Distributed heterogeneous navigation information fusion cooperative positioning method based on cnn

Abstract

The application provides a distributed heterogeneous navigation information fusion cooperative positioning method based on a CNN, and belongs to the technical field of navigation positioning. The method comprises the following steps: constructing a system model of a vehicle distributed cooperative positioning system; acquiring original multi-source heterogeneous navigation data of each vehicle respectively, and performing pretreatment to obtain corresponding standard multi-source heterogeneous navigation data; arranging all the standard multi-source heterogeneous navigation data according to time sequence and the type of sensors to obtain a multi-source cooperative feature tensor; constructing a convolutional neural network, and sequentially performing local feature extraction, cooperative feature fusion and cooperative positioning processing on the multi-source cooperative feature tensor by using the convolutional neural network to obtain the cooperative positioning result of each vehicle. The application can improve the positioning accuracy of the vehicle distributed cooperative positioning system in a complex traffic environment.

Description

Technical Field

[0001] This application relates to the field of navigation and positioning technology, and in particular to a distributed heterogeneous navigation information fusion and collaborative positioning method based on CNN. Background Technology

[0002] With the rapid development of the Internet of Things (IoT) and vehicle-to-everything (V2X) technologies, vehicles have evolved into intelligent mobile terminals integrating sensing, communication, and computing. In intelligent transportation systems, high-precision and highly reliable navigation and positioning capabilities are the core foundation for achieving autonomous driving, cooperative obstacle avoidance, and path planning.

[0003] Currently, mainstream navigation and positioning solutions primarily rely on Global Navigation Satellite Systems (GNSS) and Inertial Navigation Systems (INS). While GNSS provides long-term, absolute positioning information, its signals are susceptible to multipath effects or physical obstructions in complex environments such as urban canyons, tunnels, and overpasses, leading to signal interruptions or a sharp decline in positioning accuracy. INS, while offering advantages such as short-term accuracy, high precision, and complete autonomy, suffers from inherent sensor drift characteristics that cause errors to accumulate over time, making it difficult to maintain high accuracy independently for extended periods without external correction. To address this, some research has attempted to introduce auxiliary sensing technologies such as vision, millimeter-wave radar, or lidar to enhance the system's environmental adaptability. However, these sensing technologies are significantly affected by lighting conditions and adverse weather conditions such as rain, snow, and fog, lacking robustness and making them unsuitable as stable and reliable independent navigation sources.

[0004] In related technologies, cooperative positioning technology shares vehicle location, speed, and status information through dedicated short-range communication or cellular networks within the vehicle group, effectively suppressing positioning errors of individual vehicles by leveraging the complementary information of multiple nodes within the vehicle group. This has become a research hotspot in the field of intelligent transportation. However, in-depth analysis reveals the following shortcomings in existing cooperative positioning methods:

[0005] First, there are inherent flaws in the result-based fusion model. Most existing technologies use the final positioning results output by each navigation source as the fusion object. This fusion model only combines data at the high-level result level. Once a navigation source produces single-axis data anomalies or outliers due to sudden environmental changes or its own malfunction, the error will directly enter the fusion process and spread and amplify rapidly in the network through the topology of the vehicle network, resulting in a sharp decline in the overall fusion accuracy and reliability.

[0006] Secondly, there is insufficient capability to process heterogeneous data. Different navigation sources exhibit significant heterogeneity in sampling rate, measurement quality, and spatial characteristics. Traditional cooperative localization methods based on extended Kalman filtering or factor graphs are complex to model and have poor real-time performance when processing such nonlinear, high-dimensional heterogeneous data.

[0007] Finally, the shortcomings of traditional neural network models. Existing cooperative localization methods based on backpropagation neural network models are prone to getting trapped in local optima when dealing with large-scale time-series sequences and multi-dimensional features. They are also sensitive to redundant noise, have slow training speeds, and limited generalization capabilities. Especially in scenarios where vehicle network topology changes frequently, cooperative localization methods based on backpropagation neural network models struggle to quickly and adaptively adjust the fusion strategy, and their accuracy and robustness cannot meet the stringent requirements of high-reliability navigation and positioning for autonomous driving. Therefore, it is necessary to propose a solution to improve one or more of the problems existing in the aforementioned related technical solutions.

[0008] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0009] This application provides a CNN-based distributed heterogeneous navigation information fusion cooperative localization method, which includes the following steps:

[0010] A system model for a vehicle distributed cooperative positioning system is constructed. The system model includes a distributed cooperative network composed of multiple vehicles. Each vehicle is equipped with multiple sensors, and each vehicle is regarded as an independent network node.

[0011] The original multi-source heterogeneous navigation data of each vehicle is acquired separately, and each original multi-source heterogeneous navigation data is preprocessed to obtain the corresponding standard multi-source heterogeneous navigation data.

[0012] All standard multi-source heterogeneous navigation data are arranged in chronological order and according to sensor type to obtain the multi-source cooperative feature tensor;

[0013] A convolutional neural network is constructed, and the convolutional neural network is used to sequentially perform local feature extraction, collaborative feature fusion, and collaborative localization processing on the multi-source collaborative feature tensor to obtain the collaborative localization results for each vehicle.

[0014] Furthermore, the types of sensors include global satellite navigation systems and inertial navigation systems.

[0015] Furthermore, the preprocessing includes performing time alignment, data repair, and normalization on each original multi-source heterogeneous navigation data in sequence to obtain the corresponding standard multi-source heterogeneous navigation data.

[0016] Each set of raw multi-source heterogeneous navigation data includes raw vehicle position data measured by the global satellite navigation system and raw vehicle state data measured by the inertial navigation system;

[0017] Each standard multi-source heterogeneous navigation data includes standard vehicle position data and standard vehicle status data.

[0018] Furthermore, the expression for the multi-source collaborative feature tensor is:

[0019] (1)

[0020] in, Represents the multi-source collaborative feature tensor. express 3D real space, This represents the total length of the time window for the multi-source collaborative feature tensor. This represents the total number of vehicles in the system model. This represents the number of all sensors in the system model. Indicates the first Time of the first Standard multi-source heterogeneous navigation data for each vehicle , Indicates the first Time of the first Standard vehicle location data for each vehicle Indicates the first Time of the first Standard vehicle speed data for each vehicle Indicates the first Time of the first Standard vehicle acceleration data for each vehicle express 3D real space.

