Trajectory fusion method for vehicle and ship and storage medium

Abstract

The application discloses a kind of track fusion methods of vehicle, ship: first, the measurement data and real data of target track are preprocessed, the size of data graph is determined by stepwise regression algorithm, then one-dimensional track point data is assembled into two-dimensional track data graph, to achieve the purpose of batch processing measurement data, also can capture the motion correlation between track point data.Second, the interaction between features and the spatial features between data are mined through CIN network structure, then the input of CIN is taken as the input of Informer module, the temporal features between data are mined through Informer module, and finally the fusion result is obtained through LSTM module.Finally, the model is trained and the model file is obtained, and the track fusion task is completed on the validation set.The application can quickly process large track measurement data and accurately calculate the track fusion result, improve the accuracy of track fusion.

Description

Technical Field

[0001] This invention relates to the field of multi-source information fusion technology, specifically to a method and storage medium for fusion of vehicle and ship tracks. Background Technology

[0002] Vehicle and ship track fusion is essentially the fusion of state estimates from multiple sensors. With multiple reconnaissance devices and multiple sensors, target track measurement data is characterized by large data volume and small data deviation. How to quickly process this massive amount of measurement data and accurately calculate the track fusion results, thereby improving the accuracy of track fusion, is a crucial issue in the current field of multi-source information fusion technology. Summary of the Invention

[0003] This invention provides a method for track fusion of vehicles and ships, the main purpose of which is to quickly process massive track measurement data and accurately calculate track fusion results, thereby improving the accuracy of track fusion.

[0004] First, the measured and real data of the target track are preprocessed. A stepwise regression algorithm is used to determine the size of the data map, and then the one-dimensional track point data is assembled into a two-dimensional track data map to achieve batch processing of the measured data and capture the motion correlations between track points. Second, the CIN network structure is used to mine the interaction relationships between features and the spatial features between data points. The input of CIN is then used as the input of the Informer module to mine the temporal features between data points. Finally, an LSTM module is used for fine-tuning to obtain the fusion result. Finally, the model is trained and the model file is obtained, and the track fusion task is completed on the validation set.

[0005] The vehicle and ship trajectory fusion method provided by this invention includes the following steps:

[0006] (1) Given the target's original track coordinate dataset: including the longitude coordinates, latitude coordinates and the target's true coordinates obtained from the target observation (the target's true coordinates are the data labels used for model training), divided into training set, validation set and test set.

[0007] (2) Preprocess the original target track coordinate data to construct a track data map: Track coordinate data is data collected by sensors in chronological order and at a certain sampling frequency, including multi-source observation coordinate data and target real coordinate data. Track data maps are constructed for longitude and latitude coordinates respectively. The size of the track data map is s*n, where s represents the number of data records in a track data map, which is determined by the stepwise regression algorithm, and n represents the number of features, i.e., the number of tracks participating in the fusion, to obtain the size of the track data map, thereby assembling the one-dimensional track data into a two-dimensional track data map.

[0008] (3) The constructed two-dimensional trajectory data map is first input into the CIN module. CIN is a compressed interactive network that allows for the full combination of features in the data through vector-level feature interactions. CIN multiplies the feature matrix output from each hidden layer with the original matrix through cross operations to obtain a two-dimensional result. Multiple multiplications are stacked to form a three-dimensional feature matrix. Then, the three-dimensional feature matrix obtained by cross operations is compressed by convolution and pooling operations to transform it into a two-dimensional feature matrix. Finally, one-dimensional features are obtained through Sum Pooling. Finally, the results obtained from each Sum Pooling layer are concatenated to obtain the final output.

[0009] (4) The output of the CIN module is used as the input of the Informer+LSTM module to further mine the temporal features between the data and fine-tune the results to obtain the final fusion result. The Informer module can effectively mine the temporal relationships between the data and effectively reduce the time complexity and memory usage compared with the Transformer, thus improving the inference speed. The LSTM module takes the output of the Informer module as input, fine-tunes it, and finally uses the output of the LSTM layer as the final fusion result. The model uses MSE as the loss function. By calculating the difference between the output result and the label, the model can improve its learning ability through iterative training, thereby improving the model's representation ability.

