Online vector map construction method based on hybrid expert model and relation encoding

Abstract

This invention provides an online vector map construction method based on hybrid expert models and relational encoding. By constructing multiple expert models and iteratively training them on a feature set containing multimodal features, it achieves refined modeling of multimodal interactions, fully releasing the information value of different modal feature data. Furthermore, a relational query mechanism including instance-level and point-level queries is constructed to simultaneously capture spatial relationships between map elements, connection patterns of points within the same map element, and the attribution relationships between map elements and points, avoiding element distortion or missing elements. Simultaneously, location encoding technology is used to enhance the utilization of map element reference points, effectively improving the accuracy of location representation. This method provides a solid theoretical foundation and technical support for real-time updates of online high-precision vector maps for autonomous driving, eliminating the need for expensive additional sensors and balancing lightweight deployment with engineering practicality.

Description

Technical Field

[0001] This invention relates to the field of vector map construction technology in automatic navigation, and more specifically, to an online vector map construction method based on a hybrid expert model and relational coding. Background Technology

[0002] In recent years, autonomous driving technology has become increasingly reliant on high-precision maps, requiring them for navigation. As a core infrastructure of autonomous driving systems, high-precision maps need to provide accurate semantic and location information of road elements to support key functions such as perception, localization, and path planning. However, traditional high-precision map construction methods heavily depend on surveying vehicles equipped with specialized equipment, which is not only costly but also inefficient in updating maps, making it difficult to meet the core requirements of autonomous driving for maps that are "real-time, have high coverage, and are highly accurate."

[0003] With the development of computer vision and deep learning technologies, online vector map construction technology based on multimodal sensors has gradually become a research hotspot. Most of the current mainstream methods are based on the DETR architecture, but in practical applications, existing technologies have significant drawbacks such as coarse multimodal interaction modeling and insufficient representation of map element relationships. Summary of the Invention

[0004] The purpose of this invention is to provide an online vector map construction method based on a hybrid expert model and relational coding, so as to improve the accuracy, robustness and real-time performance of online mapping.

[0005] In a first aspect, the present invention provides an online vector map construction method based on a hybrid expert model and relation coding, the method comprising: RGB images and LiDAR point clouds of the road environment are acquired. The corresponding encoders are used to extract the feature information of the RGB images and LiDAR point clouds respectively, and they are mapped to the bird's-eye view space to obtain BEV features. Based on the feature information and BEV features, a feature set containing multimodal features is constructed. Multiple expert models are constructed, and preliminary interaction features are generated by performing corresponding task processing based on each expert model and according to the feature set. Each expert model is iteratively trained according to the constructed loss function to obtain an expert model that meets the preset requirements. Based on the BEV features and preliminary interaction features, a relational query including instance-level query and point-level query is constructed. The location encoding technology is used to convert map element reference points into learnable location features. The converted location features are then fused with the relational query to obtain a relational query with fused location information. The relationship query between the feature information and the fused location information is concatenated to obtain the joint input feature. The weight coefficients of each expert model are determined based on the joint input feature. The vector map prediction result is obtained by combining the weight coefficients of each expert model and the prediction result.

[0006] In an optional implementation, the loss function is obtained by summing the task loss terms corresponding to each expert model according to their respective weights, and then adding them to the penalty term. The method further includes: The vector map prediction results are aligned and corrected with the prior map data to generate corrected vector map prediction results. Based on the corrected vector map prediction results and the prior map data, the evaluation results of the corrected vector map prediction results under multiple preset evaluation indicators are obtained. Based on the analysis of the vector map prediction results, the expert performance information of each expert model under the corresponding weight coefficients is obtained; The weights of each task loss term in the loss function are adjusted based on the expert performance information, and the penalty coefficients of the penalty terms in the loss function are adjusted based on the evaluation results of the multiple evaluation indicators. The expert models are then optimized and trained using the adjusted loss function.

[0007] In an optional implementation, the RGB image is used to record visual information of the road environment, and the lidar point cloud is used to record spatial structure information of the road environment. The step of extracting feature information from the RGB image and the LiDAR point cloud using corresponding encoders and mapping it to the bird's-eye view space to obtain BEV features includes: An encoder composed of a convolutional neural network is used to extract features from the RGB image to obtain visual semantic features; An encoder composed of a deep learning network is used to extract features from the lidar point cloud to obtain spatial structure features; The visual semantic features and spatial structural features are converted to the same viewpoint through a viewpoint conversion algorithm, and then mapped to the bird's-eye view space through feature projection operation to generate BEV features.

[0008] In an optional implementation, the plurality of expert models include expert models corresponding to visual semantic features and spatial structural features respectively, expert models that capture shared redundant information among multimodal features, expert models that capture emerging collaborative information after multimodal combination, and expert models that integrate BEV features with other modal feature interaction information. Each of the expert models includes a fusion model and a prediction head. The fusion model is used to perform feature fusion on the input feature set and output fusion features adapted to the corresponding task. The prediction head is used to output preliminary interaction features corresponding to the interaction type based on the fusion features. The preliminary interaction features include map element categories and point coordinate information. When iteratively training an expert model for corresponding visual semantic features or spatial structure features, the modal features other than the target modal features in the feature set are replaced with random vectors. The prediction results under the input of the target modal features are taken as positive examples and the prediction results under the input of other modal features are taken as negative examples. Iterative training is carried out by constraining the direction in which the similarity between the expert model output and the positive example is higher than the similarity with the negative example. When iteratively training an expert model that captures shared redundant information among multiple modalities, visual semantic features and spatial structural features are replaced with random vectors respectively. The prediction results under all single-modal feature inputs are taken as positive examples, and the deviation between the expert model output and all positive examples is constrained within a preset range by the mean squared error loss function for iterative training. When iteratively training an expert model that captures emerging collaborative information after multimodal combination, each modal feature is replaced with a random vector. The prediction results under all single-modal feature inputs are taken as negative examples. The cosine similarity loss function is used to constrain the direction in which the similarity between the expert model output and all negative examples is lower than a preset threshold for iterative training. When iteratively training an expert model that integrates BEV features with interaction information from other modal features, the BEV features and other modal features are masked separately. The prediction results with the masked features as common input are considered positive examples, and the prediction results with single modal feature input are considered negative examples. The expert model is iteratively trained using the cross-entropy loss function.