[0021] Furthermore, the convolutional neural network includes a local collaborative feature extraction unit, a distributed collaborative feature fusion unit, and a collaborative localization unit connected in sequence; the local collaborative feature extraction unit includes a feature extraction layer, a cross-channel convolutional fusion layer, a max pooling layer, and a flattening layer connected in sequence.

[0022] Furthermore, the steps of using a convolutional neural network to sequentially perform local feature extraction, collaborative feature fusion, and collaborative localization processing on the multi-source collaborative feature tensor to obtain the collaborative localization results for each vehicle include:

[0023] The local collaborative feature extraction unit is used to extract local features from the multi-source collaborative feature tensor to obtain the local collaborative feature vector for each vehicle.

[0024] The following operations are performed sequentially using the distributed collaborative feature fusion unit:

[0025] Each vehicle shares its local collaborative feature vector with all other neighboring vehicles within its communication range, and performs nonlinear aggregation of all received local collaborative feature vectors with its own local collaborative feature vector to obtain the corresponding vehicle's fused collaborative feature vector.

[0026] The steps of obtaining the fusion collaborative feature vector are iteratively performed. In each iteration, each vehicle updates its local collaborative feature vector according to the information injection coefficient of this iteration until the changes in the local collaborative feature vectors of all vehicles converge or the maximum number of iterations is reached, and the global collaborative feature vector of each vehicle is obtained.

[0027] The cooperative localization unit is used to perform cooperative localization processing on each global cooperative feature vector to obtain the corresponding vehicle cooperative localization result.

[0028] Furthermore, the steps of using the local collaborative feature extraction unit to extract local features from the multi-source collaborative feature tensor to obtain the local collaborative feature vector for each vehicle include:

[0029] The feature extraction layer performs multi-layer one-dimensional convolution operation on the multi-source collaborative feature tensor to extract the temporal dynamic feature map of each vehicle. Each temporal dynamic feature map includes multiple dynamic feature values ​​arranged in time order.

[0030] The expression for dynamic eigenvalues ​​is:

[0031] (2)

[0032] in, Indicates the first Time of the first The dynamic characteristic values ​​of each vehicle This represents the total length of the time window for the one-dimensional convolution kernel. Indicates the first The weights of a one-dimensional convolution kernel, Indicates the first Time of the first The acceleration of each vehicle, Indicates the first bias term, the... The temporal dynamic feature map of the vehicle is represented as follows: , , Indicates the first Time of the first The dynamic characteristic values ​​of each vehicle This represents the total length of the time window for the multi-source collaborative feature tensor.

[0033] By using a cross-channel convolutional fusion layer to perform cross-channel convolution operations on all temporal dynamic feature maps, a multi-layer spatial correlation feature map representing the relationship between all sensors is obtained.

[0034] The expression for the spatial association feature map is:

[0035] (3)

[0036] in, Indicates the first Layer spatial correlation feature map Indicates the first Cross-channel convolution operation for each sensor;

[0037] Max pooling is applied to each spatial correlation feature map using a max pooling layer to obtain the corresponding pooling co-feature map. Each pooling co-feature map includes multiple pooling feature values.

[0038] The expression for pooling eigenvalues ​​is:

[0039] (4)

[0040] in, Indicates the first The first layer in the spatial correlation feature graph Time, Number The vehicle, the Pooling feature values ​​corresponding to each sensor This indicates taking the maximum value. Indicates a pooled window. Indicates the first The first layer in the spatial correlation feature graph Time, Number The vehicle, the Pooling feature values ​​corresponding to each sensor This represents the time offset within the pooling window. This represents the vehicle offset within the pooling window, the first... The pooling collaborative feature map corresponding to the layer spatial correlation feature map is represented as follows: ;

[0041] The dimensionality of each pooled collaborative feature map is flattened using a flattening layer to obtain the corresponding local collaborative feature vector of the vehicle.

[0042] The expression for the local collaborative feature vector is:

[0043] (5)

[0044] in, Indicates the first Local collaborative feature vectors of individual vehicles This indicates the flattening process of the dimensions. express 3D real space, This represents the total length of the time window for pooling collaborative feature maps. This represents the total number of vehicles in the system model. This represents the total number of sensors in the system model.

[0045] Furthermore, the steps of each vehicle sharing its local cooperative feature vector with all other neighboring vehicles within its communication range, and nonlinearly aggregating all received local cooperative feature vectors with its own local cooperative feature vector to obtain the corresponding vehicle's fused cooperative feature vector include:

[0046] Each vehicle sends its local collaborative feature vector to all other neighboring vehicles within its communication range via the inter-vehicle communication link;

[0047] Each vehicle receives local cooperative feature vectors sent by all its neighboring vehicles, thus obtaining the corresponding set of neighborhood cooperative feature vectors for that vehicle.

[0048] For each vehicle, a nonlinear transformation is performed on each local cooperative feature vector in the neighborhood cooperative feature vector set to obtain the corresponding nonlinear feature vector;

[0049] The expression for the nonlinear eigenvector is:

[0050] (6)

[0051] in, Indicates the first In the neighborhood collaborative feature vector set of vehicle n, the th The nonlinear feature vector corresponding to the local cooperative feature vector of each neighboring vehicle. Represents a non-linear activation function. Indicates the first In the neighborhood collaborative feature vector set of vehicle n, the th Local collaborative feature vectors of neighboring vehicles;

[0052] Calculate the mean vector and covariance matrix of all nonlinear eigenvectors for each vehicle in turn;

[0053] The expression for the mean vector of all nonlinear eigenvectors corresponding to the vehicle is:

[0054] (7)

[0055] in, Indicates the first The mean vector of all nonlinear eigenvectors corresponding to each vehicle. Indicates the first The number of all neighboring vehicles of a vehicle;

[0056] The expression for the covariance matrix of all nonlinear eigenvectors corresponding to the vehicle is:

[0057] (8)

[0058] in, Indicates the first The covariance matrix of all nonlinear eigenvectors corresponding to each vehicle Indicates transpose;

[0059] The Mahalanobis distance between the corresponding vehicle and each neighboring vehicle within its communication range is calculated by using the mean vector and covariance matrix of all nonlinear eigenvectors of each vehicle.

[0060] The expression for Mahalanobis distance is:

[0061] (9)

[0062] in, Indicates the first The vehicle within its own communication range Mahalanobis distance between neighboring vehicles;

[0063] Each vehicle constructs a corresponding confidence weight using the Mahalanobis distance between itself and each neighboring vehicle within its communication range, and normalizes each confidence weight to obtain the corresponding normalized confidence weight.

