Robust online vector map construction method for space-time simplified query

Abstract

The invention discloses a robust online vector map construction method for space-time simplified query, and relates to the technical field of automatic driving perception, and the method comprises the steps: fusing multi-view features into a unified bird's-eye view for representation; matching instance features of the current frame and the historical frame based on time association; updating query features and optimizing query representations; a high-quality vector map construction model is used as a teacher model, a map construction model focused on taking an instance as a center is used as a student model, and features of a teacher network are adaptively distilled to a student network; and designing a loss function to evaluate the student model, performing training optimization, and if the output requirement is not met, further optimizing the student model through a middle distance perception mechanism. According to the method, the robustness, the precision and the efficiency of online vector map construction are remarkably improved, especially the performance in a dynamic environment and a complex scene, and by means of the mode, multi-view, space-time correlation and complex 3D target detection tasks can be better processed.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving perception technology, and in particular to a robust online vector map construction method for spatiotemporal simplified query. Background Art

[0002] With the rapid development of artificial intelligence technology, the requirements for autonomous driving technology in terms of perception accuracy, environmental adaptability and safety are constantly increasing. In complex dynamic scenarios, especially driving environments with high dynamics and multi-target interactions, achieving high-precision, real-time environmental perception and map construction has become a major challenge for autonomous driving technology. Vector map construction, as a method of representing environmental instances through unified geometric and semantic information, provides key support for scene understanding and decision-making in autonomous driving. However, existing technologies still have many shortcomings when processing dynamic environments and large-scale multi-view data.

[0003] Currently, vector map construction methods based on multi-view 2D images can achieve some success in simple static scenes. However, when dealing with complex scenes with occlusion, dynamic objects, and perspective changes, they often suffer from insufficient accuracy due to insufficient feature fusion. On the other hand, 3D reconstruction methods based on LiDAR and image fusion, while offering higher geometric accuracy, suffer from complex data processing and significant computational resource consumption, making them difficult to meet real-time requirements. Furthermore, existing methods lack the ability to consistently associate object instances and model spatiotemporal features during multi-frame data fusion, resulting in discontinuities and decreased accuracy in map construction results in dynamic scenes. Furthermore, existing methods lack redundant query optimization and lightweight design in map construction tasks, further limiting their applicability in resource-constrained scenarios.

[0004] In summary, there is an urgent need for a vector map construction method that combines streamlined spatiotemporal querying with efficient instance-based feature modeling to achieve high-precision, continuous, and real-time map construction in complex and dynamic scenarios. Furthermore, by introducing knowledge distillation and lightweight optimization mechanisms, we can address the high requirements for resource adaptability and model robustness in practical applications, further promoting the popularization and application of autonomous driving technology in real-world scenarios. Summary of the Invention

[0005] In order to solve the above technical problems, a robust online vector map construction method for spatiotemporal simplified query is provided. This technical solution solves the above problems.

[0006] In order to achieve the above objects, the technical solution adopted by the present invention is: A robust online vector map construction method for spatiotemporal simplified query, characterized by comprising: S1: Based on multi-view image preprocessing, feature extraction and geometric correction, multi-view features are fused into a unified bird's-eye view representation to describe the geometric structure and semantic details of the scene; S2: Matches instance features of the current frame and historical frames based on temporal association. Simultaneously, the spatial fusion module combines the cross-attention mechanism to fuse the historical point set with the current point set to construct a globally continuous and high-precision map representation. S3: Based on query classification analysis, it identifies and prunes queries with low contributions, retains high-quality queries, interacts with the self-attention module, and integrates the cross-attention mechanism with image features to update query features and optimize query representation. S4: In the knowledge distillation stage, a high-quality vector map construction model is used as the teacher model, and an instance-centric map construction model is used as the student model to adaptively distill the features of the teacher network to the student network. S5: Design a loss function to evaluate the student model and perform training optimization. If the output requirements are not met, further optimize the student model through the mid-range perception mechanism.

