An AR guide method and system for buildings based on visual recognition

Abstract

This invention discloses a visual recognition-based AR navigation method and system for buildings, comprising: pre-collecting data on a target building area to generate a set of building observation primitives; constructing a path-reversible observation embryo based on the observation primitives and establishing a competitive constraint map based on path response differences; collecting the current frame image and inertial data and inputting them into a network, performing obfuscation separation and path association encoding to generate candidates; associating the candidates with historical primitives and embryos to generate a matching chain, and performing forward-looking proof to obtain a stable matching result; establishing spatial anchor points based on the stable matching and mapping them to the navigation, constructing a pose debt chain and performing closure correction; projecting and displaying the navigation content based on the corrected anchor points, and dynamically updating the output result according to the viewing angle and occlusion relationship. This invention achieves stable recognition and continuous display of AR navigation in complex building environments through path-conditional reversible observation embryos, competitive constraint maps, forward-looking proof, and pose closure correction.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a visual recognition-based AR navigation method and system for buildings. Background Technology

[0002] With the development of augmented reality technology and mobile terminal computing capabilities, visual recognition-based AR navigation technology for buildings is gradually being applied in scenarios such as large public buildings, hospitals, shopping malls, and exhibition halls. Existing technologies typically use cameras to capture real-time images and combine them with visual positioning and mapping techniques to overlay navigation information onto the real-world scene, thereby providing path guidance and information display. In practice, most solutions rely on image feature point matching, semantic segmentation, or simple visual anchor point recognition to identify and locate structures such as door frames, corners, and columns in the architectural environment to achieve spatial alignment of the AR content.

[0003] However, in actual building interiors, there are numerous repetitive or highly similar structural components, such as continuous door frames, similar corridors, regularly arranged columns, and symmetrical spatial layouts. These structures exhibit high consistency in appearance, making it difficult for existing recognition methods based on single-frame image features or local semantic information to effectively distinguish them. When terminal devices rely solely on the current viewpoint image for recognition, they are prone to misidentifying similar components in different locations as the same object, leading to spatial anchor point mismatches. This causes the AR navigation content to shift or even jump, severely impacting the continuity and accuracy of the navigation.

[0004] Existing technologies often employ re-identification or simple position update strategies when dealing with changes in perspective, occlusion interference, and short-term recognition failures during user movement, lacking a constraint mechanism for the continuity of the recognition process. When visual anchor points are lost for a short time or the recognition results are unstable, the system struggles to restore the original spatial relationships in a timely manner, easily leading to error accumulation and further exacerbating the problem of navigation content drift.

[0005] Therefore, how to provide a visual recognition-based AR navigation method and system for buildings is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a visual recognition-based AR navigation method and system for buildings. This invention constructs a path-condition reversible observation embryo, introduces a building observation confusion separation and path observation association coding mechanism based on an improved DDRNet, and combines short-range forward-looking counter-evidence matching and pose debt chain closure correction methods to effectively distinguish and stably identify highly similar components in complex building environments. This achieves accurate spatial anchoring and continuous display of AR navigation content, and has the advantages of strong identification uniqueness, high cross-view stability, and navigation display that is not prone to drifting or jumping.

[0007] An AR-guided building navigation method based on visual recognition according to an embodiment of the present invention includes: Pre-collection is performed on the target building area to acquire continuous image data and inertial sensor data, generating a set of building observation primitives; Based on the set of building observation primitives, a path-condition reversible observation embryo is constructed for each building observation primitive. The observation embryo competition constraint map is established based on the difference in the observation response of each path-condition reversible observation embryo under the preset movement path conditions. During the guided tour, the current frame image data and inertial sensor data of the collected terminal device are input into the improved DDRNet network to perform building observation confusion separation processing and path observation association coding processing, and generate a candidate set of building observation primitives by combining the observation primitive competition constraint map; The candidate set of building observation primitives in the current frame is associated with the building observation primitives, path-condition reversible observation embryos and observation embryo competition constraint maps of historical key frames to generate candidate matching chains. The candidate matching chains are then subjected to forward proof by contradiction to obtain stable matching results. Based on the stable matching results, a set of stable spatial anchor points is established and guided information is mapped. When the spatial anchor points become unstable, pose debt units are generated and pose debt chains are constructed. When the building observation primitive corresponding to the unstable spatial anchor point is detected, a closure correction is performed on the set of stable spatial anchor points. Based on a stable set of spatial anchor points after closure correction, the AR navigation content is spatially projected and displayed. The AR navigation content is dynamically updated according to changes in the current viewpoint and the building occlusion relationship, and a stable building AR navigation result is output.

