation, which in turn helps improve map matching accuracy. Further data augmentation is performed on the spatial coding sequence, further improving map matching accuracy. Finally, an accurate map matching result is obtained using a map matching model.
[0020] Furthermore, when the difference in connection time between the first and last base stations in the last sub-base station trajectory sequence is less than T, the last sub-base station trajectory sequence is merged with the previous sub-base station trajectory sequence into a single sub-base station trajectory sequence.
[0021] Furthermore, when the difference in connection time between the first and last base stations in a certain sub-base station trajectory sequence is less than T / 2, the sub-base station trajectory sequence is merged with the previous sub-base station trajectory sequence into a single sub-base station trajectory sequence.
[0022] Furthermore, the specific implementation process for spatial encoding of each of the sub-base station trajectory sequences is as follows:
[0023] The latitude and longitude coordinates of each base station are determined based on the base station IDs in the sub-base station trajectory sequence;
[0024] The latitude and longitude coordinates of each base station are converted into a grid ID, and the specific calculation formula is as follows:
[0025]
[0026]
[0027] Where GID represents the grid ID corresponding to the base station, lon represents the longitude of the base station, lat represents the latitude of the base station, and LON represents the longitude of the base station. max LON min LAT represents the gridded longitude range. max LAT min C represents the range of gridded latitudes. O R represents the number of grid columns. O D represents the number of grid rows. LO D represents the Earth's distance corresponding to a gridded range of longitudes. LA L represents the Earth's distance corresponding to a gridded range of latitudes. G This represents the side length of each grid cell;
[0028] The spatial coding sequence is composed of all grid IDs of the sub-base station trajectory sequence.
[0029] Furthermore, the Earth distance D corresponding to the gridded longitude range LO The Earth's distance D corresponding to the gridded latitude range LA The calculation formula is:
[0030] D LO = 2 * radius * Distance(d(LON) min ),d(LAT min ),d(LON max ),d(LAT min ))
[0031] D LA = 2 * radius * Distance(d(LON) min ),d(LAT min ),d(LON min ),d(LAT max ))
[0032] Distance(x1,y1,x2,y2)=2*radius*c(x1,y1,x2,y2)
[0033]
[0034]
[0035]
[0036] Where radius represents the Earth's radius, and Distance() represents the distance calculation function. Indicates the angle Convert to radians, c() represents an intermediate function.
[0037] Furthermore, data augmentation is performed on each of the spatially encoded sequences using a data augmentation model. The data augmentation model includes a first encoder and a first decoder. The first encoder includes an embedding layer and a GRU network, and the first decoder includes a GRU network, an attention layer, a fully connected layer, and an argmax layer. The specific implementation process of the data augmentation is as follows:
[0038] The spatial coding sequence is encoded using an embedding layer to obtain a multidimensional output vector;
[0039] The multidimensional output vector is further encoded using the GRU network of the first encoder to obtain the encoded feature vector;
[0040] The encoded feature vector is decoded using the first decoder to obtain a dense representation sequence.
[0041] Furthermore, the map matching model includes a second encoder and a second decoder. The second encoder includes an embedding layer and a GRU network, and the second decoder includes a GRU network, an attention layer, a fully connected layer, and an argmax layer. The specific process for obtaining the map matching results of the sub-base station trajectory sequence is as follows:
[0042] The dense representation sequence is encoded using a second encoder to obtain a hidden state sequence;
[0043] The effective feature vector is obtained by using an Attention layer to perform a weighted summation of all hidden states in the hidden state sequence.
[0044] The effective feature vector is decoded using the GRU network of the second decoder to obtain a multidimensional vector;
[0045] The multidimensional vector is processed sequentially using a fully connected layer and an argmax layer to obtain the map matching result of the sub-base station trajectory sequence corresponding to the dense representation sequence.
[0046] Based on the same inventive concept, the present invention also provides an electronic device, comprising:
[0047] Memory, used to store computer programs;
[0048] A processor is configured to implement the map matching method based on sparse cellular signaling data as described above when executing the computer program.
[0049] Based on the same inventive concept, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the map matching method based on sparse cellular signaling data as described above.