[0007] Preferably, the S1 specifically includes: Input the synchronously acquired multi-view camera image sequences into the system to obtain scene data containing dynamic changes in the surrounding environment; The scene data is input and the image is preprocessed, including distortion correction and time synchronization, to ensure alignment between multi-view images. Based on the intrinsic and extrinsic parameters of the camera, the image of each view is geometrically calibrated to construct a globally consistent environment representation. Based on the deep learning backbone network, the image of each perspective is extracted, and the image data is converted into a high-dimensional feature tensor. The low-order features are extracted through the convolution layer to capture the texture and edge details in the image. High-order features are modeled through a multi-head attention mechanism, focusing on global relationships and contextual associations between different objects in the scene, generating a high-dimensional feature representation tensor corresponding to each viewpoint; Based on the vehicle's geometric information, the extracted high-dimensional features are uniformly aligned in space through projection transformation, and the multi-view features are projected into a unified bird's-eye view coordinate system using geometric correction. Through multi-view feature fusion and coordinate system conversion, the features from multiple cameras are aligned into a two-dimensional plane representation. In the BEV coordinate system, each pixel contains high-dimensional feature information generated by multi-view fusion.

[0008] Preferably, the S2 specifically includes: Based on the BEV features, the instance-centric time association module is used to match the instance features of the current moment with those of the historical moments. Feature alignment technology is used to associate and transform the historical instances in the geometric and feature dimensions, unifying them and the current instance into the same semantic space. Through a learnable temporal association mechanism, the matching relationship between historical instances and current instances is dynamically modeled in a high-dimensional feature space, capturing feature association changes across time series. Based on the instance-centered spatial fusion module, the point set features of the historical frame are spatially aligned with the point set features of the current frame, and the local features of the historical point set and the current point set are efficiently aligned and fused through the cross-attention mechanism; Perform spatial sampling on the historical global map point set, apply the cross-attention operation to the fusion process of detection results and historical features, strengthen the interaction relationship between instances in the spatial domain, and optimize the fusion result through curve fitting technology; Based on the temporal correlation features and spatial fusion features, the temporal and spatial dimensions are jointly modeled, and the changing trends of instances in the temporal and spatial domains are captured through a multi-layer network design; In the time dimension, based on the dynamically updated instance memory buffer, the historical information of multi-frame features is combined to perform real-time association with the current instance and generate a unified instance identifier; In the spatial dimension, a three-layer cross-attention mechanism is used to deeply interactively model current and historical features, strengthening the spatial consistency of global map features. Combined with a classification optimization strategy, different types of map instances are finely processed. Based on the optimized instance features, a global vectorized map representation is generated, integrating geometric, temporal, and semantic information to describe the details and spatial relationships of each instance in the global map.

[0009] Preferably, the S3 specifically includes: Based on query classification analysis, we identify queries with smaller contributions and gradually prune them, removing redundant queries to optimize computational efficiency while retaining high-quality queries that contribute significantly to map construction. By dynamically sorting and performing multi-layer analysis on the query classification scores, redundant queries with less impact on map building goals are screened out; A gradual refinement strategy is introduced during pruning, and the retained high-quality queries interact with each other through the self-attention module, which effectively captures the global relationship between queries. The optimized query features are fused with the historical map sampling points, and the current query features and the historical features are unified into the same spatial domain based on geometric alignment and feature enhancement strategies; Perform in-depth analysis of the feature distribution of historical sampling points, extract key points that are representative and contain high information, correct the position deviation of historical features in the current coordinate system, and strengthen the semantic consistency with the current query features; Through the feature enhancement strategy, the model dynamically adjusts the fusion weights of historical features and current features, introduces a dynamic feature weighting mechanism, and fine-grainedly controls the fusion process based on the semantic importance and contribution of features; Through a recursive update strategy, the feedforward optimized features are integrated layer by layer to maintain consistency and globality in both temporal and spatial dimensions. The geometric semantic information of historical sampling points is combined to optimize the generated map representation, thereby ensuring its high accuracy and smoothness in complex scenarios.

