A large scene positioning method and system based on a double-layer scene semantic topology graph
By constructing a hierarchical semantic topology map and adopting a progressive matching strategy, the problem of poor positioning accuracy and efficiency in large-scale indoor environments is solved, achieving efficient and stable object-level positioning that adapts to complex environmental changes.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HEFEI INSTITUTE OF PHYSICAL SCIENCE CHINESE ACADEMY OF SCIENCES
- Filing Date
- 2026-03-30
- Publication Date
- 2026-07-14
AI Technical Summary
Existing semantic topological map localization methods suffer from problems such as map size expansion, high computational complexity, sensitivity to environmental changes, and matching ambiguity in large-scale indoor environments, making it difficult to achieve a balance between localization accuracy and computational efficiency.
A hierarchical semantic topology map is constructed, including an image layer and an object layer. Hierarchical relationships are established through topological edges. A progressive matching strategy is adopted, first screening candidate nodes in the image layer, and then performing precise matching in the object layer. Combined with a visual language model and an instance segmentation network, efficient object-level localization is achieved.
It significantly reduces computational complexity, improves the robustness and efficiency of positioning, can adapt to environmental changes, and enhances the stability and accuracy of positioning.
Smart Images
Figure CN121933025B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of visual environment perception and autonomous navigation technology for unmanned systems, and particularly to a large-scene localization method and system based on a two-layer scene semantic topology map. Background Technology
[0002] As the application of robots in large-scale indoor environments continues to expand, achieving stable and accurate autonomous localization in complex environments has become a fundamental problem for robot navigation. Traditional localization methods are typically based on geometric features or grid maps, but these methods are prone to map bloat and high computational complexity in large-scale indoor environments. Furthermore, these methods are highly sensitive to changes in ambient lighting or layout. To improve the robustness of localization systems, existing technologies are beginning to incorporate object-level semantic information and topological structures, constructing semantic topological maps to provide a structured representation of the scene.
[0003] Current semantic topological mapping localization methods typically rely on a single semantic level or are function-oriented. As the scene scale expands, the topological structure easily introduces a large amount of redundant data, thus increasing the computational burden of online matching. Furthermore, in indoor environments with numerous repeating objects and similar spatial structures, the lack of effective global semantic constraints makes existing object-level localization methods prone to matching ambiguities, making it difficult to balance localization accuracy and computational efficiency. Simultaneously, the unclear division of roles among different semantic scales in the localization process leads to a high degree of coupling between global search and fine-grained matching processes.
[0004] Therefore, how to achieve efficient, stable, and environmentally adaptable object-level positioning has become a pressing technical challenge. Summary of the Invention
[0005] The main objective of this invention is to provide a large-scene localization method and system based on a two-layer scene semantic topology map, aiming to achieve efficient, stable, and environmentally adaptable object-level localization.
[0006] To achieve the above objectives, this invention proposes a large-scene localization method based on a two-layer scene semantic topology map, comprising: S1, constructing a hierarchical semantic topology map, wherein the hierarchical semantic topology map includes at least an image layer and an object layer, and the image layer and the object layer establish a hierarchical association relationship through topological edges; S2, generating image layer nodes and object layer nodes based on scene image data, wherein the object layer nodes contain semantic features and appearance features of object instances; S3, establishing the hierarchical association relationship so that the image layer nodes serve as indexes for the object layer nodes; S4, receiving a query image and performing progressive matching based on the hierarchical semantic topology map: firstly, candidate image nodes are selected in the image layer, and then, according to the hierarchical association relationship, matching is performed in the object layer nodes associated with the candidate image nodes to determine the localization result.
[0007] Preferably, the image layer nodes are generated by aggregating image features from similar viewpoints or repeated observations that meet the fusion conditions; the generation process of the image layer nodes includes: aggregating image features from similar viewpoints or repeated observations that meet the fusion conditions. Features of Frame Observation Images Aggregation is performed to generate aggregated features of image layer nodes. The calculation formula is as follows:
[0008] .
[0009] Preferably, the object layer nodes are generated by parsing the scene image through an instance segmentation network, and the attributes stored in the object layer nodes include at least the pixel-level mask, category label, bounding box, and semantic-appearance feature representation of the object.
[0010] Preferably, when generating object layer nodes, if multiple object nodes are detected as multi-view observations of the same semantic entity, an object node fusion strategy is executed; the object node fusion strategy includes: extracting cropped image feature points of the first object node and the second object node and matching them; if the matching is successful, the first object node and the second object node are merged into a unified object node.