[0050] Beneficial effects
[0051] Compared with the prior art, the advantages of the present invention are as follows:
[0052] The present invention provides a map matching method, device and storage medium based on sparse cellular signaling data. Compared with GPS data, cellular signaling data almost completely covers the user's activity range, is continuously available, has no additional cost, has lower energy consumption requirements, and is more secure than GPS data because of the passive collection method of cellular signaling data (i.e., passively acquired by the user while using the mobile phone).
[0053] This invention significantly improves the quality of cellular signal trajectories by dividing the base station trajectory sequence into multiple sub-base station trajectory sequences and performing data augmentation on the spatial coding sequence of the sub-base station trajectory sequences. This reduces the difficulty of map matching for sparse cellular signaling data while greatly improving the accuracy of map matching. Attached Figure Description
[0054] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0055] Figure 1 This is a flowchart of the map matching method in an embodiment of the present invention;
[0056] Figure 2(a) is a schematic diagram of the average accuracy results in the comparative test in the embodiment of the present invention;
[0057] Figure 2(b) is a schematic diagram of the average recall results in the comparative experiment in the embodiment of the present invention;
[0058] Figure 2(c) is a schematic diagram of the average F1 score results in the comparative experiment in the embodiment of the present invention;
[0059] Figure 2(d) is a schematic diagram of the average time results in the comparative experiment in the embodiment of the present invention;
[0060] Figure 3 This is a schematic diagram of the ablation test results in an embodiment of the present invention. Detailed Implementation
[0061] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0062] The technical solutions of this application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.
[0063] Definitions:
[0064] Embedding: A vector representation method that uses a low-dimensional dense vector to represent an object, so that the vector can express certain features of the corresponding object, and the distance between vectors can reflect the similarity between objects.
[0065] Encoder-decoder: A deep learning model framework consisting of two parts: an encoder and a decoder. The encoder encodes the input into an intermediate semantic representation, and the decoder decodes the intermediate state and then outputs the result.
[0066] Gated Recurrent Neural Network (GRU): A variant of recurrent neural network (RNN) suitable for sequentially related sequences.
[0067] Attention mechanism: A mechanism for focusing limited attention on key information.
[0068] Hidden Markov Model (HMM): A machine learning model that can be used for map matching. It uses location sampling points as observation points and real road segments as hidden states, transforming map matching into a decoding problem.
[0069] To address the challenges of achieving satisfactory map matching performance using cellular data, such as... Figure 1 As shown in the figure, the map matching method based on sparse cellular signaling data provided by the embodiment of the present invention includes the following steps:
[0070] Step S100: Obtain the base station trajectory sequence of cellular signaling data. Each base station trajectory sequence consists of the base station IDs and corresponding connection times connected by the passenger during a certain mode of transportation.
[0071] When a user uses their mobile phone while traveling, the phone automatically connects to various base stations and obtains the base station trajectory sequence of cellular signaling data. For example, data from over two hundred users traveling by car totals over two hundred hours and over six thousand kilometers; data from approximately fifty users traveling by bus totals over ten hours and over two hundred kilometers; and data from approximately thirty users traveling by subway totals over forty hours and over one thousand kilometers. The base station trajectory sequence obtained for a particular user traveling to their destination using a certain mode of transportation can be represented as follows: Among them, P i This represents the i-th base station trajectory sequence. This represents the information of the k-th base station connected in the i-th base station trajectory sequence. Including base station ID and the time of connection to the base station (i.e., connection time), the latitude and longitude coordinates (lon,lat) of the base station can be found based on the base station ID.
[0072] Step S200: Divide the base station trajectory sequence into trajectories to obtain the trajectory sequences of each sub-base station.
[0073] To expand the original data volume and achieve data augmentation, a sliding window is used to divide each base station trajectory sequence into sub-base station trajectory sequences. Let the start time of the sliding window be t. begin The termination time is t end =t begin +T, meaning the width of the sliding window is T, the specific implementation process of trajectory division is as follows:
[0074] Step 2.1: Using the base station trajectory sequence P i Connection time of the first base station in China The start time t of the sliding window begin Thus, the trajectory sequence P of the first sub-base station is obtained. i1.
[0075] For example, suppose T = 5 minutes, the connection time of the first base station is 8:05, the connection time of the second base station is 8:07, the connection time of the third base station is 8:08, the connection time of the fourth base station is 8:09, the connection time of the fifth base station is 8:11, and the connection time of the sixth base station is 8:13. Then the sequence trajectory of the first sub-base station includes the information of the first to fourth base stations (all base stations within the range of start time 8:05 to end time 8:10).