[0010] Preferably, the S4 specifically includes: In the knowledge distillation phase, a model is built based on a high-quality vector map and used as a teacher model to provide high-precision feature guidance of the global map to the student model. Through temporal association, spatial fusion, and efficient pruning, a high-quality vector map containing geometric, semantic, and temporal information is generated. During the distillation process, an adaptive weight mechanism is used to extract high-quality global feature representations from the teacher network, including multi-frame temporal information, global geometric semantics, and spatial relationships between instances; The teacher model's features are divided into three parts: global features, regional features, and instance features. A specific distillation loss function is used to optimize the knowledge transfer of each part. Global feature distillation guides the student model to capture the entire scene by constructing a holistic contextual representation. Regional feature distillation helps the student network form accurate representations in key areas. Instance feature distillation focuses on improving the student model's robust detection capabilities for target instances. Adopting the instance center alignment strategy, the features of the student model are embedded into the semantic space consistent with the teacher model through geometric and semantic mapping mechanisms; During the distillation process, an adaptive loss balancing module is designed to dynamically adjust the distillation weights of global features and instance features. This module allocates weights based on scene complexity, number of instances, and feature contribution, allowing the student model to prioritize learning key instance information in complex scenarios while taking into account globally consistent expression in simple scenarios. In addition, in response to the unique lightweight requirements of the student model, the distillation process uses multi-layer feature compression technology to transfer the deep semantic features of the teacher model to the student model in a lightweight representation, thereby significantly reducing computing resource usage while retaining key feature information.

[0011] Through knowledge distillation, the student model is constructed around instance-centered key regions without relying on complex global feature modeling, achieving efficient and robust 3D object detection.

[0012] Preferably, the S5 specifically includes: Based on the student model, a loss function is designed to comprehensively evaluate and optimize the student model to ensure its robustness and high-precision 3D object detection capabilities; The loss function includes classification loss and regression loss, which are used to evaluate the target category prediction ability and 3D bounding box regression accuracy of the student model respectively; For classification accuracy, the quality focus loss function is used to calculate the deviation between the target category probability distribution predicted by the student model and the true category label; For regression accuracy, a smooth L1 loss function is used to evaluate the difference between the 3D bounding box parameters predicted by the student model and the true value; The classification loss and regression loss are combined into a total loss through weighted summation to guide the optimization of the student model. The total loss calculation formula is: L total =L cls +L reg According to the loss function evaluation index, if the output requirements are not met, the mid-range perception mechanism is further used to optimize the student model. Through distance perception response distillation, the loss weights of different targets are dynamically adjusted, focusing on optimizing the detection accuracy of medium and short-range targets. The specific formula is: Among them, d i is the distance between target i and sensor, w i is the normalized weight of the target, μ is the fixed weight of the distant target, and A, B, C and δ are adjustment parameters.

[0013] According to the distance perception response distillation strategy, the student model is trained and optimized multiple times, and the evaluation index of the loss function is recalculated. If L total If the threshold requirement is ≤0.1, it is considered that the output accuracy requirement is met, which ensures that the student model can efficiently perform 3D object detection tasks in resource-constrained environments, while maintaining a high level of robustness and accuracy in different complex scenarios, and finally outputs the instance-centric vector map construction result constructed by the student model.

[0014] Preferably, the method of using a quality focus loss function to calculate the deviation between the target category probability distribution predicted by the student model and the true category label for classification accuracy specifically includes: Among them, the calculation formula of quality focus loss function is: Where S is the total number of targets, p ti and p si are the predicted probabilities of the teacher model and the student model, wi For target i.

[0015] Preferably, the use of a smooth L1 loss function to evaluate the difference between the 3D bounding box parameters predicted by the student model and the true value for regression accuracy specifically includes: Among them, the formula of smooth L1 loss function is: Where r ti and r si are the bounding box regression parameters of the teacher model and the student model respectively, S is the total number of targets, and SmoothL1 is the smooth L1 loss function.

[0016] Preferably, if the output requirement is not met according to the evaluation indicator of the loss function, the student model is further optimized using the medium-distance perception mechanism, and the loss weights of different targets are dynamically adjusted through distance perception response distillation, focusing on optimizing the detection accuracy of medium- and short-distance targets. Specifically, the following steps are included: The dynamic adjustment formula is: Among them, d i is the distance between target i and sensor, w i is the normalized weight of the target, μ is the fixed weight of the distant target, and A, B, C and δ are adjustment parameters.

[0017] Compared with the prior art, the present invention has the following beneficial effects: The present invention proposes to significantly improve the robustness, accuracy and efficiency of online vector map construction, especially in dynamic environments and complex scenes. In this way, it can better handle multi-view, spatiotemporal correlation and complex 3D object detection tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a step flow chart of the present invention; Figure 2 This is a step flow chart of S2 in the present invention. DETAILED DESCRIPTION

[0019] The following description is intended to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are merely examples, and those skilled in the art may conceive of other obvious variations.