[0008] Optionally, the continuous image data includes a sequence of multiple frames of building scene images acquired in chronological order, timestamp information corresponding to each frame image, and pixel resolution information corresponding to each frame image. The inertial sensing data includes triaxial acceleration data, triaxial angular velocity data corresponding to the timestamps of each frame image, and terminal attitude angle data calculated from the triaxial acceleration data and triaxial angular velocity data.

[0009] Optionally, the generated set of building observation primitives includes Boundary detection and geometric structure analysis are performed on continuous image data to extract boundary features corresponding to door frame boundaries, wall corner intersections, column turning edges, stair edges, sign borders, elevator door frame boundaries, and handrail end edges. Based on the boundary features, the intersection positions, connection relationships, and relative orientation relationships between each boundary are calculated to determine candidate structural units with stable spatial indication significance. The candidate structural units are then labeled with type and spatial location, forming a set of building observation primitives containing structural type identifiers, spatial coordinate information, and adjacency relationship information.

[0010] Optionally, the step of establishing a competition constraint map of observed embryos based on the differences in the observed responses of reversibly observed embryos under preset movement path conditions includes: For each building observation primitive, a local observation domain corresponding to the building observation primitive is determined. Auxiliary building observation primitives with stable spatial correlation with the building observation primitive are extracted within the local observation domain, and a local accompanying observation unit group corresponding to the building observation primitive is generated. For each local accompanying observation unit group, a set of preset movement paths corresponding to the guided walking mode is set. According to each preset movement path, the order of appearance, order of occlusion, order of contour opening and closing, and order of relative position change of the main building observation unit and each auxiliary building observation unit in continuous image data are recorded to generate a path response sequence. Based on the path response sequence, a path-condition reversible observation embryo is constructed for each building observation element. The path-condition reversible observation embryo includes the main building observation element's body identification information, the adjacency arrangement information of the local accompanying observation unit group, the display-concealment switching trajectory information under each preset movement path, the outline opening and closing trajectory information, and the position reordering information of the accompanying observation units. Competitive trigger analysis is performed on the reversible observation embryos under each path condition. Building observation primitive pairs with the same component category and overlapping features in local morphology under static view are selected. The differences in the display and concealment switching order, the speed of contour opening and closing, and the changes in the position of the accompanying observation unit are compared under each preset movement path. Competitive trigger markers for the corresponding building observation primitive pairs are generated. Based on each competition trigger marker, a competition exclusion relationship is established between building observation primitives. Each competition exclusion relationship is associated with the corresponding path condition reversible observation embryo to generate building observation primitive-level competition constraint units. All building observation primitive-level competitive constraint units are summarized, and an observation embryo competitive constraint map is established according to the correlation between the building observation primitive pairs that constitute the building observation primitive-level competitive constraint units, the corresponding preset movement paths, and the corresponding observation response differences.

[0011] Optionally, the step of generating a candidate set of building observation primitives by combining the observation embryo competition constraint map includes: An improved DDRNet network is constructed using the DDRNet backbone network as the basic feature extraction network. A contour steady-state preservation layer, an observation confusion separation layer, and a path observation association coding layer are sequentially set on the output path of the DDRNet backbone network. The current frame image data and inertial sensing data are input into the improved DDRNet network. The DDRNet backbone network performs basic feature extraction and scene semantic integration on the current frame image to generate basic feature results of the current frame building scene. The basic feature results of the current frame building scene are input into the contour steady-state preservation layer. Joint constraints are applied to the boundary continuity, boundary transition consistency and local adjacency integrity of the door frame boundary, wall corner intersection, column turning edge, stair edge, sign frame, elevator door frame boundary and handrail end edge to generate the contour preservation response corresponding to each building observation primitive. The contour-preserving response is input into the observation confusion separation layer. Contour-preserving responses with the same component category and adjacent positions are selected within a preset spatial range. Overlapping response regions, common boundary segments, and independent boundary segments are determined. Confusion attribution is performed by combining inertial sensing data and observation embryo competition constraint map. Overlapping response regions are split and common boundary segments are redistributed to generate candidate observation response clusters. The candidate observation response clusters are input into the path observation association coding layer. The path condition reversible observation embryo and the observation embryo competition constraint map are retrieved. The adjacency arrangement information, visible-hidden switching trajectory information, contour opening and closing trajectory information and the position reorder information of the local accompanying observation unit group of each candidate observation response cluster and the corresponding building observation primitive are associated and coded to form an observation coding result with path differentiation mark and competition exclusion mark. The observation coding results are merged and filtered according to the building observation primitive category, local adjacency relationship, path differentiation relationship and competition and exclusion relationship to generate a candidate set of building observation primitives.