[0064] Each normalized confidence weight is used to perform weighted fusion of the local collaborative feature vectors in the neighborhood collaborative feature vector set of the corresponding vehicle to obtain the fused collaborative feature vector of the corresponding vehicle.

[0065] The expression for the fused collaborative feature vector is:

[0066] (10)

[0067] in, Indicates the first The fusion and collaborative feature vector of each vehicle Indicates the first The vehicle within its own communication range Normalized confidence weights corresponding to each neighbor vehicle , Indicates the first The vehicle within its own communication range The credibility weight of each neighbor's vehicle , This represents the regularization term.

[0068] Further, the step of iteratively obtaining the fused collaborative feature vector involves each vehicle updating its local collaborative feature vector based on the information injection coefficient of that iteration during each iteration, until the changes in the local collaborative feature vectors of all vehicles converge or the maximum number of iterations is reached. The steps for obtaining the global collaborative feature vector for each vehicle include:

[0069] Set the convergence threshold for the change amount and the maximum number of iterations, and set the information injection coefficient for each iteration. The convergence threshold for the change amount is... express;

[0070] The expression for the change in the local cooperative feature vector of the vehicle is:

[0071] (11)

[0072] in, Indicates the first During the nth iteration The change in the local collaborative feature vector of each vehicle Indicates the first During the nth iteration Local collaborative feature vectors of individual vehicles Indicates the first During the nth iteration Local collaborative feature vectors of individual vehicles , Indicates the first Information injection coefficients at the next iteration Indicates the first The number of all neighboring vehicles of a given vehicle. Indicates the first During the nth iteration In the neighborhood collaborative feature vector set of vehicle n, the th Local collaborative feature vectors of neighboring vehicles;

[0073] when or When the maximum number of iterations is reached, the fusion collaborative feature vector of the corresponding vehicle is used as the corresponding global collaborative feature vector.

[0074] Furthermore, the steps of using the cooperative localization unit to perform cooperative localization processing on each global cooperative feature vector to obtain the corresponding vehicle cooperative localization result include:

[0075] The cooperative localization unit includes a location coding network, which includes a fully connected layer;

[0076] The global collaborative feature vector of each vehicle is input into the location encoding network, and the actual location information of the corresponding vehicle is obtained by mapping using a fully connected layer;

[0077] The expression for the vehicle's actual location information is:

[0078] (12)

[0079] in, Indicates the first Time of the first The actual location information of each vehicle. Represents the regression weight matrix. Indicates the first Global collaborative feature vectors of individual vehicles Indicates the second bias term;

[0080] The observed positioning information of each vehicle at the current moment is used to verify the corresponding actual positioning information to obtain the corresponding cooperative positioning information;

[0081] The expression for the vehicle's cooperative positioning information is:

[0082] (13)

[0083] in, Indicates the first Time of the first Coordinated location information of individual vehicles Indicates the correction factor. Indicates the first Time of the first The observation and positioning information of each vehicle.

[0084] This application provides a CNN-based distributed heterogeneous navigation information fusion cooperative localization method, which has at least the following beneficial effects:

[0085] (1) This application constructs a system model of a vehicle distributed cooperative positioning system, obtains the original multi-source heterogeneous navigation data of each vehicle, performs preprocessing, and arranges all the obtained standard multi-source heterogeneous navigation data into a multi-source cooperative feature tensor according to the time order and sensor type, so as to unify the differences in sampling frequency, data type and accuracy of different sensors, and provide a data input basis for subsequent deep learning processing;

[0086] (2) This application utilizes the local cooperative feature extraction unit in a convolutional neural network (CNN) to extract the temporal dynamic feature map of each vehicle in the time dimension and extract the spatial correlation feature map in the sensor dimension, thereby obtaining the local cooperative feature vector of each vehicle. This process does not rely on manually designed motion models or noise statistical assumptions, but automatically learns deep fusion features in a data-driven manner, thereby being able to capture cooperative motion patterns in complex dynamic environments more accurately;

[0087] (3) This application utilizes the distributed collaborative feature fusion unit of a convolutional neural network to iteratively obtain the fused collaborative feature vector of each vehicle, and obtains the global collaborative feature vector of the corresponding vehicle when the change in the fused collaborative feature vector is less than the convergence threshold of the change or when the maximum number of iterations is reached. Since the communication content is features rather than raw data, this can reduce the communication bandwidth requirement, protect the privacy of the local information of each vehicle, and dynamically adjust the confidence weight according to the features of neighboring vehicles during the weighted fusion process, so that when individual vehicles experience sensor abnormalities or communication interruptions, the influence of abnormal features can be effectively suppressed, thereby significantly improving the robustness of the entire vehicle distributed collaborative positioning system under non-ideal conditions such as signal occlusion and sensor drift;

[0088] (4) This application utilizes the cooperative localization unit of a convolutional neural network to input the global cooperative feature vector of each vehicle into the location coding network, and uses a fully connected layer to map and obtain the actual positioning information of the corresponding vehicle. Then, the observed positioning information of each vehicle at the current time is used to perform feedback verification on the corresponding actual positioning information to obtain the corresponding cooperative positioning information. This achieves end-to-end positioning regression, avoiding the problem of gradual error accumulation in traditional filtering or integration methods. Moreover, the entire vehicle distributed cooperative localization system supports parallel processing of multiple vehicles. Each vehicle independently completes feature extraction and local fusion, and the computational load is balanced, which can meet the real-time requirements of the Internet of Vehicles. Thus, while significantly improving positioning accuracy, it can effectively reduce the interference of data anomalies of individual vehicles on the overall positioning results. Attached Figure Description

[0089] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0090] Figure 1This illustration shows the steps of a CNN-based distributed heterogeneous navigation information fusion cooperative localization method in an exemplary embodiment of this application.

[0091] Figure 2 This illustration shows a flowchart of vehicle cooperative localization using a convolutional neural network in an exemplary embodiment of this application.

[0092] Figure 3 A schematic diagram showing the global trajectories of all vehicles in a simulation experiment of an exemplary embodiment of this application;

[0093] Figure 4 This diagram illustrates the thermal distribution of vehicle positioning error in a simulation experiment of an exemplary embodiment of this application.

[0094] Figure 5 This diagram illustrates a comparison of positioning errors obtained under ideal conditions using different prediction error methods in a simulation experiment of an exemplary embodiment of this application.

[0095] Figure 6 This diagram illustrates a comparison of positioning errors obtained using different prediction error methods under abrupt changes in a simulation experiment of an exemplary embodiment of this application.