[0076] Step 2.2: Slide the sliding window forward H base stations to obtain the trajectory sequence P of the second sub-base station. i2 .
[0077] For example, if T=5 and H=2, then the trajectory sequence of the second sub-base station includes information from the third to the sixth base stations (all base stations within the range of start time 8:08 to end time 8:13), meaning the starting point has moved forward by two base stations. When moving forward by H base stations, the start time corresponds to the connection time of the (1+H)th base station, and the start and end times of the sliding window also change accordingly. Represents the base station trajectory sequence P i The time of the (1+H)th trajectory point (base station). The sliding window continues to slide, starting with the time of the n+Hth trajectory point, where n is the number of base stations in the previous sub-base station trajectory sequence.
[0078] Step 2.3: Slide the sliding window forward H base stations to obtain the trajectory sequence P of the next sub-base station. ij .
[0079] Step 2.4: Following this pattern, the sub-base station trajectory sequences of the base station trajectory sequence are obtained, which can be represented as P. i ={P i1 ,P i2 ,...,P ij ,...}, where P ij Represents the base station trajectory sequence P i The trajectory sequence of the j-th sub-base station.
[0080] Since the width T of the sliding window is a fixed value, and the connection times of each base station in the base station trajectory sequence are not equally spaced, the lengths of the trajectory sequences of each sub-base station may not be equal. When dividing the trajectory, the number of base stations H that the sliding window slides each time can be the same or different.
[0081] When the difference in connection time between the first and last base stations in the trajectory sequence of the last sub-base station is less than T, the trajectory sequence of the last sub-base station is merged with the trajectory sequence of the previous sub-base station.
[0082] When the difference in connection time between the first and last base stations in a certain sub-base station trajectory sequence is less than T / 2, the sub-base station trajectory sequence is merged with the previous sub-base station trajectory sequence.
[0083] Step S300: Spatial encoding is performed on each of the sub-base station trajectory sequences to obtain the corresponding spatial encoding sequence.
[0084] Each base station ID corresponds to the latitude and longitude coordinates (lon, lat) of the base station. The latitude and longitude coordinates of the base station are converted into grid IDs to realize the spatial encoding of the sub-base station trajectory sequence. The specific implementation process is as follows:
[0085] Step 3.1: Determine the latitude and longitude coordinates of each base station based on the base station ID in the sub-base station trajectory sequence.
[0086] Step 3.2: Convert the latitude and longitude coordinates of each base station into a grid ID. The specific calculation formula for converting latitude and longitude coordinates into a grid ID is as follows:
[0087]
[0088]
[0089] Where GID represents the grid ID corresponding to the base station, lon represents the longitude of the base station, lat represents the latitude of the base station, and LON represents the longitude of the base station. max LON min LAT represents the gridded longitude range. max LAT min C represents the range of gridded latitudes. O R represents the number of grid columns. O D represents the number of grid rows. LO D represents the Earth's distance corresponding to a gridded range of longitudes. LA L represents the Earth's distance corresponding to a gridded range of latitudes. G This represents the side length of each grid cell.
[0090] Calculate Earth distances within a latitude and longitude range using distance calculation functions:
[0091] Distance(x1,y1,x2,y2)=2*radius*c(x1,y1,x2,y2)
[0092] Where radius represents the Earth's radius (6,371,000 meters), and Distance() represents the distance calculation function.
[0093] c(x1,y1,x2,y2) is defined as:
[0094]
[0095] The Haversine formula is defined as follows:
[0096]
[0097] therefore:
[0098] D LO = 2 * radius * Distance(d(LON) min ),d(LAT min ),d(LON max ),d(LAT min ))
[0099] D LA = 2 * radius * Distance(d(LON) min ),d(LAT min ),d(LON min ),d(LAT max ))
[0100] in, Indicates the angle Convert to radians:
[0101]
[0102] Step 3.3: The spatial coding sequence is composed of all grid IDs of the sub-base station trajectory sequence.
[0103] Step S400: Perform data augmentation on each of the spatial coding sequences to obtain a dense representation sequence of sparse cellular signaling data in the corresponding sub-base station trajectory sequence.