[0020] Reference Figure 1 As shown, a robust online vector map construction method for spatiotemporal simplified query is characterized by comprising: S1: Based on multi-view image preprocessing, feature extraction and geometric correction, multi-view features are fused into a unified bird's-eye view representation to describe the geometric structure and semantic details of the scene; S2: Matches instance features of the current frame and historical frames based on temporal association. Simultaneously, the spatial fusion module combines the cross-attention mechanism to fuse the historical point set with the current point set to construct a globally continuous and high-precision map representation. S3: Based on query classification analysis, it identifies and prunes queries with low contributions, retains high-quality queries, interacts with the self-attention module, and integrates the cross-attention mechanism with image features to update query features and optimize query representation. S4: In the knowledge distillation stage, a high-quality vector map construction model is used as the teacher model, and an instance-centric map construction model is used as the student model to adaptively distill the features of the teacher network to the student network. S5: Design a loss function to evaluate the student model and perform training optimization. If the output requirements are not met, further optimize the student model through the mid-range perception mechanism.

[0021] Reference Figure 2 As shown, S1 specifically includes: Input the synchronously acquired multi-view camera image sequences into the system to obtain scene data containing dynamic changes in the surrounding environment; The scene data is input and the image is preprocessed, including distortion correction and time synchronization, to ensure alignment between multi-view images. Based on the intrinsic and extrinsic parameters of the camera, the image of each view is geometrically calibrated to construct a globally consistent environment representation. Based on the deep learning backbone network, the image of each perspective is extracted, and the image data is converted into a high-dimensional feature tensor. The low-order features are extracted through the convolution layer to capture the texture and edge details in the image. High-order features are modeled through a multi-head attention mechanism, focusing on global relationships and contextual associations between different objects in the scene, generating a high-dimensional feature representation tensor corresponding to each viewpoint; Based on the vehicle's geometric information, the extracted high-dimensional features are uniformly aligned in space through projection transformation, and the multi-view features are projected into a unified bird's-eye view coordinate system using geometric correction. Through multi-view feature fusion and coordinate system conversion, the features from multiple cameras are aligned into a two-dimensional plane representation. In the BEV coordinate system, each pixel contains high-dimensional feature information generated by multi-view fusion. This step unifies the features of multi-view images into a bird's-eye view coordinate system, thereby achieving comprehensive information representation from multiple perspectives, which helps to improve the accuracy of global spatial information.

[0022] S2 specifically includes: Based on the BEV features, the instance-centric time association module is used to match the instance features of the current moment with those of the historical moments. Feature alignment technology is used to associate and transform the historical instances in the geometric and feature dimensions, unifying them and the current instance into the same semantic space. Through a learnable temporal association mechanism, the matching relationship between historical instances and current instances is dynamically modeled in a high-dimensional feature space, capturing feature association changes across time series. Based on the instance-centered spatial fusion module, the point set features of the historical frame are spatially aligned with the point set features of the current frame, and the local features of the historical point set and the current point set are efficiently aligned and fused through the cross-attention mechanism; Perform spatial sampling on the historical global map point set, apply the cross-attention operation to the fusion process of detection results and historical features, strengthen the interaction relationship between instances in the spatial domain, and optimize the fusion result through curve fitting technology; Based on the temporal correlation features and spatial fusion features, the spatiotemporal dimensions are jointly modeled, and the changing trends of instances in the temporal and spatial domains are captured through a multi-layer network design; In the time dimension, based on the dynamically updated instance memory buffer, the historical information of multi-frame features is combined to perform real-time association with the current instance and generate a unified instance identifier; In the spatial dimension, a three-layer cross-attention mechanism is used to deeply interactively model current and historical features, strengthening the spatial consistency of global map features. Combined with classification optimization strategies, different types of map instances are processed in detail. Based on the optimized instance features, a global vectorized map representation is generated, integrating geometric, temporal, and semantic information to describe the details and spatial relationships of each instance in the global map. This step significantly improves the feature fusion capability of historical and current frames through the design of cross-attention mechanism and instance memory buffer, and enhances the continuity and accuracy of the system in dynamic environments.

