A high-precision map dynamic detection and updating method using a general image restoration strategy

Abstract

The application discloses a high-precision map dynamic detection and updating method adopting a general image recovery strategy, belongs to the technical field of high-precision map dynamic detection and updating methods, and through a pre-training process, a degraded image is input into a network, features are extracted and enhanced layer by layer by adopting a feature generation module and a feature selection module, image recovery is realized by combining multi-dimensional feature fusion and a cross attention mechanism, the detail restoration and global consistency of the degraded image are improved, according to a BEV representation generated by the pre-training model in step S1, in combination with the historical state of the high-precision map, a dynamic detection module is adopted to identify elements that are newly added, deleted or unchanged in the map, and a difference map is output, according to the difference map generated in step S2, a map updating module is adopted to perform geometric adjustment and semantic correction on the marked added and deleted areas, an accurate and coherent local high-precision map is generated, and through a model evaluation framework, the performance of dynamic detection and map updating is comprehensively evaluated.

Description

Technical Field

[0001] This invention relates to a method for dynamic detection and updating of high-precision maps, and particularly to a method for dynamic detection and updating of high-precision maps using a general image restoration strategy, belonging to the technical field of dynamic detection and updating methods for high-precision maps. Background Technology

[0002] Existing high-precision map generation methods can provide certain geometric and semantic representations in static scenes, but when dealing with dynamic scenes, insufficient feature fusion due to occlusion, dynamic targets, and viewpoint changes leads to decreased mapping accuracy. Furthermore, traditional methods often lack effective image restoration strategies when faced with degraded images (such as those with insufficient lighting, blurriness, or noise interference), resulting in insufficient input data quality and further affecting the reliability of the generated map.

[0003] On the other hand, while map building technology based on LiDAR and image fusion can provide high geometric accuracy, it consumes a lot of computational resources and is difficult to adapt to the real-time requirements of actual autonomous driving scenarios. At the same time, it lacks the ability to model the temporal consistency of multi-frame data and does not strongly express the dynamic changes of target instances in the scene, resulting in discontinuity and inconsistency in the results during the map update process. In addition, existing map building methods do not optimize the lightweighting and resource adaptability of models, making it difficult to deploy effectively in resource-constrained scenarios, further limiting the application scenarios of autonomous driving technology.

[0004] In summary, current high-precision map dynamic detection and update technologies still face many technical bottlenecks in degraded image processing, multi-view feature fusion, dynamic change detection, and lightweight design. To address these issues, a high-precision map dynamic detection and update method using a general image restoration strategy is designed. Summary of the Invention

[0005] The main objective of this invention is to provide a method for dynamic detection and updating of high-precision maps using a general image restoration strategy.

[0006] The objective of this invention can be achieved by adopting the following technical solution:

[0007] A high-precision map dynamic detection and update method employing a general image restoration strategy includes the following steps:

[0008] Step S1: Through the pre-training process, the degraded image is input into the network. The feature generation module and feature selection module are used to extract and enhance features layer by layer. The multi-dimensional feature fusion and cross-attention mechanism are combined to realize image restoration and improve the detail restoration and global consistency of the degraded image.

[0009] Step S2: Based on the BEV representation generated by the pre-trained model in Step S1, and combined with the historical state of the high-precision map, the dynamic detection module is used to identify newly added, deleted, or unchanged elements in the map, and outputs a difference map.

[0010] Step S3: Based on the difference map generated in step S2, the map update module performs geometric adjustments and semantic corrections on the newly added and deleted areas to generate an accurate and coherent local high-precision map.

[0011] Step S4: Conduct a comprehensive evaluation of the performance of dynamic detection and map updating using the model evaluation framework.

[0012] Preferably, in step S1, the pre-training process generates a feature map from the input image through an initial convolution operation based on the characteristics of the degraded image, and introduces it into the encoder stage to complete the learning of multi-level features.