[0009] In an optional implementation, the step of constructing a relational query including instance-level and point-level queries based on the BEV features and preliminary interaction features includes: The preliminary interaction features are incorporated into the BEV features through feature concatenation operations to form the BEV basic features that integrate interaction information. Based on the BEV basic features, the association weights of different map elements in the bird's-eye view space are calculated through a self-attention mechanism to construct an instance-level query. Based on the BEV basic features, the association weights between points within the same map element are calculated using a self-attention mechanism to obtain the topological relationships of points within the map element. The association weights between points and their respective map elements are calculated using a cross-attention mechanism to obtain the subordinate relationships between instances and points, thereby constructing a point-level query.

[0010] In an optional implementation, the step of using location encoding technology to convert map element reference points into learnable location features, and fusing the converted location features with the relationship query to obtain a relationship query with fused location information, includes: Obtain the coordinate values ​​of map element reference points in the bird's-eye view space, and use a preset location encoding formula to map different dimensions of the coordinate values, converting them into learnable location features with location distinguishability; Adjust the dimensions of the transformed location features, instance-level queries, and point-level queries in the relational query to be consistent; By adding elements together, the location features are incorporated into the instance-level query and the point-level query respectively, resulting in instance-level query and point-level query with fused location information. The instance-level and point-level queries of the merged location information are integrated to form the final relational query of the merged location information.

[0011] In an optional implementation, the step of determining the weight coefficients of each expert model based on the joint input features, and combining the weight coefficients of each expert model with the prediction results to obtain the vector map prediction results includes: The joint input features are compressed to map into low-dimensional feature vectors. The initial weight coefficients of each expert model are calculated based on the low-dimensional feature vectors. The initial weight coefficients are positively correlated with the degree of adaptation of the expert model to the corresponding task. The initial weight coefficients of each expert model are normalized to obtain normalized weight coefficients. The vector map prediction result is obtained by weighting and summing the prediction results of multiple expert models and normalized weight coefficients.

[0012] In an optional implementation, the vector map prediction result includes the predicted map element category probability, predicted point coordinates and corresponding confidence scores, and the map prior data includes the baseline map element category and baseline point coordinates; The step of aligning and correcting the vector map prediction result with the prior map data to generate a corrected vector map prediction result includes: The coordinate transformation matrix between the vector map prediction result and the map prior data is calculated using the iterative nearest point algorithm, and the vector map prediction result is mapped to the coordinate system of the map prior data based on the coordinate transformation matrix; Calculate the coordinate deviation between the predicted point coordinates in the vector map prediction result and the reference point coordinates in the map prior data, and align and correct the predicted point coordinates in the vector map prediction result based on the coordinate deviation to generate a corrected vector map prediction result.

[0013] In an optional implementation, the step of obtaining the evaluation results of the corrected vector map prediction results under multiple preset evaluation indicators based on the corrected vector map prediction results and the map prior data includes: Based on the multi-frame corrected vector map prediction results, a multi-frame prediction result set is constructed. For map elements of the same type in the prediction result set, a non-maximum suppression strategy is used for fusion processing to remove redundant and overlapping map elements in the multi-frame prediction result set, and a first draft of the fused regional vector map is generated. Based on the initial draft of the regional vector map and the prior map data, the evaluation results of multiple preset evaluation indicators are calculated. The multiple evaluation metrics include element category accuracy, average point coordinate deviation, and element recall. The element category accuracy is obtained by statistically analyzing the ratio of the number of correctly predicted map elements to the total number of map elements. The average point coordinate deviation is obtained by calculating the average Euclidean distance between the predicted point coordinates and the reference point coordinates. The element recall is obtained by calculating the ratio of the number of correctly predicted reference map elements to the total number of reference map elements.

[0014] In an optional implementation, the expert performance information is obtained based on the mean of the weight coefficients, the variance of the weight coefficients, and the prediction contribution of each expert model in multi-frame sample prediction. The mean of the weight coefficients represents the overall participation of the expert model in this round of tasks, the variance of the weight coefficients represents the performance stability of the expert model in different scenarios, and the prediction contribution is calculated by the similarity between the prediction results of the expert model and the final vector map prediction results.

[0015] This invention provides an online vector map construction method based on hybrid expert models and relational encoding. By constructing multiple expert models and iteratively training them on a feature set containing multimodal features, it achieves refined modeling of multimodal interactions, fully releasing the information value of different modal feature data. Furthermore, a relational query mechanism including instance-level and point-level queries is constructed to simultaneously capture spatial relationships between map elements, connection patterns of points within the same map element, and the attribution relationships between map elements and points, avoiding element distortion or missing elements. Simultaneously, location encoding technology is used to enhance the utilization of map element reference points, effectively improving the accuracy of location representation. This method provides a solid theoretical foundation and technical support for real-time updates of online high-precision vector maps for autonomous driving, eliminating the need for expensive additional sensors and balancing lightweight deployment with engineering practicality. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments of the present invention will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart of an online vector map construction method provided in an embodiment of the present invention; Figure 2 for Figure 1 A flowchart of the sub-steps included in S11; Figure 3 for Figure 1 A flowchart of the sub-steps included in S13; Figure 4 for Figure 1 A flowchart of the sub-steps included in S14; Figure 5 for Figure 1 A flowchart of the sub-steps included in S15; Figure 6 A flowchart of the model optimization method in the online vector map construction method provided in this embodiment of the invention; Figure 7 A functional module block diagram of the online vector map construction system provided in this embodiment of the invention; Figure 8 This is a structural block diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will now be described with reference to the accompanying drawings in the embodiments of the present invention.

[0019] Please see Figure 1 The flowchart below shows the online vector map construction method based on hybrid expert model and relation coding provided in the embodiments of the present invention. The detailed steps of the online vector map construction method are described below.