[0023] S3 specifically includes: Based on query classification analysis, we identify queries with smaller contributions and gradually prune them, removing redundant queries to optimize computational efficiency while retaining high-quality queries that contribute significantly to map construction. By dynamically sorting and performing multi-layer analysis on the query classification scores, redundant queries with less impact on map building goals are screened out; A gradual refinement strategy is introduced during pruning, and the retained high-quality queries interact with each other through the self-attention module, which effectively captures the global relationship between queries. The optimized query features are fused with historical map sampling points. Based on geometric alignment and feature enhancement strategies, the current query features and historical features are unified into the same spatial domain. Perform in-depth analysis of the feature distribution of historical sampling points, extract key points that are representative and contain high information, correct the position deviation of historical features in the current coordinate system, and strengthen the semantic consistency with the current query features; Through the feature enhancement strategy, the model dynamically adjusts the fusion weights of historical features and current features, introduces a dynamic feature weighting mechanism, and fine-grainedly controls the fusion process based on the semantic importance and contribution of features; Through a recursive update strategy, the feedforward optimized features are integrated layer by layer to maintain consistency and globality in both time and space. The geometric semantic information of historical sampling points is combined to optimize the generated map representation, thus ensuring high accuracy and smoothness in complex scenes. By introducing the self-attention mechanism and dynamic feature weighting, the global correlation between queries is enhanced, and the consistency and globality of features are ensured through a recursive update strategy, thereby improving the accuracy of map construction.

[0024] S4 specifically includes: In the knowledge distillation phase, a model is built based on a high-quality vector map and used as a teacher model to provide high-precision feature guidance of the global map to the student model. Through temporal association, spatial fusion, and efficient pruning, a high-quality vector map containing geometric, semantic, and temporal information is generated. During the distillation process, an adaptive weight mechanism is used to extract high-quality global feature representations from the teacher network, including multi-frame temporal information, global geometric semantics, and spatial relationships between instances; The teacher model's features are divided into three parts: global features, regional features, and instance features. A specific distillation loss function is used to optimize the knowledge transfer of each part. Global feature distillation guides the student model to capture the entire scene by constructing a holistic contextual representation. Regional feature distillation helps the student network form accurate representations in key areas. Instance feature distillation focuses on improving the student model's robust detection capabilities for target instances. Adopting the instance center alignment strategy, the features of the student model are embedded into the semantic space consistent with the teacher model through geometric and semantic mapping mechanisms; During the distillation process, an adaptive loss balancing module is designed to dynamically adjust the distillation weights of global and instance features. This module assigns weights based on scene complexity, number of instances, and feature contribution, enabling the student model to prioritize learning key instance information in complex scenarios while maintaining globally consistent representation in simple scenarios. Furthermore, to address the unique lightweight requirements of the student model, the distillation process uses multi-layer feature compression technology to transfer the deep semantic features of the teacher model to the student model in a lightweight form, significantly reducing computational resources while preserving key feature information. This step reduces the computational resources consumed by introducing adaptive loss balancing and lightweight feature compression strategies, while maintaining efficient 3D object detection performance without relying on complex global modeling.

[0025] Through knowledge distillation, the student model is constructed around instance-centered key regions without relying on complex global feature modeling, achieving efficient and robust 3D object detection.

[0026] S5 specifically includes: Based on the student model, a loss function is designed to comprehensively evaluate and optimize the student model to ensure its robustness and high-precision 3D object detection capabilities; The loss function includes classification loss and regression loss, which are used to evaluate the target category prediction ability and 3D bounding box regression accuracy of the student model respectively; For classification accuracy, the quality focus loss function is used to calculate the deviation between the target category probability distribution predicted by the student model and the true category label; For regression accuracy, a smooth L1 loss function is used to evaluate the difference between the 3D bounding box parameters predicted by the student model and the true value; The classification loss and regression loss are combined into a total loss through weighted summation to guide the optimization of the student model. The total loss calculation formula is: L total =L cls +L reg According to the loss function evaluation index, if the output requirements are not met, the mid-range perception mechanism is further used to optimize the student model. Through distance perception response distillation, the loss weights of different targets are dynamically adjusted, focusing on optimizing the detection accuracy of medium and short-range targets. The specific formula is: Among them, d i is the distance between target i and sensor, w i is the normalized weight of the target, μ is the fixed weight of the distant target, and A, B, C and δ are adjustment parameters.