[0104] In this embodiment, a data augmentation model is used to augment each spatially encoded sequence. The data augmentation model includes a first encoder and a first decoder. The first encoder includes an embedding layer and a GRU network, and the first decoder includes a GRU network, an attention layer, a fully connected layer, and an argmax layer. The specific implementation process of data augmentation is as follows:
[0105] Step 4.1: In order to adapt to the input format of the GRU network, the spatial coding sequence (1*N) is encoded using the embedding layer to obtain a multidimensional output vector (M*N), where N is the length of the spatial coding sequence and M is the dimension of the multidimensional output vector, for example, M = 128 or 64, which can be set according to the requirements.
[0106] Step 4.2: Use the GRU network of the first encoder to further encode the multidimensional output vector to obtain the encoded feature vector (i.e., the final output of GRU).
[0107] Step 4.3: Use the first decoder to decode the encoded feature vector to obtain a dense representation sequence.
[0108] For the embedding layer, the input to the embedding layer is the spatial encoding sequence τ. L ={g1,g2,...,g n1 The output of the GRU network of the first encoder is a latent representation. Here, D represents the dimension of the latent space. The embedding layer transforms the spatial encoding sequence, represented by sparse one-hot vectors, into a dense vector (i.e., the multidimensional output vector is a dense vector). The GRU network of the first encoder encodes the multidimensional output vector to obtain a hidden state sequence {h1, h2, ..., h...}. n1 (Intermediate output of the GRU network). In step i, the dense vector g is input to the GRU network of the first encoder. i (i.e., multidimensional output vector) and hidden state h i-1 This is used to obtain the hidden state h in the i-th step. i The specific formula is as follows:
[0109] h i =GRU(h i-1 ,g i )
[0110] Here, GRU is a non-linear function that, after n1 steps, yields an encoded feature vector z = h. n1 .
[0111] For the first decoder, the encoded feature vector z is used as the input, and the output of the first decoder is the predicted dense feature sequence τ'. L The first decoder consists of a GRU network, a fully connected layer, and an argmax layer. The GRU network is used to obtain the output vector, the fully connected layer is used to reconstruct the dimension of the output vector, and the argmax layer is used to obtain the dense feature sequence.
[0112] Step S500: Input each of the dense representation sequences into the map matching model to obtain the map matching result of the corresponding sub-base station trajectory sequence, and then obtain the map matching result of the base station trajectory sequence.
[0113] In this embodiment, the map matching model includes a second encoder and a second decoder. The second encoder includes an embedding layer and a GRU network, and the second decoder includes a GRU network, an attention layer, a fully connected layer, and an argmax layer. The specific process for obtaining the map matching results of the sub-base station trajectory sequence is as follows:
[0114] Step 5.1: Encode the dense representation sequence using the second encoder to obtain the hidden state sequence (intermediate output of the GRU network);
[0115] Step 5.2: Use the Attention layer to perform a weighted summation of all hidden states in the hidden state sequence to obtain the effective feature vector z. i The specific formula is as follows:
[0116]
[0117] Where j represents the j-th base station of the dense representation sequence, |X| represents the length of the sequence, and h j Let a represent the j-th hidden state obtained by the GRU network of the second encoder. ij Used to measure hidden state h j The importance, or weight, is calculated using the following formula:
[0118]
[0119] Among them, e ij Let h be a fractional function used to measure the hidden state h obtained by the GRU network of the second encoder. j Hidden state h' of the second decoder on the GRU network one step i-1 .
[0120] Step 5.3: Use the GRU network of the second decoder to decode the effective feature vector to obtain a multidimensional vector;
[0121] Step 5.4: Process the multidimensional vector sequentially using a fully connected layer and an argmax layer to obtain the map matching result of the sub-base station trajectory sequence corresponding to the dense representation sequence.
[0122] For the second decoder, the effective feature vector z i The second decoder takes as input and outputs as map matching results of the sub-base station trajectory sequences. The GRU network in the second decoder is used to obtain the output vector, the fully connected layer is used to reconstruct the dimensions of the output vector, and the argmax layer is used to obtain the map matching results of the sub-base station trajectory sequences.