[0020] S11. Acquire RGB images and LiDAR point clouds of the road environment, extract feature information from the RGB images and LiDAR point clouds using corresponding encoders, and map them to the bird's-eye view space to obtain BEV features. Construct a feature set containing multimodal features based on the feature information and BEV features. S12, Construct multiple expert models, generate preliminary interactive features based on each expert model and the corresponding task processing according to the feature set, and iteratively train each expert model according to the constructed loss function to obtain an expert model that meets the preset requirements. S13, construct a relational query that includes instance-level query and point-level query based on BEV features and preliminary interaction features; S14, The location coding technology is used to convert the map element reference points into learnable location features, and the converted location features are fused with the relationship query to obtain the relationship query with fused location information; S15: The relationship query between feature information and fused location information is concatenated to obtain joint input features. The weight coefficients of each expert model are determined based on the joint input features. The vector map prediction result is obtained by combining the weight coefficients of each expert model and the prediction results.

[0021] The online vector map construction method based on hybrid expert models and relational encoding provided in this embodiment constructs multiple expert models and iteratively trains them based on a feature set containing multimodal features. This achieves refined modeling of multimodal interactions and fully releases the information value of different modal feature data. Furthermore, a relational query mechanism including instance-level and point-level queries is constructed to simultaneously capture spatial relationships between map elements, connection patterns of points within the same map element, and the attribution relationships between map elements and points, avoiding element distortion or missing elements. Simultaneously, location encoding technology is used to enhance the utilization of map element reference points, effectively improving the accuracy of location representation. This method provides a solid theoretical foundation and technical support for real-time updates of online high-precision vector maps for autonomous driving, eliminating the need for expensive additional sensors and balancing lightweight deployment with engineering practicality.

[0022] The specific implementation methods of each of the above steps will be explained in detail below.

[0023] In this embodiment, the vehicle is equipped with sensing devices such as an onboard camera and a LiDAR sensor to form a multimodal sensor system. The onboard camera can acquire RGB images of the road environment, while the LiDAR can collect LiDAR point clouds of the road environment. The high-precision RGB images are used to record visual information about the road environment, including road markings and the appearance of traffic facilities. The LiDAR point clouds are used to record spatial structural information about the road environment, including road width, guardrail height, and the three-dimensional positions of obstacles.

[0024] In this embodiment, corresponding encoders are used to extract feature information from the RGB image and the LiDAR point cloud, respectively, and then mapped to the bird's-eye view space to obtain BEV features. Please refer to [link to relevant documentation]. Figure 2 This step can be achieved in the following way: S111 uses an encoder composed of a convolutional neural network to extract features from RGB images and obtain visual semantic features; S112 uses an encoder composed of a deep learning network to extract features from the LiDAR point cloud to obtain spatial structure features; S113 transforms visual semantic features and spatial structural features to the same viewpoint through a viewpoint transformation algorithm, and then maps them to the bird's-eye view space through feature projection operation to generate BEV features.

[0025] In this embodiment, dedicated encoders are provided for the two modalities (including RGB images and LiDAR point clouds). For high-precision RGB images, a feature extraction is performed using an encoder composed of convolutional neural networks, such as a high-precision ResNet-50 network. Local detail features of the image are extracted through multi-layer convolution operations, and key semantic information is retained through max pooling operations. Finally, visual semantic features with uniform dimensions are output.

[0026] Feature extraction is performed on the lidar point cloud using an encoder constructed from a deep learning network, such as a high-precision PointNet network. First, the point cloud coordinates are transformed to adapt to a unified coordinate system. Then, the neighborhood information of the point cloud is integrated through local feature aggregation operations. Finally, the feature dimension is compressed through global max pooling operations, and the output is a spatial structure feature that reflects the spatial structure.

[0027] Visual semantic features and spatial structural features are transformed to the same viewpoint through a perspective transformation algorithm, and then mapped to the bird's-eye view space through feature projection operation to generate high-precision BEV features containing multimodal information.

[0028] Simultaneously, a linear layer is used to adjust the dimensions of visual semantic features, spatial structure features, and high-precision BEV features, mapping the three types of features to the same dimension to construct a feature set. ,in Representing visual semantic features, Represents spatial structural characteristics, This represents high-precision BEV characteristics, ensuring compatibility with subsequent feature processing.

[0029] Based on this, a hybrid expert framework is introduced to construct an interactive expert pool with multiple expert models, using high-precision BEV features and feature set L.

[0030] Multiple expert models include expert models specifically for visual semantic features and spatial structural features (expert models for single-modal features), which are used to focus on capturing information unique to a single modality. In addition, there are expert models that capture shared redundant information among multiple modalities, expert models that capture emerging collaborative information after multimodal combination, and expert models that integrate BEV features with interaction information from other modal features. All expert models form a set. F1, F2, F3, F4, and F5 represent expert models corresponding to visual semantic features, expert models corresponding to spatial structural features, expert models that capture redundant information between multimodal features, expert models that capture emerging collaborative information after multimodal combination, and expert models that integrate BEV features with other modal feature interaction information, respectively.

[0031] Each expert model includes a fusion model and a prediction head. The fusion model is used to fuse features from the input feature set and output fused features adapted to the corresponding task. Specifically, the fusion model uses a three-layer fully connected network to fuse modal features from the dimensionally aligned input feature set and outputs fused features adapted to the corresponding task.

[0032] The prediction head is used to output preliminary interaction features corresponding to the interaction type based on the fused features. The preliminary interaction features include map element categories and point coordinate information.

[0033] Then, a weakly supervised interactive loss function is designed to enable specialized training of the expert model.

[0034] When iteratively training an expert model for corresponding visual semantic features or spatial structure features, all modal features in the feature set except for the target modal features (visual semantic features or spatial structure features) are replaced with random vectors. The prediction results under the target modal feature input are considered positive examples, and the prediction results under other modal feature inputs are considered negative examples. Iterative training is performed by constraining the direction in which the expert model output has a higher similarity to positive examples than to negative examples. This constraint can be achieved using a triplet loss function.

[0035] When iteratively training an expert model that captures shared redundant information among multiple modalities, visual semantic features and spatial structural features are replaced with random vectors respectively. The prediction results under all single-modal feature inputs are taken as positive examples, and the deviation between the expert model output and all positive examples is constrained within a preset range by the mean squared error loss function for iterative training.