[0027] According to the distance perception response distillation strategy, the student model is trained and optimized multiple times, and the evaluation index of the loss function is recalculated. If L total If the threshold requirement is ≤0.1, the output accuracy requirement is considered to be met, ensuring that the student model can efficiently perform 3D object detection tasks in resource-constrained environments while maintaining a high level of robustness and accuracy in different complex scenarios. The final output is the instance-centric vector map construction result constructed by the student model; This step optimizes the model's detection accuracy for targets at different distances by introducing a mid-range perception mechanism and quality focus loss, while ensuring high robustness and accuracy in resource-constrained environments.

[0028] For classification accuracy, the quality focus loss function is used to calculate the deviation between the target category probability distribution predicted by the student model and the true category label. Specifically, Among them, the calculation formula of quality focus loss function is: Where S is the total number of targets, p ti and p si are the predicted probabilities of the teacher model and the student model, w i For target i.

[0029] For regression accuracy, the smooth L1 loss function is used to evaluate the difference between the 3D bounding box parameters predicted by the student model and the actual value. Specifically, Among them, the formula of smooth L1 loss function is: Where r ti and r si are the bounding box regression parameters of the teacher model and the student model respectively, S is the total number of targets, and SmoothL1 is the smooth L1 loss function.

[0030] According to the loss function evaluation indicators, if the output requirements are not met, the mid-range perception mechanism is further used to optimize the student model. Through distance perception response distillation, the loss weights of different targets are dynamically adjusted, focusing on optimizing the detection accuracy of medium and short-range targets. Specifically, the following are included: The dynamic adjustment formula is: Among them, d i is the distance between target i and sensor, w i is the normalized weight of the target, μ is the fixed weight of the distant target, and A, B, C and δ are adjustment parameters.

[0031] In summary, the advantages of the present invention are: This method enables the construction of high-precision spatiotemporal vector maps. By fusion of multi-view images and optimization of temporal and spatial feature correlations, it ensures that the details and spatial relationships of each instance in the map are fully described, thereby providing an accurate representation of the environment. Through a multi-level attention mechanism and efficient fusion of spatial and temporal features, the system remains robust in dynamically changing environments and maintains high performance even in complex or noisy scenarios. Adopting query classification analysis and gradual pruning technology effectively reduces redundant queries, improves computing efficiency, ensures high system response speed in real-time applications, and reduces computing resource consumption. Through feature extraction and fusion based on self-attention and cross-attention mechanisms, it can better capture and optimize key features in various perspectives and spatiotemporal dimensions, enhancing the adaptability of map construction models to complex scenarios; By guiding the training of the student model through the teacher model, efficient knowledge transfer is achieved. Especially in resource-constrained situations, the student model can maintain high-precision object detection capabilities and effectively avoid the waste of computing resources. It is suitable for deployment on edge devices or in resource-constrained environments. The optimized student model not only builds high-precision maps but also performs efficient 3D object detection. By combining classification accuracy and regression accuracy evaluation methods, it ensures high-quality object detection and further improves the detection accuracy of medium- and short-range objects through the distance perception mechanism. This method uses adaptive loss functions and multi-layer feature compression technology to ensure that the student model can balance global consistency and instance accuracy. It is particularly suitable for efficient real-time mapping and target detection in complex environments with dynamic changes.

[0032] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A robust online vector map construction method for spatiotemporal simplified query, characterized by: include: S1: Based on multi-view image preprocessing, feature extraction and geometric correction, multi-view features are fused into a unified bird's-eye view representation to describe the geometric structure and semantic details of the scene; S2: Matches instance features of the current frame and historical frames based on temporal association. Simultaneously, the spatial fusion module combines the cross-attention mechanism to fuse the historical point set with the current point set to construct a globally continuous and high-precision map representation. S3: Based on query classification analysis, it identifies and prunes queries with low contributions, retains high-quality queries, interacts with the self-attention module, and integrates the cross-attention mechanism with image features to update query features and optimize query representation. S4: In the knowledge distillation stage, a high-quality vector map construction model is used as the teacher model, and an instance-centric map construction model is used as the student model to adaptively distill the features of the teacher network to the student network. S5: Design a loss function to evaluate the student model and perform training optimization. If the output requirements are not met, further optimize the student model through the mid-range perception mechanism.