[0013] Preferably, the encoder uses a cross-channel attention module to model global semantic features, while combining a spatial self-attention module to enhance the extraction of local detail features;

[0014] At each layer, the spatial resolution of the feature map is gradually reduced through a downsampling mechanism to ensure layer-by-layer learning and extraction of multi-scale features. By combining global and local information, multiple degradation characteristics in the input image are captured.

[0015] Pre-training fuses multi-scale features from the encoder layers through a skip connection mechanism and combines upsampling operations to restore image resolution;

[0016] The decoder applies the TSAB module layer by layer to enhance global information modeling and uses the SSAB module to optimize local detail recovery capabilities, ensuring the generation of high-quality feature representations.

[0017] While progressively upsampling the features to the original input resolution, the multi-scale features are jointly modeled through a concatenation operation to enhance the overall performance of the output features;

[0018] Based on the requirements of global information modeling, the Multi-Dconv Transpose Attention module is used to capture long-distance feature interactions in the channel dimension;

[0019] The overlapping cross-attention module is used to jointly learn local and global information in the spatial dimension;

[0020] During the pre-training process, the alternating effects of consistency regularization constraints TSAB and SSAB are introduced to ensure the temporal stability and structural consistency of features.

[0021] Preferably, in step S2, based on the pre-trained model, the differences between the recovered image obtained from the input and the outdated high-precision map are identified as follows:

[0022] First, the geometric and semantic information of the expired high-precision maps is encoded;

[0023] This process uses dedicated geometric and semantic encoders to transform complex map descriptions into compact, high-dimensional feature representations while preserving core information.

[0024] The geometric shape and semantic information of each map element are converted into a high-dimensional feature tensor for subsequent fusion.

[0025] The topology of expired maps is processed to ensure their spatial consistency and compatibility.

[0026] After the expired map is encoded, the sensor images are sent to the preprocessing module; the preprocessing module uses geometric correction and time synchronization technology to ensure the consistency of data from different perspectives in the spatiotemporal dimensions.

[0027] A deep learning backbone network is used to extract features from image data. Low-order features are captured by a multi-layer convolutional network, while semantic information is globally modeled through a multi-head self-attention mechanism to generate a high-dimensional feature tensor with multiple perspectives.

[0028] By using a cross-attention mechanism, the real-time features of sensor data are integrated with the features of expired high-precision maps. The model's classification head independently predicts the state of each map element and integrates these predictions into a difference map, marking all elements in the scene that have changed.

[0029] The final difference map, as output, provides detailed change information, clearly identifying whether each element is newly added, deleted, or unchanged. Dynamic detection performance is quantified using an accuracy formula:

[0030]

[0031] Where, Acc± represents the dynamic detection performance;

[0032] S represents the sequence number;

[0033] Fi is the number of frames in the i-th sequence;

[0034] c and These represent the changed and unchanged states, respectively.

[0035] and These represent the predicted elements in the difference map and the actual values ​​of the difference map elements, respectively.

[0036] Preferably, in step S3, based on the output of the difference map and sensor data, the map update module is used to correct and improve the local high-precision map to ensure that the updated map accurately reflects the current scene.

[0037] First, for elements marked as new in the difference map, the system generates a geometric description of the new area using sensor data;

[0038] Each element in the newly added region contains detailed shape information, semantic labels, and topological relationships;

[0039] The attributes of newly added elements are further optimized through the semantic analysis module to ensure their semantic consistency with the overall map.

[0040] For elements marked as deleted in the difference map, the system removes their geometric descriptions and related attributes from the map;

[0041] The deletion operation is performed using a precise positioning algorithm to avoid erroneous impact on unchanged elements. The IoU precision of element deletion is quantified using the intersection-union ratio formula:

[0042]

[0043] After completing the addition and deletion operations, the system uses the geometric correction module to spatially align all map elements, ensuring that the added and deleted areas are seamlessly integrated with the unchanged areas in a unified coordinate system.