[0096] Figure 7 This diagram illustrates a comparison of positioning errors obtained under occlusion conditions using different prediction error methods in a simulation experiment of an exemplary embodiment of this application. Detailed Implementation

[0097] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0098] Furthermore, the accompanying drawings are merely illustrative of this application and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0099] In the Internet of Vehicles (IoV), vehicle clusters typically exhibit highly free and decentralized movement characteristics. Each vehicle not only independently performs its own navigation and motion control tasks but also interacts with surrounding vehicles. This decentralized clustering characteristic means that each vehicle receives a large amount of heterogeneous navigation information from its own sensors and those of surrounding vehicles during its movement, including data from GPS, inertial navigation systems, vehicle radar, lidar, and visual sensors. Each sensor provides different data types, data precision, and sampling frequencies, forming a complex multi-source, multi-dimensional information network. To achieve high-precision and high-efficiency cooperative positioning, it is necessary to uniformly organize and standardize this heterogeneous, multi-source, and multi-temporal data, and establish a tensor structure suitable for deep learning processing.

[0100] Based on this, this application proposes a distributed heterogeneous navigation information fusion cooperative localization method based on CNN, such as... Figure 1 As shown, the method may include the following steps:

[0101] Step S101: Construct a system model for a vehicle distributed cooperative positioning system. This system model includes a distributed cooperative network composed of multiple vehicles. Each vehicle is equipped with multiple sensors, and each vehicle is regarded as an independent network node.

[0102] Furthermore, the types of sensors configured on each vehicle correspond to the raw, multi-source heterogeneous navigation data received. The types of sensors configured on each vehicle include global navigation satellite systems (GPS) and inertial navigation systems (INS), and may also include other types such as radio navigation systems to acquire a wider variety of multi-source heterogeneous navigation data. Here, the radio navigation system is primarily used to measure relative distance data between vehicles.

[0103] Furthermore, each set of raw multi-source heterogeneous navigation data includes raw vehicle position data measured by a global satellite navigation system and raw vehicle state data measured by an inertial navigation system. The raw vehicle state data includes raw vehicle speed data and raw vehicle acceleration data, etc. Each set of standard multi-source heterogeneous navigation data includes standard vehicle position data and standard vehicle state data.

[0104] Step S102: Obtain the original multi-source heterogeneous navigation data for each vehicle, and preprocess each original multi-source heterogeneous navigation data to obtain the corresponding standard multi-source heterogeneous navigation data.

[0105] Furthermore, the preprocessing includes sequentially performing time alignment, data repair, and normalization on each original multi-source heterogeneous navigation data set to obtain the corresponding standard multi-source heterogeneous navigation data. These preprocessing procedures are conventional techniques in the field and will not be elaborated upon in this embodiment.

[0106] Step S103: Arrange all standard multi-source heterogeneous navigation data according to time order and sensor type to obtain multi-source cooperative feature tensor.

[0107] Furthermore, the expression for the multi-source collaborative feature tensor is:

[0108] (1)

[0109] in, Represents the multi-source collaborative feature tensor. express 3D real space, This represents the total length of the time window for the multi-source collaborative feature tensor. This represents the total number of vehicles in the system model. This represents the number of all sensors in the system model. Indicates the first Time of the first Standard multi-source heterogeneous navigation data for each vehicle , Indicates the first Time of the first Standard vehicle location data for each vehicle Indicates the first Time of the first Standard vehicle speed data for each vehicle Indicates the first Time of the first Standard vehicle acceleration data for each vehicle express 3D real space.

[0110] This multi-source collaborative feature tensor contains not only the historical motion information of each vehicle, but also preserves the spatial relationships between all vehicles in the vehicle cluster. It also uniformly describes the temporal characteristics of a single vehicle and the spatial relationships between the vehicle cluster, providing a complete data foundation for subsequent feature extraction by convolutional neural networks.

[0111] Step S104: As Figure 2 As shown, a convolutional neural network is constructed, and the convolutional neural network is used to sequentially perform local feature extraction, collaborative feature fusion, and collaborative localization processing on the multi-source collaborative feature tensor to obtain the collaborative localization result for each vehicle. Step S104 of this embodiment may include the following sub-steps:

[0112] Sub-step S1041: Construct a convolutional neural network.

[0113] Furthermore, the convolutional neural network includes a local collaborative feature extraction unit, a distributed collaborative feature fusion unit, and a collaborative localization unit connected in sequence; the local collaborative feature extraction unit includes a feature extraction layer, a cross-channel convolutional fusion layer, a max pooling layer, and a flattening layer connected in sequence.

[0114] Sub-step S1042: Utilize the local collaborative feature extraction unit to extract local features from the multi-source collaborative feature tensor, obtaining the local collaborative feature vector for each vehicle. The specific process includes:

[0115] The first step involves performing multi-layer one-dimensional convolution operations on the multi-source collaborative feature tensor using a feature extraction layer to extract the temporal dynamic feature map for each vehicle. This temporal dynamic feature map includes multiple dynamic feature values ​​arranged in chronological order. Here, multi-layer one-dimensional convolution operations effectively capture the collaborative behavior patterns, local dynamic changes, and temporal trends between vehicles, thus providing a fundamental feature representation for high-precision vehicle localization. Through multi-layer one-dimensional convolution operations, the network can model motion features at different time scales. Shallow one-dimensional convolution operations can capture high-frequency local changes and suppress instantaneous noise. Deep one-dimensional convolution operations can be applied to model temporal dependencies with longer timeframes and can identify systematic drift and dynamic error trends.

[0116] Compared to traditional filtering methods, convolutional neural networks do not rely on strict motion models and noise statistics assumptions. Instead, they learn the error evolution law through data-driven methods, which can show stronger adaptability and stability when facing dynamic and complex traffic environments.

[0117] Furthermore, the expression for the dynamic eigenvalue is:

[0118] (2)

[0119] in, Indicates the first Time of the first The dynamic characteristic values ​​of each vehicle This represents the total length of the time window for the one-dimensional convolution kernel. Indicates the first The weights of a one-dimensional convolution kernel, Indicates the first Time of the first The acceleration of each vehicle, Indicates the first bias term, the... The temporal dynamic feature map of the vehicle is represented as follows: , , Indicates the first Time of the first The dynamic characteristic values ​​of each vehicle This represents the total length of the time window for the multi-source collaborative feature tensor.