[0011] Preferably, the process of merging into a unified object node includes: averaging the semantic-appearance feature vectors of the first object node and the second object node, and using them as the feature attributes of the merged object node; for image source, pixel-level mask, and bounding box attributes, only the one with the highest quality score is retained.
[0012] Preferably, the progressive matching includes a region-level coarse screening process, which is as follows: extracting visual features of the query image using a visual language model, calculating the similarity between the visual features and predefined region category text prompts, selecting the category corresponding to the highest similarity as the region-level label of the query image, and selecting nodes with the same region-level label in the image layer as the location search range.
[0013] Preferably, the progressive matching further includes image-level fine-grained screening, the specific process of which is: visual feature encoding of candidate image nodes within the location search range, and calculation of the query image features. With candidate image node features The cosine similarity is calculated, and the top scorers are selected. Image nodes as candidate set It is represented as: ;in, This represents the similarity calculation function. Represents the features of candidate images.
[0014] Preferably, in the candidate set The object-level matching process based on this includes: extracting the appearance features and text features of the query object from the query image, and matching them with the candidate set. Features of associated object layer nodes are fused and matched to calculate the final matching score. : ;in, These are the weighting coefficients, and ; For object image feature similarity, Score for semantic consistency.
[0015] Preferably, the object image feature similarity The semantic consistency score is obtained by calculating the cosine similarity between the stored object image features and the query appearance features. The matching degree between the features of the object and the features of the query text is obtained by calculating the visual language model.
[0016] Preferably, the method further includes step S5, a location-driven map incremental maintenance step: matching the queried object with existing object layer nodes in the hierarchical semantic topology map; if the matching fails, the queried object is determined to be a new semantic entity, and an incremental update process is triggered to construct and introduce new object layer nodes.
[0017] This application also discloses a large-scene localization system based on a two-layer scene semantic topology map, comprising: a map construction module for constructing a hierarchical semantic topology map, wherein the hierarchical semantic topology map includes at least an image layer and an object layer, and the image layer and the object layer establish a hierarchical association relationship through topological edges; a node generation module for generating image layer nodes and object layer nodes based on scene image data, wherein the object layer nodes contain semantic features and appearance features of object instances; an association construction module for establishing the hierarchical association relationship so that the image layer nodes serve as indexes of the object layer nodes; and a localization matching module for receiving a query image and performing progressive matching based on the hierarchical semantic topology map: firstly, candidate image nodes are selected in the image layer, and then, according to the hierarchical association relationship, matching is performed in the object layer nodes associated with the candidate image nodes to determine the localization result.
[0018] The above technical solution has the following advantages:
[0019] This invention achieves structured organization of information at different semantic scales by constructing a hierarchical semantic topology map, effectively avoiding information redundancy caused by single-level mapping and significantly reducing overall computational complexity while ensuring object-level localization accuracy. By establishing hierarchical associations between the image layer and the object layer, image layer nodes can serve as indexes for object layer nodes, thus providing an efficient path for subsequent localization retrieval. Employing a coarse-to-fine progressive localization strategy, by filtering candidate regions at the image layer and combining hierarchical associations for fine-grained matching at the object layer, the localization search space is progressively compressed and invalid matches are reduced, significantly improving the robot's localization robustness and operational efficiency in complex, large-scale environments. Attached Figure Description
[0020] The present invention will now be described in detail with reference to specific embodiments and accompanying drawings, wherein:
[0021] Figure 1 This is a flowchart illustrating the large-scene localization method based on a two-layer scene semantic topology graph provided in an embodiment of the present invention.
[0022] Figure 2 This is a schematic diagram of the structure of a hierarchical semantic topology map provided in an embodiment of the present invention. Detailed Implementation
[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0024] Example 1
[0025] To address the challenges of robot localization in large-scale indoor scenes, such as map scaling, high computational complexity, sensitivity to environmental changes, and matching ambiguities caused by similar spatial structures, this embodiment provides a large-scene localization method based on a two-layer scene semantic topology map. This method constructs a hierarchical semantic topology map to achieve structured organization of information at different semantic scales and employs a coarse-to-fine progressive localization strategy, significantly improving the robustness and efficiency of robot localization in complex large-scale scenes.
[0026] This embodiment presents a large-scene localization method based on a two-layer scene semantic topology graph, as follows: Figure 1 As shown, it includes the following steps.