[0012] Optionally, performing lookahead proof on the candidate matching chain to obtain a stable matching result includes: Each candidate building observation primitive in the current frame building observation primitive candidate set is associated with the historical keyframe building observation primitive, the corresponding path condition reversible observation embryo, and the corresponding local accompanying observation unit group to generate multiple candidate matching chains. For each candidate matching chain, a forward-looking observation projection sequence is constructed along a preset short-range forward-looking path, combining the movement direction, turning state, and displacement changes represented by the current frame inertial sensing data. According to the forward-looking observation projection sequence, corresponding observations are extracted from the continuously acquired verification frame images, and the actual observation results in the verification frame images are compared with the forward-looking observation projection sequence position by position to generate the counter-proof verification record for each candidate matching chain. Based on the counter-evidence verification records and observation embryo competition constraint map corresponding to each candidate matching chain, competition exclusion verification is performed on each candidate matching chain to identify candidate matching chains that call the same accompanying observation unit, occupy the same path distinction relationship, or violate the competition exclusion relationship during the forward observation process. The corresponding conflict records are written into the forward counter-evidence results of the candidate matching chains to form the forward counter-evidence judgment results corresponding to each candidate matching chain. The forward-looking proof results of all candidate matching chains are merged, and candidate matching chains with forward-looking proof conflicts are removed. The candidate matching chains that continuously satisfy the following conditions within the preset short-range forward-looking path are retained as stable matching results.

[0013] Optionally, the construction of the pose debt chain, upon detecting a building observation primitive corresponding to an unstable spatial anchor point, involves performing closure correction on the set of stable spatial anchor points, including: Based on the stable matching results, each stable matching building observation primitive is bound to its corresponding spatial location to generate a stable spatial anchor point set, and each stable spatial anchor point is associated with the corresponding path condition reversible observation embryo identifier, local accompanying observation unit group identifier, and guide information identifier. Visibility detection and observation consistency detection are performed on the set of stable spatial anchor points. When a stable spatial anchor point is not identified or the observation result is inconsistent with the path condition reversible observation embryo, the stable spatial anchor point is marked as an unstable spatial anchor point. During the period when the unstable spatial anchor point is in an unstable state, the pose change process is divided into segments based on inertial sensing data and the direction of the navigation path, generating pose debt segments arranged in chronological order, and recording the cumulative pose change and corresponding path interval identifier for each pose debt segment; For each pose debt segment, determine the front stable spatial anchor point and the rear stable spatial anchor point that are spatially continuous with the pose debt segment. Then, connect each pose debt segment sequentially according to the connection relationship between the front stable spatial anchor point, the pose debt segment, and the rear stable spatial anchor point to form a pose debt chain. When a building observation primitive corresponding to an unstable spatial anchor point is detected, the current observation pose is obtained and compared with the predicted pose at the end of the pose debt chain. Based on the pose deviation, reverse write-back correction is performed on the pose debt segment, and closure correction is performed on the set of stable spatial anchor points.

[0014] Optionally, the output of stable building AR navigation results includes: Based on the set of stable spatial anchor points after closure correction, the guide information identifier, spatial position and attitude direction corresponding to each stable spatial anchor point are read, and each stable spatial anchor point is bound to the corresponding guide information to generate an AR guide content display set. Based on the current pose of the terminal device, the spatial position and orientation of each stable spatial anchor point, spatial projection positioning processing is performed on each guide content in the AR guide content display set to determine the display position, display orientation and display range of each guide content in the current view. Based on the building boundary information and scene area information in the current frame image data, the spatial occlusion relationship between each guide content and the building observation primitive is determined. Occlusion suppression processing is performed on the guide content within the occlusion area, and visible display processing is performed on the guide content not within the occlusion area, generating the AR guide display result of the current frame. Based on the changes in the terminal device's pose, the changes in the stable spatial anchor point's position, and the current frame's AR navigation display results, the display position, orientation, and status of each navigation content are continuously updated to output stable architectural AR navigation results.