[0120] The second step involves using a cross-channel convolutional fusion layer to perform cross-channel convolution operations on all temporal dynamic feature maps, resulting in a multi-layer spatial correlation feature map that characterizes the relationships between all sensors.

[0121] Furthermore, the expression for the spatial association feature map is:

[0122] (3)

[0123] in, Indicates the first Layer spatial correlation feature map Indicates the first Cross-channel convolution operation for each sensor.

[0124] Thirdly, to further compress the feature dimension and enhance robustness to noise, a max-pooling operation is introduced. Max-pooling layers are used to perform max-pooling on each spatially correlated feature map, resulting in a corresponding pooled co-feature map. This pooled co-feature map includes multiple pooled feature values. Max-pooling reduces the computational complexity of the feature tensor while preserving the most salient local features, thus reducing the interference of sensor noise and abrupt changes on the feature representation.

[0125] Furthermore, the expression for the pooling eigenvalues ​​is:

[0126] (4)

[0127] in, Indicates the first The first layer in the spatial correlation feature graph Time, Number The vehicle, the Pooling feature values ​​corresponding to each sensor This indicates taking the maximum value. Indicates a pooled window. Indicates the first The first layer in the spatial correlation feature graph Time, Number The vehicle, the Pooling feature values ​​corresponding to each sensor This represents the time offset within the pooling window. This represents the vehicle offset within the pooling window, the first... The pooling collaborative feature map corresponding to the layer spatial correlation feature map is represented as follows: .

[0128] The fourth step involves using a flattening layer to flatten the dimensions of each pooled collaborative feature map, thereby obtaining the corresponding local collaborative feature vector of the vehicle.

[0129] Furthermore, the expression for the local collaborative feature vector is:

[0130] (5)

[0131] in, Indicates the first Local collaborative feature vectors of individual vehicles This indicates the flattening process of the dimensions. express 3D real space, This represents the total length of the time window for pooling collaborative feature maps. This represents the total number of vehicles in the system model. This represents the total number of sensors in the system model.

[0132] Sub-step S1043: Utilize the distributed collaborative feature fusion unit to perform the following operations sequentially:

[0133] In Operation 1, each vehicle shares its local cooperative feature vector with all other neighboring vehicles within its communication range. Then, it performs a non-linear aggregation of all received local cooperative feature vectors with its own local cooperative feature vector to obtain the corresponding vehicle's fused cooperative feature vector. The specific process of Operation 1 is as follows:

[0134] The first step is for each vehicle to send its local collaborative feature vector to all other neighboring vehicles within its communication range via the inter-vehicle communication link.

[0135] The second step is for each vehicle to receive local cooperative feature vectors sent by all its neighboring vehicles, thereby obtaining the corresponding set of neighborhood cooperative feature vectors for each vehicle.

[0136] The third step is to perform a nonlinear transformation on each local cooperative feature vector in the neighborhood cooperative feature vector set of each vehicle to obtain the corresponding nonlinear feature vector.

[0137] Furthermore, the expression for the nonlinear eigenvector is:

[0138] (6)

[0139] in, Indicates the first In the neighborhood collaborative feature vector set of vehicle n, the th The nonlinear feature vector corresponding to the local cooperative feature vector of each neighboring vehicle. Represents a non-linear activation function. Indicates the first In the neighborhood collaborative feature vector set of vehicle n, the th Local collaborative feature vectors of neighboring vehicles.

[0140] The fourth step is to calculate the mean vector and covariance matrix of all nonlinear eigenvectors corresponding to each vehicle in turn.

[0141] Furthermore, the expression for the mean vector of all nonlinear eigenvectors corresponding to the vehicle is:

[0142] (7)

[0143] in, Indicates the first The mean vector of all nonlinear eigenvectors corresponding to each vehicle. Indicates the first The number of all neighboring vehicles of a vehicle.

[0144] Furthermore, the expression for the covariance matrix of all nonlinear eigenvectors corresponding to the vehicle is:

[0145] (8)

[0146] in, Indicates the first The covariance matrix of all nonlinear eigenvectors corresponding to each vehicle This indicates transpose.

[0147] The fifth step involves using the mean vector and covariance matrix of all nonlinear eigenvectors of each vehicle to calculate the Mahalanobis distance between the corresponding vehicle and each neighboring vehicle within its communication range.

[0148] Furthermore, the expression for the Mahalanobis distance is:

[0149] (9)

[0150] in, Indicates the first The vehicle within its own communication range Mahalanobis distance between neighboring vehicles.

[0151] The sixth step involves constructing corresponding confidence weights using the Mahalanobis distance between each vehicle and each neighboring vehicle within its communication range, and then normalizing each confidence weight to obtain the corresponding normalized confidence weights.

[0152] Step 7: Use each normalized confidence weight to perform weighted fusion of the corresponding local collaborative feature vectors in the neighborhood collaborative feature vector set of the corresponding vehicle to obtain the corresponding vehicle's fused collaborative feature vector.

[0153] Furthermore, the expression for the fused collaborative feature vector is:

[0154] (10)

[0155] in, Indicates the first The fusion and collaborative feature vector of each vehicle Indicates the first The vehicle within its own communication range Normalized confidence weights corresponding to each neighbor vehicle , Indicates the first The vehicle within its own communication range The credibility weight of each neighbor's vehicle , This represents the regularization term.

[0156] Weighted fusion can not only improve the positioning accuracy of individual vehicles, but also enhance the global consistency of the overall positioning results through information sharing within the vehicle cluster.

[0157] Operation two involves iteratively obtaining the fused collaborative feature vector. In each iteration, each vehicle updates its local collaborative feature vector based on the information injection coefficients of that iteration, until the changes in the local collaborative feature vectors of all vehicles converge or the maximum number of iterations is reached, thus obtaining the global collaborative feature vector for each vehicle. The specific process of operation two is as follows:

[0158] The first step is to set the convergence threshold for the change and the maximum number of iterations, and then set the information injection coefficient for each iteration. The convergence threshold for the change is... express.

[0159] The second step is to express the change in the vehicle's local cooperative feature vector as follows:

[0160] (11)

[0161] in, Indicates the first During the nth iteration The change in the local collaborative feature vector of each vehicle Indicates the first During the nth iteration Local collaborative feature vectors of individual vehicles Indicates the first During the nth iteration Local collaborative feature vectors of individual vehicles , Indicates the first Information injection coefficients at the next iteration Indicates the first The number of all neighboring vehicles of a given vehicle. Indicates the first During the nth iteration In the neighborhood collaborative feature vector set of vehicle n, the th Local collaborative feature vectors of neighboring vehicles.