[0044] The system uses spatial projection and coordinate transformation techniques to align the updated map with the global scene in order to eliminate potential spatial errors;

[0045] By integrating unchanged areas with newly added and updated elements, a complete local high-precision map is generated. The system uses the following formula to evaluate the average accuracy of the updated map in order to quantify its overall performance:

[0046]

[0047] in,

[0048] Where V and V′ represent the predicted and actual map elements, respectively;

[0049] Chamfer distance measures the error between boundaries;

[0050] Fréchet distance assessment of the similarity of center lines;

[0051] V left and V right V' represents the map elements on the left and right sides of the predicted map. left ,V'right V represents the map elements on the left and right sides of the real map. ctr ,V' ctr These represent the central elements of the predicted map and the central elements of the actual map, respectively.

[0052] Preferably, step S4 includes, in the final evaluation stage of the model, using multiple methods to comprehensively verify the model performance in conjunction with change data in the real scene;

[0053] First, the model's ability to detect changes in each frame is evaluated using a single-frame mode, while the model's temporal stability and consistency are analyzed using a multi-frame mode. This dual-mode evaluation method can determine the model's performance in single and continuous scenes.

[0054] Among them, the single-frame mode evaluation model can be considered to meet the requirements if it satisfies Acc+≥0.85 and Acc-≥0.90, and the multi-frame mode analysis model satisfies Acc+≥0.90 and Acc-≥0.92.

[0055] For different types of changes, type-related and type-independent indices are calculated to evaluate the model's detection performance for specific change categories and its generalization ability to overall changes. The model's localization performance for local changes is evaluated using the localization intersection-union ratio (IUU), the formula of which is:

[0056]

[0057] When Accloca ≥ 0.8, it is considered to meet the requirements;

[0058] The overall performance of the model in updating the map is further quantified by the average accuracy of each map element to ensure the reliability of the model output. When the average accuracy AP is ≥ 0.8, it is considered to meet the requirements.

[0059] Beneficial technical effects of the present invention:

[0060] This invention provides a high-precision map dynamic detection and update method employing a general image restoration strategy. Through feature generation and selection modules during pre-training, as well as multi-dimensional feature fusion and cross-attention mechanisms, it effectively handles degraded images, significantly improving detail restoration and global consistency. Regardless of whether the image is poorly lit, blurry, or subject to noise interference, the encoder and decoder, through the collaborative efforts of cross-channel attention modules and spatial self-attention modules, achieve multi-scale feature learning and fusion, generating high-quality feature representations. This provides an accurate and clear data foundation for subsequent map detection and update, greatly improving input data quality and reducing map construction errors caused by image quality issues compared to traditional methods.

[0061] By combining the BEV representation generated by the pre-trained model with the historical state of the high-precision map and the dynamic detection module, it is possible to accurately identify newly added, deleted, or unchanged elements in the map.

[0062] By encoding the geometric and semantic information of expired high-precision maps and fusing features with sensor images, this method can effectively capture real-time changes and accurately output difference maps even in complex scenarios such as multi-target interaction and dynamic occlusion. Its dynamic detection accuracy is ≥0.85 and ≥0.90 in single-frame mode, and ≥0.90 and ≥0.92 in multi-frame (MF) mode, providing a reliable basis for timely map updates and enabling high-precision maps to better adapt to dynamic environmental changes.

[0063] Based on the difference map, the map update module performs geometric adjustments and semantic corrections on newly added and deleted areas. For newly added elements, it generates detailed geometric descriptions and optimizes attributes; for deleted elements, it removes them precisely without affecting other parts. Through geometric correction and spatial projection techniques, the updated map is aligned with the global scene, resulting in a local high-precision map with an average accuracy (AP) ≥ 0.8. This ensures the accuracy and consistency of the map, providing high-precision map support for tasks such as autonomous driving path planning, and improving the reliability and safety of autonomous driving systems.