[0123] The map matching result of the base station trajectory sequence is the union of the map matching results of all sub-base station trajectory sequences. For base station trajectory sequence Pi ={P i1 ,P i2 ,...,P ij ,...},P ij Represents the base station trajectory sequence P i The trajectory sequence of the j-th sub-base station has a map matching result of Q. ij The map matching result for the base station trajectory sequence is:
[0124] Q i =Q i1 ∪Q i2 ∪...∪Q ij ∪...
[0125] Data augmentation models and map matching models require training before application. The training sample data in this embodiment comes from data collection conducted by over two hundred volunteers using different modes of transportation to obtain base station trajectory sequences. The dataset includes cellular signaling data and real GPS data from users in a certain province using three different modes of transportation. Specifically, the data from over two hundred users traveling by car totals over two hundred hours and over six thousand kilometers; the data from approximately fifty users traveling by bus totals over ten hours and over two hundred kilometers; and the data from approximately thirty users traveling by subway totals over forty hours and over one thousand kilometers.
[0126] The data labels for the data augmentation model are real GPS trajectory sequences, acquired during data collection activities. The data labels for the map matching model are real road ID sequences, manually labeled.
[0127] Data augmentation models enable the learning of dense representations from real GPS trajectory sequences. The data augmentation model consists of an encoder and a decoder. An encoder is constructed using an embedding layer and a GRU network to obtain a fixed-dimensional feature vector of the spatially encoded sequence. A decoder is constructed using a GRU network, fully connected layers, and an argmax layer to decode the feature vector obtained from the encoder into a dense representation sequence. The cross-entropy loss function is used as the objective function to calculate the loss value between the decoder's output and the real GPS trajectory sequence.
[0128] The input to the data augmentation model is the spatially encoded sequence of the sub-base station trajectory sequence (see steps S200 and S300 for the method of obtaining the spatially encoded sequence), and the data label is the real GPS trajectory sequence. The model training parameters are: batch size = 256, training epochs = 200, embedding encoding size = 512, and GRU hidden state units of encoder and decoder = 512.
[0129] During the training of the data augmentation model, the spatial encoding sequence is input to the encoder to obtain a fixed-dimensional feature vector. The feature vector is then used as the input to the decoder to obtain a dense representation sequence. Finally, the cross-entropy loss between the dense representation sequence and the real GPS trajectory sequence is calculated, and the parameters of the data augmentation model are updated by backpropagation to make the dense representation sequence closer to the real GPS trajectory sequence.
[0130] Map matching models can be used to obtain map matching results for sub-base station trajectory sequences. The map matching model consists of an encoder and a decoder. An encoder is constructed using an embedding layer and a GRU network to obtain a fixed-dimensional feature vector representing the dense representation sequence. The decoder incorporates an attention mechanism, employing an attention layer, a GRU network, a fully connected layer, and an argmax layer to obtain the map matching results. The cross-entropy loss function is used as the objective function to calculate the loss value between the road ID sequence predicted by the map matching model and the actual road ID sequence.
[0131] The input to the map matching model is a dense representation sequence of the sub-base station trajectory sequence (see steps S200-S400 for the method of obtaining the dense representation sequence), and the data label is the real road ID sequence. The model training parameters are: batch size = 64, training epochs = 200, embedding encoding size = 512, GRU hidden state units = 512, and attention units = 10.
[0132] During the training of the map matching model, a dense representation sequence is input to the encoder to obtain a fixed-dimensional feature vector. This feature vector is then used as the input to the Attention layer to obtain an effective feature vector, which is then used as the input to the GRU network to obtain an output vector. The dimension of the output vector is reconstructed through a fully connected layer, and the map matching result of the sub-base station trajectory sequence is obtained through an argmax layer. Finally, the cross-entropy loss between the predicted map matching result and the real road ID sequence is calculated, and the model parameters are updated through backpropagation to make the map matching result closer to the real road ID sequence.
[0133] To better illustrate the embodiments of the present invention, the following experiments were conducted in this embodiment:
[0134] The sample set used in this embodiment of the invention contains approximately 400 base station trajectory sequences, which are divided into training and testing sets for the data augmentation model and the map matching model in an 8:2 ratio. The training parameters are: batch size = 64, epochs = 200, embedding size = 512, and GRU hidden state units = 512. The specific experimental process is as follows:
[0135] I. Performance Indicators
[0136] The embodiments of this invention use precision, recall, F1 score, time, and P-value as evaluation metrics for model performance. These five metrics are widely used in the evaluation of classification models.