[0036] When iteratively training an expert model that captures emerging collaborative information after multimodal combination, each modal feature is replaced with a random vector. The prediction results under all single-modal feature inputs are taken as negative examples. The cosine similarity loss function is used to constrain the direction in which the similarity between the expert model output and all negative examples is lower than a preset threshold for iterative training.

[0037] When iteratively training an expert model that integrates BEV features with interaction information from other modal features, the BEV features and other modal features are masked separately. The prediction results with the masked features as common input are considered positive examples, and the prediction results with single modal feature input are considered negative examples. The expert model is iteratively trained using the cross-entropy loss function.

[0038] In this embodiment, a high-precision AdamW optimizer is used to construct the overall loss function by combining the prediction task loss terms (each loss function) of each expert model with the weakly supervised interaction loss term. Using the modal features in the feature set L as input, all parameters of the interaction expert pool are iteratively trained. After each iteration, the overall loss value is calculated. If the decrease in the loss value is less than a preset threshold, training continues until the loss value converges to a stable range, which is considered to meet the preset requirements. After training, the parameter sets of each expert model are output for use in the subsequent multimodal interaction information extraction process during vector map construction.

[0039] Based on this, a relationship query is constructed using the aforementioned high-precision BEV features and preliminary interaction features. For details, please refer to [link to relevant documentation]. Figure 3 This step can be achieved in the following ways: S131, the preliminary interactive features are integrated into the BEV features through feature splicing operations to form the BEV basic features that integrate interactive information; S132, Based on the BEV basic features, the association weights of different map elements in the bird's-eye view space are calculated through a self-attention mechanism to construct an instance-level query; S133: Based on the BEV basic features, the association weight between points within the same map element is calculated through a self-attention mechanism to obtain the topological relationship of points within the map element. The association weight between a point and its associated map element is calculated through a cross-attention mechanism to obtain the membership relationship between instances and points, so as to construct a point-level query.

[0040] In this embodiment, the high-precision BEV features generated above are used as the basic carrier. The preliminary interaction features are integrated into the high-precision BEV features through feature splicing operations to form high-precision BEV basic features with integrated interaction information, providing data support for the construction of subsequent relationship query mechanisms.

[0041] Based on the fundamental features of high-precision BEV (bird's-eye view) images with integrated interactive information, a relationship query mechanism is constructed. Relationship queries are divided into two categories: instance-level queries and point-level queries. Instance-level queries target the map elements as a whole, calculating the association weights of different map elements in the high-precision BEV (bird's-eye view) image space through a self-attention mechanism to capture the spatial relationships between map elements, thereby constructing instance-level queries.

[0042] Point-level queries target individual points that make up map elements. On the one hand, they use a self-attention mechanism to calculate the association weights between points within the same map element, capturing the topological relationships between points within the element. On the other hand, they use a cross-attention mechanism to calculate the association weights between a point and its associated map element, capturing the hierarchical relationships between instances and points, thus constructing a point-level query.

[0043] Instance-level queries and point-level queries together constitute a complete set of relational queries. ,in Represents an instance-level query. This represents a point-level query.

[0044] Furthermore, location encoding technology is used to convert map element reference points into learnable location features, and these converted location features are then fused with a relationship query to obtain a relationship query with fused location information. (See [link to relevant documentation]). Figure 4 This step can be achieved in the following way: S141, obtain the coordinate values ​​of the map element reference points in the bird's-eye view space, and use the preset position encoding formula to map the different dimensions of the coordinate values ​​and convert them into learnable position features with position discrimination. S142, adjust the dimensions of the transformed location features, instance-level queries and point-level queries in the relational query to be consistent; S143, by adding elements together, the location features are incorporated into the instance-level query and the point-level query respectively, resulting in instance-level query and point-level query with fused location information; S144 integrates instance-level and point-level queries that fuse location information to form the final relational query that fuses location information.

[0045] In this embodiment, a sine-cosine position coding technique is introduced to process the high-precision BEV coordinates of map element reference points. The coordinate values ​​of the map element reference points in the high-precision BEV (bird's-eye view) space are obtained, and these coordinate values ​​are input into a preset sine-cosine position coding formula. Different dimensions of the coordinates are mapped using sine and cosine functions to generate learnable position features with location discrimination. These position features contain the lateral and longitudinal position information of the reference point in the high-precision BEV space, accurately representing the spatial position attributes of the reference point.

[0046] Simultaneously, the generated learnable location features are fused with the relational query. A multilayer perceptron is used to process the location features and relational query. , Each feature dimension is adjusted individually to ensure consistency across all three dimensions. Then, element-wise addition is used to integrate the positional features. and In the process, an instance-level query that obtains fused location information is obtained. Point-level query Finally, and Integrate and form the final relationship query of merged location information. = This provides accurate information on both relationships and locations for subsequent vector map predictions.

[0047] Then, the relational query of feature information (visual semantic features and spatial structure features) and fused location information is concatenated to obtain joint input features. Based on the joint input features, the weight coefficients of each expert model are determined. Finally, the vector map prediction result is obtained by combining the weight coefficients of each expert model with the prediction results. For details, please refer to [link to relevant documentation]. Figure 5 This step can be achieved in the following way: S151, the joint input features are compressed in dimension to map them into low-dimensional feature vectors. The initial weight coefficients of each expert model are calculated based on the low-dimensional feature vectors. The initial weight coefficients are positively correlated with the degree of adaptation of the expert model to the corresponding task. S152, normalize the initial weight coefficients of each expert model to obtain normalized weight coefficients; S153 calculates the weighted sum of the prediction results from multiple expert models and the normalized weight coefficients to obtain the vector map prediction result.

[0048] First, obtain the visual semantic features and spatial structural features extracted by the two modality-specific encoders, as well as the relationship query of the fused location information generated above. Corresponding relational location fusion features. The three types of features are preprocessed. Visual semantic features, spatial structure features, and relational location fusion features are mapped to the same feature dimension through a linear layer. Then, the three types of features are integrated into a joint input feature through feature concatenation. This joint input feature covers multimodal basic information and optimized relational location information, providing comprehensive data input for the subsequent reweighting module.