[0162] when or When the maximum number of iterations is reached, the fusion collaborative feature vector of the corresponding vehicle is used as the corresponding global collaborative feature vector.

[0163] Sub-step S1044: The cooperative localization unit performs cooperative localization processing on each global cooperative feature vector to obtain the corresponding vehicle cooperative localization result. The specific process of sub-step S1044 is as follows:

[0164] The first step involves a cooperative localization unit comprising a location coding network, which includes fully connected layers. The global cooperative feature vector for each vehicle is input into the location coding network, and the actual location information of the corresponding vehicle is obtained by mapping using the fully connected layers.

[0165] Furthermore, the expression for the vehicle's actual location information is as follows:

[0166] (12)

[0167] in, Indicates the first Time of the first The actual location information of each vehicle. Represents the regression weight matrix. Indicates the first Global collaborative feature vectors of individual vehicles This indicates the second bias term.

[0168] Through this mapping, the high-dimensional spatiotemporal features extracted by the convolutional neural network can be effectively converted into three-dimensional spatial locations, while preserving the temporal dynamics and inter-vehicle cooperative behavior features contained in the high-dimensional spatiotemporal features. Compared to directly using the observation data of a single vehicle for localization, this method fully utilizes the cooperative information of the cluster of vehicles, improving the localization accuracy and robustness of individual vehicles.

[0169] The second step is to use the observed positioning information of each vehicle at the current moment to verify the corresponding actual positioning information, thereby obtaining the corresponding collaborative positioning information.

[0170] Furthermore, the expression for the vehicle's cooperative localization information is as follows:

[0171] (13)

[0172] in, Indicates the first Time of the first Coordinated location information of individual vehicles Indicates the correction factor. Indicates the first Time of the first The observation and positioning information of each vehicle.

[0173] Through feedback verification, the vehicle distributed cooperative positioning system can dynamically adjust the prediction results to make them closer to the actual location, while suppressing the impact of abnormal observations on the final positioning accuracy.

[0174] To verify the superiority of the CNN-based distributed heterogeneous navigation information fusion cooperative localization method proposed in this application, the following simulation experiments were conducted.

[0175] The simulation experiment consists of three parts: The first part represents the ideal scenario, where no navigation source signals are lost, and analyzes the impact of key simulation parameters on the accuracy of fusion and collaborative localization. The second part represents the abrupt change scenario, where some navigation source signals are interfered with or even interrupted, to verify the anti-interference capability of the CNN-based distributed heterogeneous navigation information fusion and collaborative localization method. The third part represents the occlusion scenario, verifying the anti-interference capability of the CNN-based distributed heterogeneous navigation information fusion and collaborative localization method under occlusion conditions.

[0176] Specifically, the simulation experiment mainly evaluates the following two aspects: positioning accuracy and model convergence speed.

[0177] like Figure 3 As shown, the simulation experiment was conducted in a multi-vehicle cooperative localization environment based on simulation data. Through a convolutional neural network, each vehicle extracts its own features using its navigation source information and exchanges features with neighboring vehicles to achieve distributed feature fusion. The fused features are then used for cooperative localization prediction, and local error correction and iterative mechanisms are combined to further improve localization accuracy and robustness. Figure 3In this diagram, the X-axis is defined as a road running from north to south with a distance of -200 meters to 200 meters, and the Y-axis is defined as a road running from west to east with a distance of -200 meters to 200 meters. Ramp 1 is a road that starts from the south, turns right, and then travels east; Ramp 2 is a road that starts from the west, turns right, and then travels south; Ramp 3 is a road that starts from the east, turns right, and then travels north; Ramp 4 is a road that starts from the north, turns right, and then travels west; and Ramp 5 is a road that starts from the east, loops around, and then travels south.

[0178] To ensure convergence and predictive performance during training, the training parameters of the convolutional neural network model were carefully set. In this simulation, an adaptive estimation optimizer was employed, which improved training efficiency and convergence stability by adaptively estimating and adjusting the learning rates of the first and second moments of the gradient. An initial learning rate of 0.001 was set to allow for larger weight updates in the early stages of training, quickly approaching the optimal solution. The maximum number of training epochs was set to 500, and the batch size to 128, thus achieving a balance between training speed and model performance. Furthermore, to avoid getting trapped in local optima and improve the generalization ability of the convolutional neural network model, the training samples were randomly shuffled at the beginning of each training cycle. A dynamic visualization module was enabled during training to monitor changes in training and validation errors in real time, enabling timely detection of overfitting or underfitting issues and real-time adjustment of the model training strategy based on feedback.

[0179] Through the above design, the experiment can realistically reflect the localization performance of the distributed cooperative network based on convolutional neural network in multiple vehicle simulation environments, and at the same time verify the robustness and real-time performance of the method under sudden changes and multi-source heterogeneous data conditions.

[0180] To quantitatively evaluate the vehicle localization accuracy of the convolutional neural network model, we further provide a heatmap of the vehicle localization error, such as... Figure 4 As shown, the spatial distribution characteristics and density of the deviation between the predicted and actual positions are intuitively displayed. The results show that the high-error region is concentrated in a few discrete points, while most areas exhibit a cool color tone with low error, indicating that the overall deviation between the predicted and actual trajectories is small and concentrated. This observation result corroborates the aforementioned trajectory overlap phenomenon, jointly verifying that the method proposed in this application can still achieve high-precision and high-stability trajectory prediction in complex dynamic environments.

[0181] like Figure 5 As shown, under normal circumstances, there are significant differences in positioning accuracy when using four different methods. This application achieves the best performance using a convolutional neural network, with an average positioning error of approximately [missing value]. This indicates that it can effectively extract multi-source spatiotemporal features and improve positioning accuracy. The average positioning error when using a backpropagation neural network is approximately... m, second only to convolutional neural networks but better than other methods, indicates that backpropagation neural networks have relatively stable error convergence capabilities, but their expressive power is limited when facing complex nonlinear features. The average localization error using factor maps is approximately While modeling with factor graphs can maintain good consistency, its adaptability in dynamic scenarios is insufficient due to limitations in structural design and real-time performance. The average positioning error using extended Kalman filtering is approximately... m performs the weakest, mainly due to the accumulation of linearization errors under strong nonlinear conditions and its sensitivity to noise.