[0027] First, step S1 is executed to construct a hierarchical semantic topology map. The hierarchical semantic topology map uses a topological graph structure to model the indoor environment, organizing information at different semantic scales into different levels of topological structures. Specifically, the hierarchical semantic topology map uses a topological graph... To represent, where Represents a set of nodes. This represents the set of edges. This hierarchical semantic topology map includes at least an image layer and an object layer, as illustrated in the diagram below. Figure 2 As shown. The image layer contains multiple image layer nodes, used to represent the visual continuity of the scene; the object layer contains multiple object layer nodes, used to represent specific object instances in the scene and their spatial relationships. A hierarchical relationship is established between the image layer and the object layer through topological edges. This relationship is used during the graph construction phase to bind image observations to object entities, allowing image layer nodes to serve as indexes for object layer nodes.
[0028] Next, step S2 is executed to generate image layer nodes and object layer nodes based on the collected scene image data.
[0029] In the process of generating image layer nodes, to reduce map data redundancy and improve the stability of feature representation, this embodiment aggregates features from similar viewpoints or repeatedly observed images that meet the fusion conditions. Specifically, when the robot acquires multiple frames of observation images, the system extracts features from each frame. The system will combine the number of object instances in the image to make similarity judgments, for example, for an image... and Their image features are respectively and Image similarity The calculation formula is as follows:
[0030]
[0031] If adjacent or similar viewpoints If a frame image meets a preset fusion condition, such as an image feature cosine similarity greater than 0.9 or a corresponding camera pose distance less than 0.5m, then the fusion is performed on that frame. The features of the frame images are aggregated to generate aggregated features for the image layer nodes. The calculation formula is as follows:
[0032]
[0033] This average fusion method can effectively smooth out noise or short-term occlusion interference that may exist during single-frame image acquisition, forming a more robust image layer node representation.
[0034] In the process of generating object-layer nodes, an instance segmentation network is used to perform instance-level parsing of the scene image. The attributes stored in the object-layer nodes include at least the object's pixel-level mask, category label, bounding box, and semantic appearance feature representation. To construct the object-layer nodes, the system processes each input image frame to identify specific objects in the image, such as sofas, chairs, and tables. Pixel-level masks of these objects are extracted to accurately describe their shapes, bounding boxes are extracted to locate their positions in the image, and category labels are extracted for semantic understanding. Subsequently, the instance images cropped from the segmented regions are processed by a feature extraction network to extract object-level semantic appearance representations, which serve as the core attributes of the object-layer node.
[0035] Specifically, in this embodiment, when multiple object nodes are detected as multi-view observations of the same semantic entity, an object node fusion strategy is executed. This fusion strategy first extracts and matches the cropped image feature points of the first and second object nodes. If the feature point matching is successful and the geometric consistency constraint is satisfied, it is determined that the two nodes correspond to the same real object in the real environment, and the first and second object nodes are merged into a unified object node. During the merging process, the semantic appearance feature vectors of the first and second object nodes are averaged and used as the feature attributes of the merged object node. For attributes such as image source, pixel-level mask, and bounding box, to ensure data quality, the system uses a quality scoring mechanism, such as evaluating image sharpness and the area ratio of the object in the image, retaining only the one with the highest quality score, thereby ensuring the compactness of the object layer topology and the high accuracy of the attributes.
[0036] Then, step S3 is executed to establish hierarchical relationships. Simultaneously with object node generation, the system establishes an undirected topological edge between each object layer node and its source image layer node. This relationship indicates that the object can be observed from the image, allowing image layer nodes to effectively act as indexes for object layer nodes. In this way, during the localization and retrieval phase, a rapid coarse screening can be performed at the image layer, followed by precise object localization based on hierarchical relationships, significantly reducing the search space.
[0037] Next, step S4 is executed, where the query image is received, and progressive matching is performed based on a hierarchical semantic topology map. This progressive matching process includes region-level coarse screening, image-level fine-grained screening, and object-level precise matching.
[0038] The regional coarse screening process is as follows: A visual language model, specifically the CLIP model, is used to extract visual features from the query image. The system calculates the similarity between these visual features and predefined regional category text prompts, such as "bedroom," "living room," and "study room." The category with the highest similarity is selected as the regional label for the query image. Since nodes in the image layer are pre-labeled with corresponding regional labels during the mapping phase, the system only needs to select nodes with the same regional labels as the localization search range, thus eliminating a large number of irrelevant interference regions globally.