[0137] Precision, recall, and F1 score are calculated as follows:
[0138]
[0139]
[0140]
[0141] Among them, Y m For the map matching results, Y g The true value is represented by time; time is the average inference time for map matching of the base station trajectory sequence; the p-value represents the significance of the difference and is used to determine the probability that the observed result is a random result.
[0142] II. Comparative Experiment
[0143] This invention uses Hidden Markov Models (HMM), Gated Recurrent Neural Networks (GRU), and GRU with Attention Mechanism (DeepMM) to compare and evaluate the overall performance of the technical solutions in the embodiments of this invention. The test samples are divided into 5 groups (the specific performance index values for the 5 groups are shown in Figure 2 for the corresponding values of the five black dots for each method). The average value of the 5 groups is used as the performance index value of the test samples. Refer to Figure 2 and... Figure 3 We can draw four conclusions:
[0144] 1) The technical solution of this invention (represented by AMM) is significantly superior to other methods in all metrics. In particular, in terms of F1-Score evaluation, the average F1-Score values of HMM, GRU, and DeepMM are 38.18%, 46.04%, and 66.82%, respectively, while the technical solution of this invention can reach 79.73%, which is an improvement of 41.55%, 33.69%, and 12.91% compared with other models.
[0145] 2) Because the base station trajectory sequence data used has larger errors than GPS data, HMM models, which perform well on GPS data, perform poorly on base station trajectory sequence data. However, due to the advantages of deep learning methods in modeling nonlinear relationships, these models generally outperform HMMs.
[0146] 3) Compared with other methods, the technical solution of the embodiments of the present invention has the lowest deviation, which proves the robustness of the performance of the technical solution of the embodiments of the present invention;
[0147] 4) Observe Figure 2. Figure 3 It can be seen that the technical solution of the present invention presents the best effect. Although the GRU with attention mechanism can obtain effective semantic information of the base station trajectory sequence, due to the sparsity of the base station trajectory sequence, the attention mechanism alone cannot obtain enough contextual information. The technical solution of the present invention performs data augmentation on the base station trajectory sequence, transforming the sparse sequence into a dense feature vector, and thus can obtain better performance.
[0148] III. Ablation Experiment
[0149] To evaluate the effectiveness of different components in the AMM model, this invention verifies the results through ablation experiments. Figure 3 The analysis results for three metrics—Precision, Recall, and P-value—are presented. Here, w / o DA indicates no trajectory segmentation or data augmentation model, w / o TD indicates no trajectory segmentation, and w / o TRE indicates no data augmentation model.
[0150] The results show that removing each component significantly impacts the overall performance of the method. Specifically, when trajectory partitioning and the data augmentation model are removed simultaneously, and when trajectory partitioning and the data augmentation model are removed separately, Precision decreases from 79.78% to 67.78% (removing trajectory partitioning and the data augmentation model), 72.06% (removing trajectory partitioning), and 71.93% (removing the data augmentation model), respectively; Recall decreases from 79.68% to 66.01%, 72.92%, and 72.70%, respectively. The results indicate that when the technical solution of this embodiment lacks dense features, map matching performance significantly decreases, demonstrating the significant impact of dense features on map matching.
[0151] This invention utilizes base station trajectory sequences based on cellular signaling data. Compared to GPS data, cellular signaling data almost completely covers the user's activity range, is continuously available, incurs no additional cost, has lower energy consumption requirements, and is more secure than GPS data due to its passive collection method. This invention divides the base station trajectory sequence into multiple sub-base station trajectory sequences and performs data augmentation on the spatial coding sequence, employing deep learning methods to significantly improve the accuracy of map matching while reducing the difficulty of map matching for sparse cellular signaling data.