2. A robust online vector map construction method for spatiotemporal simplified query according to claim 1, characterized in that: Said S1 specifically includes: Input the synchronously acquired multi-view camera image sequences into the system to obtain scene data containing dynamic changes in the surrounding environment; The scene data is input and the image is preprocessed, including distortion correction and time synchronization, to ensure alignment between multi-view images. Based on the intrinsic and extrinsic parameters of the camera, the image of each view is geometrically calibrated to construct a globally consistent environment representation. Based on the deep learning backbone network, the image of each perspective is extracted, and the image data is converted into a high-dimensional feature tensor. The low-order features are extracted through the convolution layer to capture the texture and edge details in the image. High-order features are modeled through a multi-head attention mechanism, focusing on global relationships and contextual associations between different objects in the scene, generating a high-dimensional feature representation tensor corresponding to each viewpoint; Based on the vehicle's geometric information, the extracted high-dimensional features are uniformly aligned in space through projection transformation, and the multi-view features are projected into a unified bird's-eye view coordinate system using geometric correction. Through multi-view feature fusion and coordinate system conversion, the features from multiple cameras are aligned into a two-dimensional plane representation. In the BEV coordinate system, each pixel contains high-dimensional feature information generated by multi-view fusion.

3. The robust online vector map construction method for spatiotemporal simplified query according to claim 2 is characterized in that: The S2 specifically includes: Based on the BEV features, the instance-centric time association module is used to match the instance features of the current moment with those of the historical moments. Feature alignment technology is used to associate and transform the historical instances in the geometric and feature dimensions, unifying them and the current instance into the same semantic space. Through a learnable temporal association mechanism, the matching relationship between historical instances and current instances is dynamically modeled in a high-dimensional feature space, capturing feature association changes across time series. Based on the instance-centered spatial fusion module, the point set features of the historical frame are spatially aligned with the point set features of the current frame, and the local features of the historical point set and the current point set are efficiently aligned and fused through the cross-attention mechanism; Perform spatial sampling on the historical global map point set, apply the cross-attention operation to the fusion process of detection results and historical features, strengthen the interaction relationship between instances in the spatial domain, and optimize the fusion result through curve fitting technology; Based on the temporal correlation features and spatial fusion features, the temporal and spatial dimensions are jointly modeled, and the changing trends of instances in the temporal and spatial domains are captured through a multi-layer network design; In the time dimension, based on the dynamically updated instance memory buffer, the historical information of multi-frame features is combined to perform real-time association with the current instance and generate a unified instance identifier; In the spatial dimension, a three-layer cross-attention mechanism is used to deeply interactively model current and historical features, strengthening the spatial consistency of global map features. Combined with a classification optimization strategy, different types of map instances are finely processed. Based on the optimized instance features, a global vectorized map representation is generated, integrating geometric, temporal, and semantic information to describe the details and spatial relationships of each instance in the global map.

4. The robust online vector map construction method for spatiotemporal simplified query according to claim 3 is characterized in that: The S3 specifically includes: Based on query classification analysis, we identify queries with smaller contributions and gradually prune them, removing redundant queries to optimize computational efficiency while retaining high-quality queries that contribute significantly to map construction. By dynamically sorting and performing multi-layer analysis on the query classification scores, redundant queries with less impact on map building goals are screened out; A gradual refinement strategy is introduced during pruning, and the retained high-quality queries interact with each other through the self-attention module, which effectively captures the global relationship between queries. The optimized query features are fused with the historical map sampling points, and the current query features and the historical features are unified into the same spatial domain based on geometric alignment and feature enhancement strategies; Perform in-depth analysis of the feature distribution of historical sampling points, extract key points that are representative and contain high information, correct the position deviation of historical features in the current coordinate system, and strengthen the semantic consistency with the current query features; Through the feature enhancement strategy, the model dynamically adjusts the fusion weights of historical features and current features, introduces a dynamic feature weighting mechanism, and fine-grainedly controls the fusion process based on the semantic importance and contribution of features; Through a recursive update strategy, the feedforward optimized features are integrated layer by layer to maintain consistency and globality in both temporal and spatial dimensions. The geometric semantic information of historical sampling points is combined to optimize the generated map representation, thereby ensuring its high accuracy and smoothness in complex scenarios.