[0049] Next, a reweighting module is designed, which includes a feature compression layer, an attention calculation layer, and a weight normalization layer. The feature compression layer uses global average pooling to compress the dimensionality of the joint input features, mapping high-dimensional features to low-dimensional feature vectors while preserving core feature information. The attention calculation layer takes the compressed feature vectors as input and calculates the initial weight coefficients of each expert model through two fully connected layers and a sigmoid activation function. The magnitude of the initial weight coefficients is positively correlated with the degree of adaptation of the expert model to the corresponding task. The weight normalization layer uses the Softmax function to normalize the initial weight coefficients, ensuring that the sum of the weight coefficients of all expert models is 1, and finally outputs the set of normalized weight coefficients for each expert model. Where n is the number of modes, to These correspond to the weight coefficients of each expert model in the interactive expert pool. For example, when the number of expert models is 5, the value of n can be 3.

[0050] Then, the vector map prediction results output by each expert model in the interactive expert pool are obtained. The prediction result of each expert model includes the probability of map element category and the coordinate information of element points. The prediction results of all expert models constitute a set. ,in to Each prediction result corresponds to the prediction results of the respective expert model. A weighted summation is performed based on the normalized weight coefficient set W and the prediction result set Y. The prediction result of each expert model is multiplied by its corresponding weight coefficient, and then all products are summed element-wise to obtain the final vector map prediction result. This result integrates the prediction advantages of each expert model, ensuring the accuracy of map element categories and point coordinates.

[0051] To further improve the prediction accuracy of the model, this embodiment also acquires preset map prior data. The vector map prediction results include the predicted map element category probabilities, predicted point coordinates, and corresponding confidence scores. The map prior data includes the categories of labeled baseline map elements and the coordinates of baseline points within the target area. The corrected vector map prediction results will also be evaluated based on the map prior data, and the expert performance information of each expert model will be analyzed. This will then update the loss function during the expert model training process, thereby further optimizing the expert model training.

[0052] Specifically, the loss function used during the training of each expert model is obtained by summing the task loss terms corresponding to each expert model according to their respective weights, and then adding them to the penalty term. Please refer to [link / reference]. Figure 6 The online vector map construction method provided in this embodiment also includes the following steps: S21, Align and correct the vector map prediction results with the prior map data to generate corrected vector map prediction results; S22, Based on the corrected vector map prediction results and prior map data, obtain the evaluation results of the corrected vector map prediction results under multiple preset evaluation indicators; S23, Based on the analysis of the prediction results of the vector map, obtain the expert performance information of each expert model under the corresponding weight coefficient; S24, Adjust the weights of each task loss term in the loss function based on expert performance information, and adjust the penalty coefficient of the penalty term in the loss function according to the evaluation results of multiple evaluation indicators; S25, use the adjusted loss function to optimize and train each expert model.

[0053] In this embodiment, the vector map prediction result is aligned and corrected with the prior map data to generate a corrected vector map prediction result. Specifically, this step can be implemented in the following way: The iterative nearest point algorithm is used to calculate the coordinate transformation matrix between the vector map prediction result and the map prior data. Based on the coordinate transformation matrix, the vector map prediction result is mapped to the coordinate system of the map prior data. The coordinate deviation between the predicted point coordinates in the vector map prediction result and the reference point coordinates in the map prior data is calculated. Based on the coordinate deviation, the predicted point coordinates in the vector map prediction result are aligned and corrected to generate the corrected vector map prediction result.

[0054] In this embodiment, the vector map prediction result and the high-precision map prior data are processed using coordinate system unification. The iterative nearest point algorithm is used to calculate the coordinate transformation matrix between the two, and the vector map prediction result is mapped to the coordinate system of the map prior data through coordinate matrix transformation. Then, using the map elements in the high-precision map prior data as a reference, the coordinate deviation between the predicted point coordinates and the reference point coordinates is calculated. The predicted point coordinates are corrected according to the deviation value to obtain the aligned and corrected single-frame vector map prediction result.

[0055] In this embodiment, the step of obtaining the evaluation results of the corrected vector map prediction results under multiple preset evaluation indicators can be achieved in the following way: Based on the multi-frame corrected vector map prediction results, a multi-frame prediction result set is constructed. For map elements of the same type in the prediction result set, a non-maximum suppression strategy is used for fusion processing to remove redundant and overlapping map elements in the multi-frame prediction result set, generating a fused regional vector map draft. Based on the regional vector map draft and the map prior data, the evaluation results of multiple preset evaluation indicators are calculated.

[0056] In this embodiment, multiple consecutive frames of aligned and corrected vector map prediction results are collected to construct a multi-frame prediction result set. For map elements of the same type in the prediction result set, a non-maximum suppression strategy is used for fusion processing. Specifically, firstly, a confidence threshold is set, and predicted map elements with confidence levels higher than the threshold are selected. Then, the intersection-union ratio (IUR) between similar predicted map elements is calculated. When the IUR is greater than a preset threshold, the predicted map element with the highest confidence level is retained, and other overlapping map elements are removed. For predicted map elements with IUR lower than the threshold, they are directly retained in the fusion result. Through the above operations, redundant and overlapping elements in the multi-frame prediction results are removed, generating a preliminary draft of the fused regional vector map.

[0057] Then, based on the initial draft of the regional vector map and prior map data, the evaluation results of multiple preset evaluation indicators are calculated.

[0058] Several evaluation metrics include element category accuracy, mean deviation of point coordinates, and element recall. Element category accuracy is obtained by calculating the ratio of the number of correctly predicted map elements to the total number of map elements. Mean deviation of point coordinates is obtained by calculating the mean Euclidean distance between the predicted point coordinates and the reference point coordinates. Element recall is obtained by calculating the ratio of the number of correctly predicted reference map elements to the total number of reference map elements.

[0059] In this embodiment, the accuracy feedback draft after map update can be generated by combining the above three types of evaluation indicators.

[0060] Furthermore, in this embodiment, the expert performance information of each expert model under the corresponding weight coefficient can be obtained based on the analysis of the prediction results of the vector map.