[0182] like Figure 6 As shown, under abrupt changes, the localization accuracy of all methods is affected. Extended Kalman filtering, factor graphs, and backpropagation neural networks all exhibit varying degrees of fluctuation and recovery delay. Among them, the average localization error using factor graphs is approximately... This is due to the limitations of factor graph structure updates and nonlinear modeling, resulting in insufficient adaptability to mutations. In contrast, the performance using convolutional neural networks is optimal, with virtually no significant degradation during the mutation phase and an average localization error of approximately [missing value]. m significantly outperforms other methods, primarily due to its convolutional structure's ability to extract deep spatial features from multi-source observations, effectively suppressing mutation noise and achieving robust fusion. Further analysis reveals that extended Kalman filtering is prone to cumulative bias due to the linearization assumption, while backpropagation neural networks lack temporal modeling capabilities, resulting in insufficient adaptability to continuous mutations. Although factor graphs offer probabilistic consistency, their real-time update capability is limited.

[0183] like Figure 7 As shown, the localization errors of each method are compared under simulated occlusion conditions. The results show that when using the extended Kalman filter, the error increases significantly and recovery is slow during occlusion; while the factor map method can maintain state correlation, its optimization efficiency decreases under discontinuous observation conditions, with the average localization error approximately [missing value]. m; When using a backpropagation neural network, it is quite sensitive to temporal discontinuities, exhibiting significant fluctuations during the occlusion phase. In contrast, the performance using a convolutional neural network is optimal, with an average localization error of approximately m, and with minimal fluctuation, is mainly due to the ability to effectively compensate for missing observations through convolutional feature fusion and to improve overall robustness by utilizing collaborative information from multiple vehicles.

[0184] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0185] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0186] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application.

[0187] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

Claims

1. A distributed heterogeneous navigation information fusion and cooperative localization method based on CNN, characterized in that, The method includes the following steps: A system model for constructing a vehicle distributed cooperative positioning system is provided. The system model includes a distributed cooperative network composed of multiple vehicles, each of which is equipped with multiple sensors and is regarded as an independent network node. The original multi-source heterogeneous navigation data of each vehicle is acquired, and each original multi-source heterogeneous navigation data is preprocessed to obtain the corresponding standard multi-source heterogeneous navigation data. All the standard multi-source heterogeneous navigation data are arranged in chronological order and according to the type of the sensor to obtain a multi-source cooperative feature tensor; A convolutional neural network is constructed, and the convolutional neural network is used to sequentially perform local feature extraction, collaborative feature fusion, and collaborative localization processing on the multi-source collaborative feature tensor to obtain the collaborative localization result for each vehicle, including: The convolutional neural network includes a local collaborative feature extraction unit, a distributed collaborative feature fusion unit, and a collaborative localization unit connected in sequence; the local collaborative feature extraction unit includes a feature extraction layer, a cross-channel convolutional fusion layer, a max pooling layer, and a flattening layer connected in sequence. The local collaborative feature extraction unit is used to extract local features from the multi-source collaborative feature tensor to obtain the local collaborative feature vector for each vehicle. The distributed collaborative feature fusion unit performs the following operations sequentially: Each vehicle shares its local collaborative feature vector with all other neighboring vehicles within its communication range, and performs nonlinear aggregation of all received local collaborative feature vectors with its own local collaborative feature vector to obtain the corresponding vehicle's fused collaborative feature vector; The step of obtaining the fused collaborative feature vector is performed iteratively. In each iteration, each vehicle updates its local collaborative feature vector according to the information injection coefficient of this iteration until the change of the local collaborative feature vector of all vehicles converges or reaches the maximum number of iterations, and the global collaborative feature vector of each vehicle is obtained respectively. The cooperative localization unit is used to perform cooperative localization processing on each of the global cooperative feature vectors to obtain the corresponding cooperative localization result of the vehicle.

2. The CNN-based distributed heterogeneous navigation information fusion and cooperative localization method according to claim 1, characterized in that, The types of sensors include global satellite navigation systems and inertial navigation systems.

3. The CNN-based distributed heterogeneous navigation information fusion and cooperative localization method according to claim 2, characterized in that, The preprocessing includes performing time alignment, data repair, and normalization on each of the original multi-source heterogeneous navigation data in sequence to obtain the corresponding standard multi-source heterogeneous navigation data. Each of the original multi-source heterogeneous navigation data includes the original vehicle position data measured by the global satellite navigation system and the original vehicle state data measured by the inertial navigation system; Each of the aforementioned standard multi-source heterogeneous navigation data includes standard vehicle location data and standard vehicle status data.

4. The CNN-based distributed heterogeneous navigation information fusion and cooperative localization method according to claim 1, characterized in that, The expression for the multi-source collaborative feature tensor is: (1) in, Represents the multi-source collaborative feature tensor. express 3D real space, This represents the total length of the time window for the multi-source collaborative feature tensor. This represents the total number of vehicles in the system model. This represents the number of all sensors in the system model. Indicates the first Time of the first Standard multi-source heterogeneous navigation data for each vehicle , Indicates the first Time of the first Standard vehicle location data for each vehicle Indicates the first Time of the first Standard vehicle speed data for each vehicle Indicates the first Time of the first Standard vehicle acceleration data for each vehicle express 3D real space.

5. The CNN-based distributed heterogeneous navigation information fusion and cooperative localization method according to claim 1, characterized in that, The step of using the local collaborative feature extraction unit to extract local features from the multi-source collaborative feature tensor to obtain the local collaborative feature vector of each vehicle includes: The feature extraction layer is used to perform multi-layer one-dimensional convolution operation on the multi-source collaborative feature tensor to extract the temporal dynamic feature map of each vehicle. Each temporal dynamic feature map includes multiple dynamic feature values ​​arranged in chronological order. The expression for the dynamic feature value is: (2) in, Indicates the first Time of the first The dynamic characteristic values ​​of each vehicle This represents the total length of the time window for the one-dimensional convolution kernel. Indicates the first The weights of a one-dimensional convolution kernel, Indicates the first Time of the first The acceleration of each vehicle, Indicates the first bias term, the... The temporal dynamic feature map of the vehicle is represented as follows: , , Indicates the first Time of the first The dynamic characteristic values ​​of each vehicle This represents the total length of the time window for the multi-source collaborative feature tensor. The cross-channel convolutional fusion layer is used to perform cross-channel convolution operations on all the temporal dynamic feature maps to obtain a multi-layer spatial correlation feature map that characterizes the relationship between all the sensors; The expression for the spatial association feature map is: (3) in, Indicates the first Layer spatial correlation feature map Indicates the first Cross-channel convolution operation for each sensor; Max pooling is applied to each of the spatial correlation feature maps in the max pooling layer to obtain corresponding pooling collaborative feature maps. Each pooling collaborative feature map includes multiple pooling feature values. The expression for the pooling feature value is: (4) in, Indicates the first The first layer in the spatial correlation feature graph Time, Number The vehicle, the Pooling feature values ​​corresponding to each sensor This indicates taking the maximum value. Indicates a pooled window. Indicates the first The first layer in the spatial correlation feature graph Time, Number The vehicle, the Pooling feature values ​​corresponding to each sensor This represents the time offset within the pooling window. This represents the vehicle offset within the pooling window, the first... The pooling collaborative feature map corresponding to the spatial correlation feature map of the layer is represented as follows: ; The flattening layer is used to flatten the dimensions of each pooled collaborative feature map to obtain the corresponding local collaborative feature vector of the vehicle. The expression for the local collaborative feature vector is: (5) in, Indicates the first Local collaborative feature vectors of individual vehicles This indicates the flattening process of the dimensions. express 3D real space, This represents the total length of the time window for pooling collaborative feature maps. This represents the total number of vehicles in the system model. This represents the total number of sensors in the system model.