[0015] According to an embodiment of the present invention, a visual recognition-based building AR navigation system includes: The pre-acquisition modeling module is used to acquire continuous image data and inertial sensing data of the target building area and generate a set of building observation primitives; The observation embryo construction module is used to construct path-condition reversible observation embryos based on the set of building observation primitives and to establish a competitive constraint graph of observation embryos. The observation candidate generation module is used to collect the current frame image data and inertial sensor data and input them into the improved DDRNet network, and generate a candidate set of building observation primitives by combining the observation embryo competition constraint map; The matching and counter-proof module is used to generate candidate matching chains based on the candidate set of building observation primitives, path-condition reversible observation embryos, and observation embryo competition constraint graphs, and obtain stable matching results. The anchor point correction module is used to establish a set of stable spatial anchor points based on the stable matching results and map navigation information. When the spatial anchor points become unstable, it constructs a pose debt chain and performs closure correction. The navigation display module is used to project and display AR navigation content based on a corrected set of stable spatial anchor points and dynamically update the navigation results.

[0016] The beneficial effects of this invention are: This invention constructs path-condition reversible observation embryos and combines them with observation embryo competition constraint maps to model the observation responses of structures such as door frames, columns, and corners in architectural scenes under different movement path conditions. This enables similar architectural components that are difficult to distinguish in static images to have dynamic distinguishability, thereby significantly improving the uniqueness of visual recognition, avoiding the problem of misjudging similar components in different locations as the same target, and improving the accuracy and stability of recognition results in complex architectural environments.

[0017] During the recognition process, this invention introduces a building observation confusion separation and path observation association coding mechanism based on an improved DDRNet, and combines it with a short-range forward-looking reverse verification method to continuously verify candidate matching relationships across multiple frames. This makes the recognition results no longer dependent on instantaneous judgment from a single viewpoint, but rather verified and filtered through a continuous observation process, thereby improving the consistency of recognition results under cross-viewpoint conditions and enhancing the system's ability to continuously track targets during viewpoint changes, occlusion interference, and movement.

[0018] During the navigation display phase, this invention constructs a pose debt chain and performs closure correction to uniformly manage and rewrite the pose changes that occur during spatial anchor point instability. This enables the system to maintain navigation continuity even when anchor points fail temporarily or recognition is interrupted, and to quickly restore spatial consistency after re-recognition. This effectively suppresses the drift and jump of AR navigation content, improving the stability of navigation display and user experience. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a visual recognition-based AR navigation method for buildings proposed in this invention; Figure 2 This is a schematic diagram of the structure of a visual recognition-based AR navigation system for buildings proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1 A visual recognition-based AR navigation method for buildings includes: Pre-collection is performed on the target building area to acquire continuous image data and inertial sensor data, generating a set of building observation primitives; Based on the set of building observation primitives, a path-condition reversible observation embryo is constructed for each building observation primitive. The observation embryo competition constraint map is established based on the difference in the observation response of each path-condition reversible observation embryo under the preset movement path conditions. During the guided tour, the current frame image data and inertial sensor data of the collected terminal device are input into the improved DDRNet network to perform building observation confusion separation processing and path observation association coding processing, and generate a candidate set of building observation primitives by combining the observation primitive competition constraint map; The candidate set of building observation primitives in the current frame is associated with the building observation primitives, path-condition reversible observation embryos and observation embryo competition constraint maps of historical key frames to generate candidate matching chains. The candidate matching chains are then subjected to forward proof by contradiction to obtain stable matching results. Based on the stable matching results, a set of stable spatial anchor points is established and guided information is mapped. When the spatial anchor points become unstable, pose debt units are generated and pose debt chains are constructed. When the building observation primitive corresponding to the unstable spatial anchor point is detected, a closure correction is performed on the set of stable spatial anchor points. Based on a stable set of spatial anchor points after closure correction, the AR navigation content is spatially projected and displayed. The AR navigation content is dynamically updated according to changes in the current viewpoint and the building occlusion relationship, and a stable building AR navigation result is output.

[0022] In this embodiment, the continuous image data includes a sequence of multiple frames of building scene images acquired in chronological order, timestamp information corresponding to each frame image, and pixel resolution information corresponding to each frame image. The inertial sensing data includes triaxial acceleration data, triaxial angular velocity data corresponding to the timestamps of each frame image, and terminal attitude angle data calculated from the triaxial acceleration data and triaxial angular velocity data.

[0023] In this embodiment, the generation of the building observation primitive set includes Boundary detection and geometric structure analysis are performed on continuous image data to extract boundary features corresponding to door frame boundaries, wall corner intersections, column turning edges, stair edges, sign borders, elevator door frame boundaries, and handrail end edges. Based on the boundary features, the intersection positions, connection relationships, and relative orientation relationships between each boundary are calculated to determine candidate structural units with stable spatial indication significance. The candidate structural units are then labeled with type and spatial location, forming a set of building observation primitives containing structural type identifiers, spatial coordinate information, and adjacency relationship information.