[0010] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described track fusion method. Attached Figure Description

[0011] Figure 1 This is a flowchart of the target trajectory fusion process of the present invention;

[0012] Figure 2 The flowchart for determining the data graph size using the stepwise regression algorithm involved in this invention is shown below.

[0013] Figure 3 This is an assembly effect diagram of the trajectory data map in this invention;

[0014] Figure 4 This is an example diagram of track data map label processing in this invention. Detailed Implementation

[0015] The embodiments and technical effects of the present invention are described in detail below.

[0016] I. The assembly method for track data charts mainly includes the following steps:

[0017] (1) Standardization of the target original dataset: For the given original dataset, and considering the high accuracy requirements of the results of track fusion, since longitude and latitude have certain differences in range values, the data is standardized to eliminate the influence of differences in the units and value ranges between features, so as to improve the effect of subsequent model training.

[0018] (2) Handling the size of the target track data map: Track data usually has a temporal sequence. Assembling one-dimensional data records into two-dimensional track data maps can capture the state and trend of track data on the time axis, and also serve as a batch processing function for track data. Since longitude and latitude in track data do not have a direct correlation on the time axis, track data maps are constructed for the longitude and latitude coordinates of the measured data respectively. Each track data map contains past and current data of each feature. The number of records contained in a single data map reflects the temporal correlation of the features. If the number of records is too small, it is not possible to fully capture the dynamic change trend of the features; if the number of records is too large, the number of track data maps will be too small, resulting in the model not having enough examples to learn and distinguish features. Therefore, the number of records in the track data map is determined by stepwise regression algorithm, and the interrelationship between the number of records is fully considered.

[0019] (3) Processing of target track data map labels: For multiple track longitude or latitude data at the same time, there is only one longitude or latitude fusion result. Since the calculation of track fusion result belongs to regression problem, the label of each track data map is also a two-dimensional data format, that is, each data record is a combination of its own fusion result.

[0020] The specific steps for determining the size of the track data map using the stepwise regression algorithm in step (2) above are as follows:

[0021] (1) Take the initial size of the data graph as s, and use the stepwise regression algorithm to calculate the important factors of 1-s records.

[0022] (2) Add new records to the initial data graph, recalculate the importance factors of 1-s+1 records, and obtain the importance factor values ​​of the new records.

[0023] (3) Repeat step (2) until the important factor value of the new record is less than the specified threshold. Obtain the first updated s value, denoted as s1.

[0024] (4) Starting from record s+1, repeat steps (1)-(3) until all records have been counted. Average all the updated s values ​​(s1, s2, ..., sn) to get the final value of the track data map size s.

[0025] II. The main steps of implementing target track fusion are as follows:

[0026] (1) Given the target's original track coordinate dataset, which includes the longitude coordinates, latitude coordinates and the target's true coordinates obtained from the target observation (the target's true coordinates are the data labels used for model training), it is divided into training set, validation set and test set.

[0027] (2) Preprocess the original track coordinate data of the target, determine the number of track data map records by stepwise regression algorithm, i.e. the size value of the data map, and construct the track data map.

[0028] (3) The target trajectory data is time-series data collected by the sensor at a certain frequency. It has a certain trend of change and contains both spatial and temporal features. Therefore, a CIL (CIN+Informer+LSTM) network model structure was constructed to learn this feature.

[0029] (4) Load the training set and validation set from the original target track coordinate data into the constructed CIL network model for training and validation. Since track data maps are constructed for the longitude and latitude features of the track data respectively, there are two training result files for the convolutional neural network, namely the model training result files for the track longitude data map and the track latitude data map.

[0030] (5) Encapsulate the test set track coordinate data into a data graph, call the two trained model files, calculate the track fusion results for track longitude and track latitude respectively, complete the track fusion task, and evaluate the generalization ability, fusion result accuracy and processing efficiency of the model.

[0031] In step (3) above, the structure of the CIL network model is constructed as follows:

[0032] (A) The first layer of the model is the CIN module, and the model input is the constructed track data map. CIN is a compressed interactive network that enables the full combination of features in the data through vector-level feature interactions, and transforms the two-dimensional input into a one-dimensional output through cross-linking, compression, and concatenation.

[0033] (B) The second layer of the model is the Informer module, which receives the output of CIN as input and mines the temporal features between the data. Compared with Transformer, the Informer module effectively reduces time complexity and memory usage, and improves inference speed.