[0061] Specifically, based on the vector map prediction results obtained by combining the weight coefficients of each expert model with the prediction results, interpretability analysis is conducted using the weight coefficients of the expert models. This includes local interpretability analysis and global interpretability analysis. Local interpretability analysis calculates the proportion of each expert model's weight coefficient in the prediction process for a single sample, clarifying the contribution of each expert model to the prediction result and generating a distribution table of expert contributions for that single sample. Global interpretability analysis calculates the mean and variance of the weight coefficients of all expert models corresponding to all samples, summarizing the performance patterns of each expert model under different scenarios and generating a statistical report on expert performance at the dataset level. The results of both local and global interpretability analysis together constitute the interpretability basis for model decision-making, providing support for model optimization and troubleshooting.

[0062] Based on this, the performance of expert models in this round of interactive expert pools can be quantitatively evaluated by combining the results of local and global interpretability analysis. Expert performance information is obtained based on the mean of weight coefficients, variance of weight coefficients, and prediction contribution of each expert model in multi-frame sample prediction.

[0063] The mean of the weight coefficients represents the overall participation of the expert model in this round of tasks, the variance of the weight coefficients represents the stability of the expert model's performance in different scenarios, and the prediction contribution is calculated by the similarity between the prediction results of the expert model and the final vector map prediction results.

[0064] Based on the multi-dimensional information in the expert performance data, each expert model is scored, an expert performance evaluation table is generated, and the types of expert models with excellent performance and those that need to be optimized are identified.

[0065] Based on this, the weights of each task loss term in the loss function during the expert model training process can be adjusted according to expert performance information, and the penalty coefficient of the penalty term in the loss function can be adjusted according to the evaluation results of multiple evaluation indicators.

[0066] For example, based on the expert performance evaluation table, the weights of the task loss terms corresponding to different types of expert models can be adjusted. For expert models with poor performance, the weights of their specialized training loss terms can be increased, while for expert models with excellent performance, the weights of the task loss terms can be maintained or appropriately reduced.

[0067] Simultaneously, based on the map update accuracy metric, the penalty coefficients in the penalty terms of the loss function are adjusted. These penalty terms include category prediction penalty and coordinate prediction penalty. When the category accuracy is low, the penalty coefficient for category prediction is increased; when the point coordinate deviation is large, the penalty coefficient for coordinate prediction is increased. Through these parameter adjustments, the training focus of the next round of interactive expert pools is clarified, providing optimization directions for subsequent iterations and driving the model to gradually improve map construction accuracy and stability.

[0068] In summary, the online vector map construction method provided in this embodiment firstly acquires high-precision RGB images of the road environment and LiDAR point clouds through multimodal sensors, and then extracts visual semantic features and spatial structural features of the two types of data using modal-specific encoders. The image modal features are mapped to the bird's-eye view space through "depth prediction-3D projection-BEV aggregation", and the LiDAR modal features are mapped to the bird's-eye view space through "point cloud encoding-height flattening-BEV matching". The two generate high-precision BEV features for the corresponding modalities. Based on the high-precision BEV features, a hybrid expert framework is introduced to construct an interactive expert pool, and a weakly supervised interactive loss function is designed to achieve specialized training of the expert model.

[0069] Secondly, based on the generated high-precision BEV features, a relationship query mechanism is constructed by combining the preliminary interaction features output by the hybrid expert pool, and the query is divided into instance level and point level; sine-cosine position coding technology is introduced to transform the high-precision BEV coordinates of map element reference points into learnable position features, which are then integrated with the relationship query.

[0070] Then, a reweighting module is designed, which takes the visual semantic features and spatial structure features extracted in the first step and the relationship and position fusion features optimized in the second step as joint inputs, calculates the weight coefficients of each expert model, and calculates the weights of the expert prediction results based on the weight coefficients to obtain the final vector map prediction results. Local and global interpretability analysis is carried out based on the expert weights.

[0071] Finally, the vector map prediction results obtained in the third step are aligned and corrected with the prior high-precision map. The prediction results of multiple frames are fused using a non-maximum suppression strategy to update the regional map. Based on the accuracy feedback and interpretability analysis results after the map update, the performance of the experts in this round of interactive expert pool is quantitatively evaluated. The evaluation results are fed back to the weakly supervised interactive loss function in the first step to adjust the parameters and focus of the next round of training.

[0072] This solution employs two main approaches. First, it designs a dedicated expert pool for interaction, comprising expert models focused on processing single-modal information, capturing shared redundant information across multiple modalities, mining emerging collaborative information generated by multimodal collaboration, and analyzing the interaction between overall BEV features and other modal features. Combined with a weakly supervised training mechanism, this achieves refined modeling of multimodal interactions, fully releasing the information value of different modal data. Second, it proposes a unified relationship query mechanism that simultaneously captures spatial relationships between map elements, connection patterns of points within the same element, and the attribution relationships between elements and points. Furthermore, it incorporates sine-cosine position encoding technology to enhance the utilization of reference point coordinates, effectively improving position representation accuracy. Moreover, this method leverages expert weights to achieve interpretability analysis at both the single-sample and dataset levels, clearly tracing the model's decision-making logic without relying on expensive additional hardware input, making it suitable for the lightweight deployment requirements of mass-produced vehicles. This method improves multimodal data utilization and map construction accuracy, while also offering interpretability, making it applicable to online map update scenarios for autonomous driving.

[0073] Based on the same inventive concept, please refer to Figure 7 This invention also provides a functional module diagram of an online vector map construction system based on a hybrid expert model and relational coding. This embodiment can divide the online vector map construction system into functional modules according to the above method embodiments. For example, each function can be divided into its own functional modules, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module. It should be noted that the module division in this embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0074] For example, when dividing functional modules according to their respective functions, Figure 7 The online vector map construction system shown is only a schematic diagram. The system may include a data acquisition and mapping module, an iterative training module, a query and construction module, and a prediction module. The functions of each module are described in detail below.