[0064] By using a model evaluation framework that combines single-frame and multi-frame modes, as well as various evaluation methods such as type-related, type-independent indicators, and location intersection-over-union ratio, the performance of the model in terms of change detection, positioning accuracy, and map updating is comprehensively evaluated. This not only provides a scientific basis for model improvement, but also ensures the reliability and generalization ability of the model in different scenarios, making high-precision maps more stable and adaptable in practical applications. Attached Figure Description

[0065] Figure 1 This is a flowchart of a preferred embodiment of a high-precision map dynamic detection and update method employing a general image restoration strategy according to the present invention. Detailed Implementation

[0066] To enable those skilled in the art to understand the technical solution of the present invention more clearly, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0067] A high-precision map dynamic detection and update method using a general image restoration strategy is characterized by the following steps: Step S1: Through a pre-training process, the degraded image is input into the network, and features are extracted and enhanced layer by layer using a feature generation module and a feature selection module. Combined with multi-dimensional feature fusion and cross-attention mechanism, image restoration is achieved, improving the detail restoration and global consistency of the degraded image.

[0068] Step S2: Based on the BEV representation generated by the pre-trained model in Step S1, and combined with the historical state of the high-precision map, the dynamic detection module is used to identify newly added, deleted, or unchanged elements in the map, and outputs a difference map.

[0069] Step S3: Based on the difference map generated in step S2, the map update module performs geometric adjustments and semantic corrections on the newly added and deleted areas to generate an accurate and coherent local high-precision map.

[0070] Step S4: Through the model evaluation framework, comprehensively evaluate the performance of dynamic detection and map updating, including change detection accuracy, positioning accuracy, and overall consistency and accuracy after map updating.

[0071] Step S1 includes generating feature maps from the input image through initial convolution operations during pre-training, based on the degraded image characteristics, and introducing these maps into the encoder stage to complete the learning of multi-level features. The encoder employs a cross-channel attention module (TSAB) to model global semantic features, while combining it with a spatial self-attention module (SSAB) to enhance the extraction of local detail features. At each layer, the spatial resolution of the feature maps is progressively reduced through a downsampling mechanism (e.g., from H×W×C to H / 2×W / 2×2C, and then expanded to 4C and 8C channel dimensions) to ensure layer-by-layer learning and extraction of multi-scale features. The goal of this stage is to capture multiple degradation characteristics in the input image by jointly using global and local information.

[0072] In the decoder stage, pre-training fuses multi-scale features from each layer of the encoder through a skip connection mechanism and combines this with upsampling operations to restore the image resolution. The decoder applies the TSAB module layer by layer to enhance global information modeling and uses the SSAB module to optimize local detail recovery capabilities, thereby ensuring the generation of high-quality feature representations. While the features are progressively upsampled to the original input resolution, the multi-scale features are jointly modeled through a concatenation operation to enhance the overall performance of the output features.

[0073] To improve the generalization ability of the pre-trained model, based on the requirement of global information modeling, a Multi-DconvTranspose Attention (MDTA) module is used to capture long-distance feature interactions in the channel dimension; based on the requirement of local detail enhancement, an Overlapping Cross-Attention (OCA) module is used to jointly learn local and global information in the spatial dimension. During pre-training, consistency regularization constraints, TSAB and SSAB, are introduced to ensure the temporal stability and structural consistency of features.

[0074] Ultimately, the pre-trained model achieves efficient learning across various degraded image tasks, generating high-quality feature representations that can adapt to different types of degraded images, providing a stable and powerful foundational model for subsequent tasks.

[0075] Step S2 includes identifying the differences between the recovered image obtained from the input and the expired high-precision map based on the pre-trained model. First, the geometric and semantic information of the expired high-precision map is encoded.

[0076] This process uses dedicated geometric and semantic encoders to transform complex map descriptions into compact, high-dimensional feature representations while preserving core information.

[0077] The geometry (such as centerline and boundary attributes) and semantic information (such as lane type) of each map element are converted into high-dimensional feature tensors for subsequent fusion. Furthermore, the topology of expired maps is processed to ensure spatial consistency and fusion compatibility.