[0034] (C) The third layer of the model is an LSTM module, which receives the output of the Informer module as input, fine-tunes it, and finally uses the output of the LSTM layer as the final fusion result.

[0035] The model uses MSE as the loss function. By calculating the difference between the output and the label, the model improves its learning ability through iterative training, thereby enhancing its representation ability.

[0036] In addition, the CIL network model training and testing process in steps (4) and (5) above specifically includes the following steps:

[0037] (1) Input a batch of track data records;

[0038] (2) Read data records and perform data preprocessing, including data standardization, etc.;

[0039] (3) Use the stepwise regression algorithm to determine the value of the size s of the data plot;

[0040] (4) Assemble the track data map according to parameters such as the size s*n of the track data map;

[0041] (5) Read in the training trajectory data map and generate initial network parameters in a random manner;

[0042] (6) If the number of iterations Epoch has not reached the set threshold Threshold, continue with step (6); otherwise, stop training.

[0043] (7) First, perform forward propagation to obtain the error between the current model prediction and the true value. Then, propagate the error through backpropagation to update the network weights;

[0044] (8) Repeat the above steps, and through repeated iterative training, continuously reduce the error and optimize the network weights. Stop training when the error is less than the threshold.

[0045] (9) Output the model to a local file;

[0046] (10) Read the model file and input the test track data diagram;

[0047] (11) Obtain the prediction results.

[0048] III. Technical Effects:

[0049] The final fusion result of this invention has a 10*1 dimension. In the target track fusion task, the track fusion CIL network model provided by this invention, compared with traditional track fusion algorithms, exhibits the smallest mean square error (MSE), mean absolute error (MAE), and root mean square error (RMSE) on a given track measurement data test set, demonstrating the accuracy of this invention in track fusion. Furthermore, given a large volume of track measurement data test set, this invention achieves the lowest processing time for batch measurement data fusion, indicating its high efficiency in batch track fusion processing.

Claims

1. A method for fusion of vehicle and ship tracks, characterized in that, The method includes the following steps: (1) Given the target's original track coordinate dataset, including the longitude coordinates, latitude coordinates and the target's true coordinates obtained from the target observation, it is divided into training set, validation set and test set; (2) Preprocess the original track coordinate data of the target, determine the number of track data map records by stepwise regression algorithm, that is, the size value of the data map, and construct a two-dimensional track data map; (3) Load the training set and validation set from the original target track coordinate data into the constructed CIL network model for training and validation, and obtain two model training result files for the track longitude data map and the track latitude data map; (4) Encapsulate the test set track coordinate data into a data graph, call the training result files of the two trained models, calculate the track fusion results for track longitude and track latitude respectively, complete the track fusion task, and evaluate the generalization ability, fusion result accuracy and processing efficiency of the model. The structure of the CIL network model is as follows: (A) The first layer of the model is the CIN module, and the model input is the constructed two-dimensional track data map; (B) The second layer of the model is the Informer module, which receives the output of CIN as input and mines the temporal features between data through the Informer module; (C) The third layer of the model is an LSTM module, which receives the output of the Informer module as input, fine-tunes it, and finally the output of the LSTM layer is used as the final target track fusion result.

2. The trajectory fusion method according to claim 1, characterized in that, The specific steps of the CIL-based network model training and testing process in steps (3) and (4) are as follows: (1) Input a batch of track data records; (2) Read the data records and perform data preprocessing, including data standardization; (3) Use the stepwise regression algorithm to determine the value of the size s of the data map, where s represents the number of data records in a two-dimensional track data map; (4) Assemble a two-dimensional track data map according to the size s*n of the track data map, where n represents the number of tracks participating in the fusion; (5) Read in the training trajectory data map and generate initial network parameters in a random manner; (6) If the number of iterations Epoch has not reached the set threshold Threshold, continue with step (6); otherwise, stop training. (7) First, perform forward propagation to obtain the error between the current model prediction and the true value; then, propagate the error through backpropagation to update the network weights. (8) Repeat the above steps, and through repeated iterative training, continuously reduce the error and optimize the network weights. Stop training when the error is less than the threshold. (9) Output the model to a local file; (10) Read the model file and input the test two-dimensional track data map; (11) Obtain the prediction results.

3. A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the track fusion method of claim 1 or 2.

Images