5. The robust online vector map construction method for spatiotemporal simplified query according to claim 4 is characterized in that: The S4 specifically includes: In the knowledge distillation phase, a model is built based on a high-quality vector map and used as a teacher model to provide high-precision feature guidance of the global map to the student model. Through temporal association, spatial fusion, and efficient pruning, a high-quality vector map containing geometric, semantic, and temporal information is generated. During the distillation process, an adaptive weight mechanism is used to extract high-quality global feature representations from the teacher network, including multi-frame temporal information, global geometric semantics, and spatial relationships between instances; The teacher model's features are divided into three parts: global features, regional features, and instance features. A specific distillation loss function is used to optimize the knowledge transfer of each part. Global feature distillation guides the student model to capture the entire scene by constructing a holistic contextual representation. Regional feature distillation helps the student network form accurate representations in key areas. Instance feature distillation focuses on improving the student model's robust detection capabilities for target instances. Adopting the instance center alignment strategy, the features in the student model are embedded into the semantic space consistent with the teacher model through geometric and semantic mapping mechanisms; During the distillation process, an adaptive loss balancing module is designed to dynamically adjust the distillation weights of global features and instance features. This module allocates weights based on scene complexity, number of instances, and feature contribution, allowing the student model to prioritize learning key instance information in complex scenarios while taking into account globally consistent expression in simple scenarios. In addition, in response to the unique lightweight requirements of the student model, the distillation process uses multi-layer feature compression technology to transfer the deep semantic features of the teacher model to the student model in a lightweight representation, thereby significantly reducing computing resource usage while retaining key feature information.

6. Through knowledge distillation, the student model is constructed with instance-centered key regions without relying on complex global feature modeling, achieving efficient and robust 3D object detection.

7. The robust online vector map construction method for spatiotemporal simplified query according to claim 5 is characterized in that: The S5 specifically includes: Based on the student model, a loss function is designed to comprehensively evaluate and optimize the student model to ensure its robustness and high-precision 3D object detection capabilities; The loss function includes classification loss and regression loss, which are used to evaluate the target category prediction ability and 3D bounding box regression accuracy of the student model respectively; For classification accuracy, the quality focus loss function is used to calculate the deviation between the target category probability distribution predicted by the student model and the true category label; For regression accuracy, a smooth L1 loss function is used to evaluate the difference between the 3D bounding box parameters predicted by the student model and the true value; The classification loss and regression loss are combined into a total loss through weighted summation to guide the optimization of the student model. The total loss calculation formula is: L total =L cls +L reg According to the loss function evaluation index, if the output requirements are not met, the mid-range perception mechanism is further used to optimize the student model. Through distance perception response distillation, the loss weights of different targets are dynamically adjusted, focusing on optimizing the detection accuracy of medium and short-range targets. The specific formula is: Among them, d i is the distance between target i and sensor, w i is the normalized weight of the target, μ is the fixed weight of the distant target, and A, B, C and δ are adjustment parameters.

8. According to the distance perception response distillation strategy, the student model is trained and optimized multiple times, and the evaluation index of the loss function is recalculated. If L total If the threshold requirement is ≤0.1, it is considered that the output accuracy requirement is met, which ensures that the student model can efficiently perform 3D object detection tasks in resource-constrained environments, while maintaining a high level of robustness and accuracy in different complex scenarios, and finally outputs the instance-centric vector map construction result constructed by the student model.

9. The robust online vector map construction method for spatiotemporal simplified query according to claim 6, characterized in that: The classification accuracy is calculated by using the quality focus loss function to calculate the deviation between the target category probability distribution predicted by the student model and the true category label. include: Among them, the calculation formula of quality focus loss function is: Where S is the total number of targets, p ti and p si are the predicted probabilities of the teacher model and the student model, w i For target i.

10. The robust online vector map construction method for spatiotemporal simplified query according to claim 7, characterized in that: For regression accuracy, the smooth L1 loss function is used to evaluate the difference between the 3D bounding box parameters predicted by the student model and the actual value. include: Among them, the formula of smooth L1 loss function is: Where r ti and r si are the bounding box regression parameters of the teacher model and the student model respectively, S is the total number of targets, and SmoothL1 is the smooth L1 loss function.

11. The robust online vector map construction method for spatiotemporal simplified query according to claim 8, characterized in that: According to the loss function evaluation index, if the output requirements are not met, the mid-range perception mechanism is further used to optimize the student model. Through distance perception response distillation, the loss weights of different targets are dynamically adjusted, focusing on optimizing the detection accuracy of medium and short-range targets. include: The dynamic adjustment formula is: Among them, d i is the distance between target i and sensor, w i is the normalized weight of the target, μ is the fixed weight of the distant target, and A, B, C and δ are adjustment parameters.

Images