[0078] After encoding the expired maps, the sensor images are sent to the preprocessing module. This module uses geometric correction and time synchronization techniques to ensure the consistency of data from different perspectives in the spatiotemporal dimensions.

[0079] Subsequently, a deep learning backbone network (such as ResNet) is used to extract features from the image data.

[0080] Low-order features are captured through multi-layer convolutional networks, while semantic information is globally modeled using a multi-head self-attention mechanism, generating high-dimensional feature tensors from multiple perspectives. These features represent the spatial locations of objects in the scene and their semantic relationships, laying the foundation for subsequent fusion steps.

[0081] The feature fusion stage is the core step in the dynamic detection process.

[0082] By using a cross-attention mechanism, the real-time features of sensor data are integrated with the features of outdated high-precision maps.

[0083] This approach ensures that the system retains valid information from expired maps that has not changed, while also capturing real-time changes through sensor data. The model's classification head independently predicts the state (added, deleted, or unchanged) of each map element and integrates these predictions into a difference map, marking all elements in the scene that have changed.

[0084] The final difference map, as output, provides detailed change information, clearly identifying whether each element was added, deleted, or unchanged. Dynamic detection performance is quantified using an accuracy formula:

[0085]

[0086] Where Acc± represents the dynamic detection performance;

[0087] S represents the sequence number;

[0088] Fi is the number of frames in the i-th sequence;

[0089] c and These represent the changed and unchanged states, respectively.

[0090] and These represent the predicted elements in the difference map and the actual values ​​of the difference map elements, respectively.

[0091] Step S3 involves using a map update module to correct and improve the local high-precision map based on the output of the difference map and sensor data, ensuring that the updated map accurately reflects the current scene. First, for elements marked "new" in the difference map, the system generates a geometric description of the new area using sensor data. Each element in the new area contains detailed shape information, semantic labels (such as lane type or sign location), and topological relationships. Furthermore, the semantic analysis module further optimizes the attributes of the new elements to ensure semantic consistency with the overall map.

[0092] For elements marked "deleted" in the difference map, the system removes their geometric descriptions and related attributes from the map. The deletion operation is performed using a precise localization algorithm to avoid unintended impacts on unchanged elements. The IoU precision of element deletion is quantified using the intersection-union ratio formula:

[0093] After completing the addition and deletion operations, the system uses the geometric correction module to spatially align all map elements, ensuring that the added and deleted areas are seamlessly integrated with the unchanged areas in a unified coordinate system. In addition, the system uses spatial projection and coordinate transformation technology to align the updated map with the global scene to eliminate potential spatial errors.

[0094] Finally, by integrating unchanged areas with newly added and updated elements, a complete local high-precision map is generated. The system uses the following formula to evaluate the average accuracy (AP) of the updated map to quantify its overall performance:

[0095] Where V and V′ represent the predicted and actual map elements, respectively;

[0096] Chamfer distance measures the error between boundaries;

[0097] Fréchet distance assessment of the similarity of center lines;

[0098] Vleft and V right V' represents the map elements on the left and right sides of the predicted map. left ,V' right V represents the map elements on the left and right sides of the real map. ctr ,V' ctr These represent the central elements of the predicted map and the central elements of the actual map, respectively.

[0099] Through these steps, the map update module successfully generated an accurate and complete local high-precision map, providing reliable support for subsequent tasks such as autonomous driving path planning. This update process not only dynamically adapts to environmental changes but also significantly improves the system's reliability and safety through its high-precision map generation capabilities.

[0100] Step S4 involves the final evaluation phase of the model, where multiple methods are used to comprehensively validate the model's performance, incorporating change data from real-world scenarios. First, the model's ability to detect changes in each frame is evaluated using a single-frame (SF) mode, while a multi-frame (MF) mode is used to analyze the model's temporal stability and consistency. This dual-mode evaluation approach helps determine the model's performance in both single and continuous scenarios.