[0039] The image-level fine-grained screening process involves: encoding the visual features of candidate image nodes within the defined search range; and calculating the features of the query image. Features of each candidate image node The cosine similarity. To obtain the most representative candidate set, the top-scoring candidates are selected. Image nodes as candidate set It is represented as:
[0040]
[0041] in, This represents the function for calculating cosine similarity. Represents the features of candidate images.
[0042] In obtaining candidate sets Next, precise object-level matching is performed. The system extracts the appearance and text features of the query object from the query image. The appearance features are obtained by encoding the segmented object regions (i.e., instance segmentation regions) using the CLIP image encoder, while the text features are obtained by generating a semantic description of the query object using a visual language model and processing it through the CLIP text encoder. Subsequently, the appearance and text features of the query object are fused to form a unified object-level representation, thus simultaneously characterizing semantic abstract information and appearance detail features. This representation is then compared with the candidate set. Features of associated object layer nodes are fused and matched to calculate the final matching score. The specific calculation formula is as follows:
[0043] Semantic consistency score The matching degree between query text features and stored object features is obtained by calculating the matching degree through a visual language model. The formula is as follows: .
[0044] Object image feature similarity It is obtained by calculating the cosine similarity between the stored object image features and the query appearance features, and the formula is: .
[0045] Final Match Score The weighted fusion formula for the two is as follows:
[0046]
[0047] in These are the weighting coefficients, and This is used to adjust the weight of appearance similarity and semantic consistency in the final score. By combining detailed features of the object's appearance with semantic abstraction information from the text description, it can effectively solve the problem of ambiguity in recognizing similar objects in similar environments.
[0048] Finally, step S5 is executed to perform localization-driven incremental map maintenance. During the localization matching process, the system verifies the match between the queried object and existing object layer nodes in the hierarchical semantic topology map. If a valid match cannot be established, i.e., the match fails (for example, the highest matching score is lower than a set threshold), the queried object is determined to be a new semantic entity that has not yet been recorded. At this time, the system triggers an incremental update process, constructs new object layer nodes according to the established object node generation and fusion strategy, and introduces them into the map structure. This mechanism enables the semantic topology map to continuously expand and improve as the robot operates, enhancing the system's adaptability to dynamic environmental changes and the stability of long-term localization.
[0049] Example 2
[0050] This embodiment provides a large scene localization system based on a two-layer scene semantic topology graph. This system is used to implement the large scene localization method based on a two-layer scene semantic topology graph described in Embodiment 1 above.
[0051] Specifically, the large-scene positioning system in this embodiment includes a map building module, a node generation module, an association building module, a positioning matching module, and a map maintenance module for performing step S5.
[0052] The map building module is used to construct a hierarchical semantic topological map. A hierarchical semantic topological map includes at least an image layer and an object layer, with hierarchical relationships established between the image layer and the object layer through topological edges. During the map initialization phase, this module defines the basic topological skeleton of the scene, providing structural support for the subsequent filling of nodes.
[0053] The node generation module generates image-layer nodes and object-layer nodes based on scene image data. The object-layer nodes contain the semantic and appearance features of object instances. This module integrates deep neural networks, such as object detection networks and instance segmentation networks, for the digital extraction of physical entities in the environment.
[0054] The association building module is used to establish hierarchical associations so that image layer nodes serve as indexes for object layer nodes. Based on the visibility of observations, this module binds object layer nodes to the observed image layer nodes, forming cross-level topological links.
[0055] The localization and matching module receives the query image and performs progressive matching based on a hierarchical semantic topology map. This module first filters candidate image nodes in the image layer, and then, based on hierarchical relationships, matches them with object layer nodes associated with the candidate image nodes to determine the localization result.
[0056] This embodiment also provides an electronic device, including a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the large-scene localization method based on a two-layer scene semantic topology map as described in Embodiment 1. Specifically, the electronic device may be an onboard computing unit mounted on a mobile robot platform or an industrial inspection terminal.
[0057] In addition, this embodiment also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the large scene localization method based on a two-layer scene semantic topology map as described in Embodiment 1 above.
[0058] Example 3
[0059] This embodiment, based on the first embodiment described above, provides a further detailed explanation of the construction of topological relationships and hierarchical expansion of hierarchical semantic topology maps.