[0152] The above description only discloses specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or modifications 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 map matching method based on sparse cellular signaling data, characterized in that, Includes the following steps: The base station trajectory sequence for acquiring cellular signaling data, wherein each base station trajectory sequence is composed of the IDs of each base station connected to by a passenger during a certain mode of transportation and the corresponding connection time; The base station trajectory sequence is divided into trajectory segments to obtain the trajectory sequences of each sub-base station; Each of the sub-base station trajectory sequences is spatially encoded to obtain the corresponding spatially encoded sequence; Data augmentation is performed on each of the spatial coding sequences to obtain a dense representation sequence of sparse cellular signaling data in the corresponding sub-base station trajectory sequence; Each of the dense representation sequences is input into the map matching model to obtain the map matching result of the corresponding sub-base station trajectory sequence, and then the map matching result of the base station trajectory sequence is obtained. The specific implementation process of spatial encoding for each of the aforementioned sub-base station trajectory sequences is as follows: The latitude and longitude coordinates of each base station are determined based on the base station IDs in the sub-base station trajectory sequence; The latitude and longitude coordinates of each base station are converted into a grid ID, and the specific calculation formula is as follows: wherein, represents the grid ID corresponding to the base station, represents the longitude of the base station, represents the latitude of the base station, represents the grid longitude range, represents the grid latitude range, represents the number of grid columns, represents the number of grid rows, represents the earth distance corresponding to the grid longitude range, represents the earth distance corresponding to the grid latitude range, represents the side length of each grid; The spatial coding sequence is composed of all grid IDs of the sub-base station trajectory sequence; The grid longitude range corresponding earth distance The grid latitude range corresponding earth distance The calculation formula is: = = wherein, denotes the earth radius, denotes a distance calculation function, denotes the conversion of an angle into radians, denotes an intermediate function.
2. The map matching method based on sparse cellular signaling data according to claim 1, characterized in that, The specific implementation process of dividing the base station trajectory sequence using a sliding window is as follows: Using the connection time of the first base station in the base station trajectory sequence as the start time of the sliding window, and setting the width of the sliding window to T, the first sub-base station trajectory sequence is obtained. Slide the sliding window forward by H base stations to obtain the trajectory sequence of the next sub-base station; By analogy, the trajectory sequences of each sub-base station of the base station trajectory sequence are obtained.
3. The map matching method based on sparse cellular signaling data according to claim 2, characterized in that, When the difference in connection time between the first and last base stations in the last sub-base station trajectory sequence is less than T, the last sub-base station trajectory sequence is merged with the previous sub-base station trajectory sequence into a single sub-base station trajectory sequence.
4. The map matching method based on sparse cellular signaling data according to claim 2, wherein, When the difference between the connection time of the first and last base station of an intermediate sub-base station trajectory sequence is less than the sub-base station trajectory sequence is merged with the previous sub-base station trajectory sequence into a sub-base station trajectory sequence.
5. The map matching method based on sparse cellular signaling data according to any one of claims 1-4, characterized in that, Data augmentation is performed on each of the spatially encoded sequences using a data augmentation model. The data augmentation model includes a first encoder and a first decoder. The first encoder includes an embedding layer and a GRU network, and the first decoder includes a GRU network, an attention layer, a fully connected layer, and an argmax layer. The specific implementation process of the data augmentation is as follows: The spatial coding sequence is encoded using an embedding layer to obtain a multidimensional output vector; The multidimensional output vector is further encoded using the GRU network of the first encoder to obtain the encoded feature vector; The encoded feature vector is decoded using the first decoder to obtain a dense representation sequence.
6. The map matching method based on sparse cellular signaling data according to any one of claims 1-4, characterized in that, The map matching model includes a second encoder and a second decoder. The second encoder includes an embedding layer and a GRU network, and the second decoder includes a GRU network, an attention layer, a fully connected layer, and an argmax layer. The specific process for obtaining the map matching results of the sub-base station trajectory sequence is as follows: The dense representation sequence is encoded using a second encoder to obtain a hidden state sequence; The effective feature vector is obtained by using an Attention layer to perform a weighted summation of all hidden states in the hidden state sequence. The effective feature vector is decoded using the GRU network of the second decoder to obtain a multidimensional vector; The multi-dimensional vector is processed in sequence by using a full connection layer and an argmax layer to obtain a map matching result of a sub-base station trajectory sequence corresponding to the dense representation sequence.
7. An electronic device, comprising: The device comprises: a memory for storing a computer program; a processor for implementing the map matching method based on sparse cellular signaling data according to any one of claims 1-6 when executing the computer program.
8. A computer-readable storage medium, characterized in that: The computer program is stored on the computer readable storage medium, and the computer program is executed by the processor to implement the map matching method based on sparse cellular signaling data according to any one of claims 1-6.