[0101] The single-frame (SF) mode evaluation model can be considered to meet the requirements if it satisfies Acc+≥0.85 and Acc-≥0.90, and the multi-frame (MF) mode analysis model satisfies Acc+≥0.90 and Acc-≥0.92.

[0102] Secondly, for different types of changes, type-aware and type-agnostic metrics are calculated to evaluate the model's detection performance for specific change categories (such as additions or deletions) and its generalization ability to overall changes. The model's localization performance on local changes is evaluated using the localization intersection-union ratio (Accloca), the formula of which is:

[0103]

[0104] When Accloca ≥ 0.8, it can be considered to meet the requirements.

[0105] Furthermore, the overall performance of the model in updating the map is further quantified by the average precision (AP) of each map element to ensure the reliability of the model output. A mean precision (AP) ≥ 0.8 is considered to meet the requirements.

[0106] By combining multiple indicators, the model's change detection capability, map update accuracy, and interpretability of output results are comprehensively evaluated.

[0107] These evaluation results not only provide a scientific basis for model improvement, but also provide strong support for the application of high-precision maps in real-world scenarios.

[0108] The above description is merely a further embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope disclosed in the present invention, based on the technical solution and concept of the present invention, shall fall within the scope of protection of the present invention.

Claims

1. A method for dynamic detection and updating of high-precision maps using a general image restoration strategy, characterized in that: Includes the following steps: Step S1: Through the pre-training process, the degraded image is input into the network. The feature generation module and feature selection module are used to extract and enhance features layer by layer. The multi-dimensional feature fusion and cross-attention mechanism are combined to realize image restoration and improve the detail restoration and global consistency of the degraded image. Step S2: Based on the BEV representation generated by the pre-trained model in Step S1, and combined with the historical state of the high-precision map, the dynamic detection module is used to identify newly added, deleted, or unchanged elements in the map, and outputs a difference map. Step S3: Based on the difference map generated in step S2, the map update module performs geometric adjustments and semantic corrections on the newly added and deleted areas to generate an accurate and coherent local high-precision map. Step S4: Conduct a comprehensive evaluation of the performance of dynamic detection and map updating using the model evaluation framework; The encoder uses a cross-channel attention module to model global semantic features, while combining a spatial self-attention module to enhance the extraction of local detail features; At each layer, the spatial resolution of the feature map is gradually reduced through a downsampling mechanism to ensure layer-by-layer learning and extraction of multi-scale features. By combining global and local information, multiple degradation characteristics in the input image are captured. Pre-training fuses multi-scale features from the encoder layers through a skip connection mechanism and combines convolutional operations with sampling operations to restore image resolution; The decoder applies the TSAB module layer by layer to enhance global information modeling and uses the SSAB module to optimize local detail recovery capabilities, ensuring the generation of high-quality feature representations. While progressively upsampling the features to the original input resolution, the multi-scale features are jointly modeled through a concatenation operation to enhance the overall performance of the output features; Based on the requirements of global information modeling, the Multi-Dconv Transpose Attention module is used to capture long-distance feature interactions in the channel dimension; The overlapping cross-attention module is used to jointly learn local and global information in the spatial dimension; During the pre-training process, the alternating effects of consistency regularization constraints TSAB and SSAB are introduced to ensure the temporal stability and structural consistency of features. In step S2, based on the pre-trained model, the differences between the recovered image obtained from the input and the outdated high-precision map are identified as follows: First, the geometric and semantic information of the expired high-precision maps is encoded; This process uses dedicated geometric and semantic encoders to transform complex map descriptions into compact, high-dimensional feature representations while preserving core information. The geometric shape and semantic information of each map element are converted into a high-dimensional feature tensor for subsequent fusion. The topology of expired maps is processed to ensure their spatial consistency and compatibility. After the expired map is encoded, the sensor images are sent to the preprocessing module. The preprocessing module uses geometric correction and time synchronization techniques to ensure the consistency of data from different perspectives in the spatiotemporal dimensions. A deep learning backbone network is used to extract features from image data. Low-order features are captured by a multi-layer convolutional network, while semantic information is globally modeled through a multi-head self-attention mechanism to generate a high-dimensional feature tensor with multiple perspectives. By using a cross-attention mechanism, the real-time features of sensor data are integrated with the features of expired high-precision maps. The classification head of the preprocessing model independently predicts the state of each map element and integrates these prediction results into a difference map, marking all elements in the scene that have changed. The final difference map, as output, provides detailed change information, clearly identifying whether each element is newly added, deleted, or unchanged. Dynamic detection performance is quantified using an accuracy formula: ； Where Acc± represents the dynamic detection performance; S represents the sequence number; Fi is the number of frames in the i-th sequence; c and These represent the changed and unchanged states, respectively. and These represent the predicted elements in the difference map and the actual values ​​of the difference map elements, respectively.