[0060] To construct edges between nodes in the image layer, the system calculates the cosine similarity between the corresponding image features of nodes in different image layers. If the similarity score between two nodes is higher than a preset similarity threshold, for example, a similarity greater than 0.85, an undirected topological edge is established between the two nodes. This indicates that there is a significant appearance correlation between different viewpoints or positions, thus forming a sparse and connected topological structure in the image layer to characterize the visual continuity of the scene.
[0061] Regarding the construction of edges between object layer nodes, this embodiment primarily uses the visibility association between the image layer and the object layer to indirectly express the spatial neighborhood relationship of objects. While generating object nodes, the system establishes an undirected edge between each object layer node and its source image layer node to represent the visibility relationship where the object can be observed from the image. This establishment of hierarchical consistency ensures that objects can be quickly indexed from the image during localization, improving retrieval efficiency.
[0062] Furthermore, in an optional variant implementation, the association between image layer nodes and object layer nodes is modeled using a learnable weighting mechanism. Specifically, the system introduces an attention mechanism module to learn the contribution of different objects to the semantic representation of the corresponding image layer node's region. For example, for an image layer node containing both a refrigerator and a cabinet, due to the refrigerator's higher uniqueness, the system assigns a higher weight to the topological edge between this image node and the refrigerator node, thereby enhancing cross-layer semantic constraints and improving the discriminative power of the matching.
[0063] To further compress the search space in large scenes, hierarchical semantic topology maps can introduce a region-level semantic structure. The region layer, located above the image layer, expresses the functional spatial information of the scene, such as office areas, rest areas, or warehouse areas. In the initial localization phase, a visual language model preliminarily determines the macroscopic region to which the query image belongs, allowing irrelevant region branches to be skipped directly, making the search process more efficient.
[0064] Example 4
[0065] This embodiment mainly illustrates the adaptive adjustment mechanism of the progressive positioning strategy under different scene scales and the in-depth technical details of object node fusion.
[0066] During the localization and matching process, the progressive localization strategy can adaptively adjust the filtering granularity based on the scene scale or the number of candidate nodes. For example, when the number of candidate image nodes exceeds a certain threshold, the system will automatically increase the image-level fine-grained filtering threshold and decrease the value of top-K, for example, by... The similarity constraint was adjusted from 10 to 5 to ensure that the computational cost of accurate matching at subsequent object layers remains within a controllable range. Conversely, if the initially selected candidate set is small, the system will appropriately relax the similarity constraint to prevent omissions.
[0067] For object node fusion, this embodiment executes a joint update strategy after successful matching. For pixel-level masks and bounding boxes in the attributes, the system calculates the mask integrity score for each observation, using the following formula:
[0068]
[0069] in For the mask pixel area, This represents the area of the bounding box. The system only retains... The highest-ranking data is used as the final attribute to ensure the accuracy of the object's geometric description. Semantic appearance features are achieved through a joint weighted fusion of multi-source observations, with weights determined based on the object's distance from the camera or the completeness of the observation angle. This improves the stability of the object representation and reduces the uncertainty caused by a single observation.
[0070] This location-driven map maintenance method not only enables incremental map updates but also supports the optimization of existing nodes by incorporating historical observation information. When multiple positioning points to the same object node but there are slight deviations in the reported appearance features, the system will correct the stored features of that object layer node online based on the real-time features at the time of positioning, thereby improving the structural consistency of the map under long-term operating conditions.
[0071] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. The present invention, by constructing a hierarchical semantic topology map, achieves the structured organization of information at different semantic scales, effectively avoiding the information redundancy problem caused by single-level mapping, and reducing overall computational complexity while ensuring positioning accuracy. Simultaneously, the adoption of a coarse-to-fine progressive positioning strategy can gradually compress the positioning search space, reduce invalid matches, and improve the efficiency and stability of object-level positioning in large-scale scenarios. Furthermore, the present invention introduces a map incremental maintenance strategy driven by positioning results, enabling the hierarchical semantic topology map to continuously expand with environmental changes, improving semantic coverage while maintaining structural stability, and is particularly suitable for long-term operation scenarios of unmanned systems.