6. The CNN-based distributed heterogeneous navigation information fusion and cooperative localization method according to claim 1, characterized in that, The steps of each vehicle sharing its local cooperative feature vector with all other neighboring vehicles within its communication range, and nonlinearly aggregating all received local cooperative feature vectors with its own local cooperative feature vector to obtain the corresponding vehicle's fused cooperative feature vector include: Each vehicle will send its local collaborative feature vector to all other neighboring vehicles within its communication range via the inter-vehicle communication link; Each vehicle receives the local cooperative feature vectors sent by all its neighboring vehicles, thus obtaining a set of neighborhood cooperative feature vectors for the corresponding vehicle. Each of the local cooperative feature vectors in the neighborhood cooperative feature vector set of each vehicle is subjected to a nonlinear transformation to obtain the corresponding nonlinear feature vector; The expression for the nonlinear eigenvector is: (6) in, Indicates the first In the neighborhood collaborative feature vector set of vehicle n, the th The nonlinear feature vector corresponding to the local cooperative feature vector of each neighboring vehicle. Represents a non-linear activation function. Indicates the first In the neighborhood collaborative feature vector set of vehicle n, the th Local collaborative feature vectors of neighboring vehicles; Calculate the mean vector and covariance matrix of all nonlinear eigenvectors corresponding to each vehicle in turn; The expression for the mean vector of all the nonlinear feature vectors corresponding to the vehicle is: (7) in, Indicates the first The mean vector of all nonlinear eigenvectors corresponding to each vehicle. Indicates the first The number of all neighboring vehicles of a vehicle; The expression for the covariance matrix of all the nonlinear eigenvectors corresponding to the vehicle is: (8) in, Indicates the first The covariance matrix of all nonlinear eigenvectors corresponding to each vehicle Indicates transpose; The Mahalanobis distance between the corresponding vehicle and each of its neighboring vehicles within its communication range is calculated using the mean vector and covariance matrix of all the nonlinear feature vectors of each vehicle. The expression for the Mahalanobis distance is: (9) in, Indicates the first The vehicle within its own communication range Mahalanobis distance between neighboring vehicles; Each vehicle constructs a corresponding confidence weight using the Mahalanobis distance between itself and each of its neighboring vehicles within its communication range, and normalizes each confidence weight to obtain the corresponding normalized confidence weight. Each of the normalized confidence weights is used to perform weighted fusion of the local collaborative feature vectors in the neighborhood collaborative feature vector set of the corresponding vehicle to obtain the fused collaborative feature vector of the corresponding vehicle. The expression for the fusion collaborative feature vector is: (10) in, Indicates the first The fusion and collaborative feature vector of each vehicle Indicates the first The vehicle within its own communication range Normalized confidence weights corresponding to each neighbor vehicle , Indicates the first The vehicle within its own communication range The credibility weight of each neighbor's vehicle , This represents the regularization term.

7. The CNN-based distributed heterogeneous navigation information fusion and cooperative localization method according to claim 1, characterized in that, The step of iteratively obtaining the fused collaborative feature vector includes the following steps: In each iteration, each vehicle updates its local collaborative feature vector according to the information injection coefficient of this iteration, until the changes in the local collaborative feature vectors of all vehicles converge or the maximum number of iterations is reached, and the step of obtaining the global collaborative feature vector of each vehicle includes: Set the change convergence threshold and the maximum number of iterations, and set the information injection coefficient for each iteration. The change convergence threshold is used... express; The expression for the change in the local cooperative feature vector of the vehicle is: (11) in, Indicates the first During the nth iteration The change in the local collaborative feature vector of each vehicle Indicates the first During the nth iteration Local collaborative feature vectors of individual vehicles Indicates the first During the nth iteration Local collaborative feature vectors of individual vehicles , Indicates the first Information injection coefficients at the next iteration Indicates the first The number of all neighboring vehicles of a given vehicle. Indicates the first During the nth iteration In the neighborhood collaborative feature vector set of vehicle n, the th Local collaborative feature vectors of neighboring vehicles; when or When the maximum number of iterations is reached, the fusion collaborative feature vector of the corresponding vehicle is used as the corresponding global collaborative feature vector.

8. The CNN-based distributed heterogeneous navigation information fusion and cooperative localization method according to claim 1, characterized in that, The step of using the cooperative localization unit to perform cooperative localization processing on each of the global cooperative feature vectors to obtain the corresponding cooperative localization result for the vehicle includes: The cooperative localization unit includes a location coding network, and the location coding network includes a fully connected layer. The global collaborative feature vector of each vehicle is input into the location coding network, and the actual location information of the corresponding vehicle is obtained by mapping using the fully connected layer. The expression for the actual location information of the vehicle is: (12) in, Indicates the first Time of the first The actual location information of each vehicle. Represents the regression weight matrix. Indicates the first Global collaborative feature vectors of individual vehicles Indicates the second bias term; The observed positioning information of each vehicle at the current moment is used to verify the corresponding actual positioning information to obtain the corresponding cooperative positioning information; The expression for the cooperative positioning information of the vehicle is: (13) in, Indicates the first Time of the first Coordinated location information of individual vehicles Indicates the correction factor. Indicates the first Time of the first The observation and positioning information of each vehicle.

Images