2. The high-precision map dynamic detection and updating method using a general image restoration strategy according to claim 1, characterized in that: In step S1, the pre-training process generates feature maps from the input image through initial convolution operations based on the characteristics of the degraded image, and introduces them into the encoder stage to complete the learning of multi-level features.

3. The high-precision map dynamic detection and updating method using a general image restoration strategy according to claim 1, characterized in that: Step S3 uses the map update module to correct and improve the local high-precision map based on the output of the difference map and sensor data to ensure that the updated map accurately reflects the current scene. First, for elements marked as new in the difference map, the system generates a geometric description of the new area using sensor data; Each element in the newly added region contains detailed shape information, semantic labels, and topological relationships; The attributes of newly added elements are further optimized through the semantic analysis module to ensure their semantic consistency with the overall map. For elements marked as deleted in the difference map, the system removes their geometric descriptions and related attributes from the map; The deletion operation is performed using a precise positioning algorithm to avoid erroneous impact on unchanged elements. The IoU precision of element deletion is quantified using the intersection-union ratio formula: ; After completing the addition and deletion operations, the system uses the geometric correction module to spatially align all map elements, ensuring that the added and deleted areas are seamlessly integrated with the unchanged areas in a unified coordinate system. The system uses spatial projection and coordinate transformation techniques to align the updated map with the global scene in order to eliminate potential spatial errors; By integrating unchanged areas with newly added and updated elements, a complete local high-precision map is generated. The system uses the following formula to evaluate the average accuracy of the updated map in order to quantify its overall performance: ； Among them, V and These represent predicted and actual map elements, respectively; Chamfer distance measures the error between boundaries; Fréchet distance assessment of the similarity of center lines; and This represents the map elements on the left and right sides of the prediction map. These are map elements on the left and right sides of the real map. These represent the central elements of the predicted map and the central elements of the actual map, respectively.

4. The high-precision map dynamic detection and updating method using a general image restoration strategy according to claim 1, characterized in that: Step S4 includes the final evaluation stage of the model, incorporating change data from real-world scenarios and employing multiple methods to comprehensively verify the model's performance. First, the model's ability to detect changes in each frame is evaluated using a single-frame mode, while the model's temporal stability and consistency are analyzed using a multi-frame mode. This dual-mode evaluation method can determine the model's performance in single and continuous scenes. Among them, the single-frame mode evaluation model can be considered to meet the requirements if it satisfies Acc+≥0.85 and Acc-≥0.90, and the multi-frame mode analysis model satisfies Acc+≥0.90 and Acc-≥0.

92. For different types of changes, type-related and type-independent indices are calculated to evaluate the model's detection performance for specific change categories and its generalization ability to overall changes. The model's localization performance for local changes is evaluated using the localization intersection-union ratio (IUU), the formula of which is: ； When Accloca ≥ 0.8, it is considered to meet the requirements; The overall performance of the model in updating the map is further quantified by the average accuracy of each map element to ensure the reliability of the model output. When the average accuracy AP is ≥ 0.8, it is considered to meet the requirements.

Images