Claims
1. A large-scene localization method based on a two-layer scene semantic topology graph, characterized in that, include: S1. Construct a hierarchical semantic topology map, which includes at least an image layer and an object layer, and establishes a hierarchical relationship between the image layer and the object layer through topological edges; S2. Generate image layer nodes and object layer nodes based on scene image data. The object layer nodes contain semantic features and appearance features of object instances. The object layer nodes are generated by parsing the scene image through an instance segmentation network. The attributes stored in the object layer nodes include at least the pixel-level mask, category label, bounding box, and semantic-appearance feature representation of the object. S3. Establish the hierarchical association relationship so that the image layer node serves as the index of the object layer node; while generating the object node, the system establishes an undirected topological edge between each object layer node and its source image layer node. This association relationship indicates that the object can be observed from the image, enabling the image layer node to effectively serve as the index of the object layer node; the association relationship between the image layer node and the object layer node is modeled using a learnable weight mechanism. The system introduces an attention mechanism module to learn the contribution of different objects to the semantic representation of the corresponding image layer node's region, thereby enhancing the cross-layer semantic constraint capability and improving the discriminativeness of the matching; S4. Receive the query image and perform progressive matching based on the hierarchical semantic topology map: First, filter out candidate image nodes in the image layer, and then match them in the object layer nodes associated with the candidate image nodes according to the hierarchical association relationship to determine the positioning result. The progressive matching includes a region-level coarse screening process, which is as follows: the visual features of the query image are extracted using a visual language model, the similarity between the visual features and the predefined region category text prompts is calculated, the category corresponding to the highest similarity is selected as the region-level label of the query image, and nodes with the same region-level label are selected in the image layer as the location search range. The progressive matching also includes image-level fine-grained screening, the specific process of which is: visual feature encoding of candidate image nodes within the location search range, and calculation of the query image features. With candidate image node features The cosine similarity is calculated, and the top scorers are selected. Image nodes as candidate set It is represented as: ； in, This represents the similarity calculation function. Represents candidate image features; In the candidate set The object-level matching process based on this includes: extracting the appearance features and text features of the query object from the query image, and matching them with the candidate set. Features of associated object layer nodes are fused and matched to calculate the final matching score. : ； in, These are the weighting coefficients, and ; For object image feature similarity, Score the semantic consistency. The similarity of object image features The semantic consistency score is obtained by calculating the cosine similarity between the stored object image features and the query appearance features. The matching degree between the features of the object and the features of the query text is obtained by calculating the visual language model.
2. The large-scene localization method based on a two-layer scene semantic topology graph according to claim 1, characterized in that, The image layer nodes are generated by aggregating image features from similar viewpoints or repeated observations that meet the fusion conditions. The generation process of the image layer nodes includes: for nodes that meet the fusion conditions... Features of Frame Observation Images Aggregation is performed to generate aggregated features of image layer nodes. The calculation formula is as follows: 。 3. The large-scene localization method based on a two-layer scene semantic topology graph according to claim 1, characterized in that, When generating object layer nodes, if multiple object nodes are detected as multi-view observations of the same semantic entity, an object node fusion strategy is executed. The object node fusion strategy includes: extracting cropped image feature points of the first object node and the second object node and matching them; if the matching is successful, the first object node and the second object node are merged into a unified object node.
4. The large-scene localization method based on a two-layer scene semantic topology graph according to claim 3, characterized in that, The process of merging into a unified object node includes: averaging the semantic-appearance feature vectors of the first object node and the second object node, and using them as the feature attributes of the merged object node; for image source, pixel-level mask, and bounding box attributes, only the one with the highest quality score is retained.
5. The large-scene localization method based on a two-layer scene semantic topology graph according to claim 1, characterized in that, The method further includes step S5, a location-driven map incremental maintenance step: matching the queried object with existing object layer nodes in the hierarchical semantic topology map; if the matching fails, the queried object is determined to be a new semantic entity, and an incremental update process is triggered to construct and introduce new object layer nodes.
6. A large-scene localization system based on a two-layer scene semantic topology graph, characterized in that, include: The map construction module is used to construct a hierarchical semantic topology map, which includes at least an image layer and an object layer, and the image layer and the object layer are connected by a hierarchical relationship through topological edges; the node generation module is used to generate image layer nodes and object layer nodes based on scene image data, wherein the object layer nodes contain the semantic features and appearance features of object instances. The association construction module is used to establish the hierarchical association relationship so that the image layer node serves as the index of the object layer node; the positioning and matching module is used to receive the query image and perform progressive matching based on the hierarchical semantic topology map: firstly, candidate image nodes are selected in the image layer, and then, according to the hierarchical association relationship, matching is performed in the object layer nodes associated with the candidate image nodes to determine the positioning result.