[0075] The acquisition and mapping module is used to acquire RGB images and LiDAR point clouds of the road environment. It uses corresponding encoders to extract feature information from the RGB images and LiDAR point clouds respectively, and maps them to the bird's-eye view space to obtain BEV features. Based on the feature information and BEV features, a feature set containing multimodal features is constructed. The iterative training module is used to build multiple expert models, generate preliminary interactive features based on each expert model and corresponding task processing according to the feature set, and iteratively train each expert model according to the constructed loss function to obtain an expert model that meets the preset requirements. The query construction module is used to construct relational queries, including instance-level queries and point-level queries, based on BEV features and preliminary interaction features. It also uses location encoding technology to convert map element reference points into learnable location features and fuses the converted location features with the relational queries to obtain relational queries with fused location information. The prediction module is used to concatenate the relationship query of feature information and fused location information to obtain joint input features, determine the weight coefficients of each expert model based on the joint input features, and combine the weight coefficients of each expert model and the prediction results to obtain the vector map prediction results.

[0076] The online vector map construction system provided in this embodiment can be used to execute the online vector map construction method under any of the above embodiments. For details not covered in this embodiment, please refer to the corresponding descriptions in the above embodiments. This embodiment will not elaborate further here.

[0077] Please see Figure 8 This is a structural block diagram of an electronic device provided in an embodiment of the present invention. The electronic device can be an in-vehicle controller or a cloud server. The electronic device includes a memory, a processor, and a communication module. The memory, processor, and communication module are electrically connected to each other directly or indirectly to realize data transmission or interaction. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0078] The memory is used to store computer programs or data. Memory can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc.

[0079] The processor is used to read / write data or programs stored in the memory and to execute the online vector map construction method provided in any embodiment of the present invention.

[0080] The communication module is used to establish communication connections between electronic devices and other communication terminals via a network, and to send and receive data via the network.

[0081] It should be understood that, Figure 8 The structure shown is only a schematic diagram of an electronic device; the electronic device may also include components that are larger than those shown. Figure 8 The more or fewer components shown, or having the same Figure 8 The different configurations shown.

[0082] Furthermore, embodiments of the present invention also provide a computer-readable storage medium storing machine-executable instructions, which, when executed, implement the online vector map construction method provided in the above embodiments.

[0083] Specifically, the computer-readable storage medium can be a general-purpose storage medium, such as a removable disk or hard disk. When the computer program on the computer-readable storage medium is run, it can execute the above-described online vector map construction method. The processes involved in the execution of the executable instructions on the computer-readable storage medium can be referred to the relevant descriptions in the above method embodiments, and will not be detailed here.

[0084] In summary, the online vector map construction method and system provided in this invention, by integrating a hybrid expert framework and a relational coding mechanism, can not only finely distinguish the unique, redundant, and collaborative interaction types of multimodal data and fully explore the value of multimodal information, but also accurately capture the spatial relationships between instances, the topological relationships of points within instances, and the dependency relationships between instances and points, avoiding element distortion or loss. Simultaneously, relying on expert weights to achieve local and global interpretability analysis provides a clear basis for model decision-making, effectively alleviating the problems of poor robustness, position prediction drift, and "black box output" in complex scenarios of existing methods. This method provides a sound theoretical foundation and technical support for real-time updates of online high-precision maps for autonomous driving, without relying on expensive additional sensors, and balances lightweight deployment with engineering practicality.

[0085] In the embodiments provided by this invention, it should be understood that the disclosed apparatus and method can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and there may be other division methods in actual implementation. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interface; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0086] Furthermore, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0087] Furthermore, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0088] It should be noted that if the functionality is implemented as a software module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0089] The above are merely embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An online vector map construction method based on hybrid expert model and relation encoding, characterized in that, The method includes: RGB images and LiDAR point clouds of the road environment are acquired. The corresponding encoders are used to extract the feature information of the RGB images and LiDAR point clouds respectively, and they are mapped to the bird's-eye view space to obtain BEV features. Based on the feature information and BEV features, a feature set containing multimodal features is constructed. Multiple expert models are constructed, and preliminary interaction features are generated by performing corresponding task processing based on each expert model and according to the feature set. Each expert model is iteratively trained according to the constructed loss function to obtain an expert model that meets the preset requirements. Based on the BEV features and preliminary interaction features, a relational query including instance-level query and point-level query is constructed. The location encoding technology is used to convert map element reference points into learnable location features. The converted location features are then fused with the relational query to obtain a relational query with fused location information. The relationship query between the feature information and the fused location information is concatenated to obtain the joint input feature. The weight coefficients of each expert model are determined based on the joint input feature. The vector map prediction result is obtained by combining the weight coefficients of each expert model and the prediction result. 2.The online vector map construction method based on hybrid expert model and relation encoding according to claim 1, characterized in that, The loss function is obtained by summing the task loss terms corresponding to each expert model according to their respective weights, and then adding them to the penalty term. The method further includes: The vector map prediction results are aligned and corrected with the prior map data to generate corrected vector map prediction results. Based on the corrected vector map prediction results and the prior map data, the evaluation results of the corrected vector map prediction results under multiple preset evaluation indicators are obtained. Based on the analysis of the vector map prediction results, the expert performance information of each expert model under the corresponding weight coefficients is obtained; The weights of each task loss term in the loss function are adjusted based on the expert performance information, and the penalty coefficients of the penalty terms in the loss function are adjusted based on the evaluation results of the multiple evaluation indicators. The expert models are then optimized and trained using the adjusted loss function.

3. The online vector map construction method based on hybrid expert model and relational coding according to claim 1, characterized in that, The RGB image is used to record visual information about the road environment, and the lidar point cloud is used to record spatial structure information about the road environment. The step of extracting feature information from the RGB image and the LiDAR point cloud using corresponding encoders and mapping it to the bird's-eye view space to obtain BEV features includes: An encoder composed of a convolutional neural network is used to extract features from the RGB image to obtain visual semantic features; An encoder composed of a deep learning network is used to extract features from the lidar point cloud to obtain spatial structure features; The visual semantic features and spatial structural features are converted to the same viewpoint through a viewpoint conversion algorithm, and then mapped to the bird's-eye view space through feature projection operation to generate BEV features.

4. The online vector map construction method based on hybrid expert model and relational coding according to claim 3, characterized in that, The multiple expert models include expert models corresponding to visual semantic features and spatial structural features respectively, expert models that capture shared redundant information among multimodal features, expert models that capture emerging collaborative information after multimodal combination, and expert models that integrate BEV features with other modal feature interaction information. Each of the expert models includes a fusion model and a prediction head. The fusion model is used to perform feature fusion on the input feature set and output fusion features adapted to the corresponding task. The prediction head is used to output preliminary interaction features corresponding to the interaction type based on the fusion features. The preliminary interaction features include map element categories and point coordinate information. When iteratively training an expert model for corresponding visual semantic features or spatial structure features, the modal features other than the target modal features in the feature set are replaced with random vectors. The prediction results under the input of the target modal features are taken as positive examples and the prediction results under the input of other modal features are taken as negative examples. Iterative training is carried out by constraining the direction in which the similarity between the expert model output and the positive examples is higher than the similarity with the negative examples. When iteratively training an expert model that captures shared redundant information among multiple modalities, visual semantic features and spatial structural features are replaced with random vectors respectively. The prediction results under all single-modal feature inputs are taken as positive examples, and the deviation between the expert model output and all positive examples is constrained within a preset range by the mean squared error loss function for iterative training. When iteratively training an expert model that captures emerging collaborative information after multimodal combination, each modal feature is replaced with a random vector. The prediction results under all single-modal feature inputs are taken as negative examples. The cosine similarity loss function is used to constrain the direction in which the similarity between the expert model output and all negative examples is lower than a preset threshold for iterative training. When iteratively training an expert model that integrates BEV features with interaction information from other modal features, the BEV features and other modal features are masked separately. The prediction results with the masked features as common input are considered positive examples, and the prediction results with single modal feature input are considered negative examples. The expert model is iteratively trained using the cross-entropy loss function.

5. The online vector map construction method based on hybrid expert model and relational coding according to claim 1, characterized in that, The step of constructing a relational query, including instance-level and point-level queries, based on the BEV features and preliminary interaction features includes: The preliminary interaction features are incorporated into the BEV features through feature concatenation operations to form the BEV basic features that integrate interaction information. Based on the BEV basic features, the association weights of different map elements in the bird's-eye view space are calculated through a self-attention mechanism to construct an instance-level query. Based on the BEV basic features, the association weights between points within the same map element are calculated using a self-attention mechanism to obtain the topological relationships of points within the map element. The association weights between points and their respective map elements are calculated using a cross-attention mechanism to obtain the subordinate relationships between instances and points, thereby constructing a point-level query.

6. The online vector map construction method based on hybrid expert model and relation coding according to claim 1, characterized in that, The step of using location encoding technology to convert map element reference points into learnable location features, and then fusing the converted location features with the relationship query to obtain a relationship query with fused location information, includes: Obtain the coordinate values ​​of map element reference points in the bird's-eye view space, and use a preset location encoding formula to map different dimensions of the coordinate values, converting them into learnable location features with location distinguishability; Adjust the dimensions of the transformed location features, instance-level queries, and point-level queries in the relational query to be consistent; By adding elements together, the location features are incorporated into the instance-level query and the point-level query respectively, resulting in instance-level query and point-level query with fused location information. The instance-level and point-level queries of the merged location information are integrated to form the final relational query of the merged location information.

7. The online vector map construction method based on hybrid expert model and relational coding according to claim 1, characterized in that, The step of determining the weight coefficients of each expert model based on the joint input features, and combining the weight coefficients of each expert model with the prediction results to obtain the vector map prediction results includes: The joint input features are compressed to map into low-dimensional feature vectors. The initial weight coefficients of each expert model are calculated based on the low-dimensional feature vectors. The initial weight coefficients are positively correlated with the degree of adaptation of the expert model to the corresponding task. The initial weight coefficients of each expert model are normalized to obtain normalized weight coefficients. The vector map prediction result is obtained by weighting and summing the prediction results of multiple expert models and normalized weight coefficients.

8. The online vector map construction method based on hybrid expert model and relation coding according to claim 2, characterized in that, The vector map prediction results include the predicted map element category probability, predicted point coordinates and corresponding confidence scores, and the map prior data includes the baseline map element category and baseline point coordinates. The step of aligning and correcting the vector map prediction result with the prior map data to generate a corrected vector map prediction result includes: The coordinate transformation matrix between the vector map prediction result and the map prior data is calculated using the iterative nearest point algorithm, and the vector map prediction result is mapped to the coordinate system of the map prior data based on the coordinate transformation matrix; Calculate the coordinate deviation between the predicted point coordinates in the vector map prediction result and the reference point coordinates in the map prior data, and align and correct the predicted point coordinates in the vector map prediction result based on the coordinate deviation to generate a corrected vector map prediction result.

9. The online vector map construction method based on hybrid expert model and relation coding according to claim 2, characterized in that, The step of obtaining the evaluation results of the corrected vector map prediction results under multiple preset evaluation indicators based on the corrected vector map prediction results and the map prior data includes: Based on the multi-frame corrected vector map prediction results, a multi-frame prediction result set is constructed. For map elements of the same type in the prediction result set, a non-maximum suppression strategy is used for fusion processing to remove redundant and overlapping map elements in the multi-frame prediction result set, and a first draft of the fused regional vector map is generated. Based on the initial draft of the regional vector map and the prior map data, the evaluation results of multiple preset evaluation indicators are calculated. The multiple evaluation metrics include element category accuracy, average point coordinate deviation, and element recall. The element category accuracy is obtained by statistically analyzing the ratio of the number of correctly predicted map elements to the total number of map elements. The average point coordinate deviation is obtained by calculating the average Euclidean distance between the predicted point coordinates and the reference point coordinates. The element recall is obtained by calculating the ratio of the number of correctly predicted reference map elements to the total number of reference map elements.

10. The online vector map construction method based on hybrid expert model and relational coding according to claim 2, characterized in that, The expert performance information is obtained based on the mean of the weight coefficients, the variance of the weight coefficients, and the prediction contribution of each expert model in multi-frame sample prediction. The mean of the weight coefficients represents the overall participation of the expert model in this round of tasks, the variance of the weight coefficients represents the performance stability of the expert model in different scenarios, and the prediction contribution is calculated by the similarity between the prediction results of the expert model and the final vector map prediction results.

Images