[0024] In this embodiment, the step of establishing a competition constraint map of observed embryos based on the differences in the observed responses of reversibly observed embryos under preset movement path conditions includes: For each building observation primitive, a local observation domain corresponding to the building observation primitive is determined. Auxiliary building observation primitives with stable spatial correlation with the building observation primitive are extracted within the local observation domain, and a local accompanying observation unit group corresponding to the building observation primitive is generated. For each local accompanying observation unit group, a set of preset movement paths corresponding to the guided walking mode is set. The order of appearance, occlusion, outline opening and closing, and relative position changes of the main building observation unit and each auxiliary building observation unit in the continuous image data are recorded according to each preset movement path, and a path response sequence is generated. The preset movement path set includes lateral translation path, longitudinal approach path, longitudinal departure path, and turning swing path. Based on the path response sequence, a path-conditional reversible observation embryo is constructed for each building observation element. This path-conditional reversible observation embryo includes the main building observation element's body identification information, the adjacency arrangement information of local accompanying observation unit groups, the visible / hidden switching trajectory information under each preset movement path, the contour opening / closing trajectory information, and the positional reordering information of accompanying observation units. Specifically, the construction of the path-conditional reversible observation embryo for each building observation element involves: Using a single building observation primitive as the main building observation primitive, the path response sequence corresponding to the main building observation primitive is read, and auxiliary building observation primitives that maintain stable association under each preset movement path condition are extracted. The auxiliary building observation primitives are arranged according to the spatial adjacency order and directional distribution order relative to the main building observation primitives to generate the local accompanying observation unit group corresponding to the main building observation primitive. Based on the path response sequence, the main building observation element is recorded at the initial exposure position, initial occlusion position, re-exposure position, and outline opening and closing change process under each preset movement path. The order of exposure, order of occlusion, and relative position change process of each auxiliary building observation element in the local accompanying observation unit group are recorded simultaneously. The corresponding main building observation element's exposure and concealment switching trajectory information, outline opening and closing trajectory information, and accompanying observation unit position change information are generated. The main building observation element's body identification information, the adjacency arrangement information of the local accompanying observation unit group, the display and concealment switching trajectory information under each preset movement path, the outline opening and closing trajectory information, and the accompanying observation unit position reordering information are associated and encapsulated to generate the path condition reversible observation embryo of the corresponding main building observation element. Competitive trigger analysis is performed on the reversible observation embryos under each path condition. Building observation primitive pairs with the same component category and overlapping features in local morphology under static view are selected. The differences in the display and concealment switching order, the speed of contour opening and closing, and the changes in the position of the accompanying observation unit are compared under each preset movement path. Competitive trigger markers for the corresponding building observation primitive pairs are generated. Based on each competition trigger marker, a competition exclusion relationship is established between building observation primitives. Each competition exclusion relationship is associated with the corresponding path condition reversible observation embryo, generating building observation primitive-level competition constraint units. Specifically, the competition exclusion relationship between building observation primitives is established as follows: For building observation primitive pairs with competition trigger markers, read the corresponding building observation primitives’ display switching order, outline opening and closing change order and accompanying observation unit position reordering order under each preset movement path, and determine the observation response items with order differences as path distinguishing items. Based on the distinguishing items of each path, the conditions for confusion of building observation primitive pairs under the same static perspective and the conditions for the removal of distinction under the preset moving path are determined. Building observation primitive pairs that cannot be simultaneously attributed to the same observation result under the condition of confusion are identified as building observation primitive pairs with competitive and exclusive relationship. For building observation primitive pairs with competitive and exclusive relationships, record the corresponding building observation primitive identifier, the corresponding preset movement path identifier, and the corresponding path differentiation item identifier. Associate the recorded content with the corresponding path condition reversible observation embryo to form a competitive and exclusive relationship between building observation primitives. All building observation primitive-level competitive constraint units are summarized, and an observation embryo competitive constraint map is established according to the correlation between the building observation primitive pairs that constitute the building observation primitive-level competitive constraint units, the corresponding preset movement paths, and the corresponding observation response differences.

[0025] In this embodiment, the step of generating a candidate set of building observation primitives by combining the observation embryo competition constraint map includes: An improved DDRNet network is constructed using the DDRNet backbone network as the basic feature extraction network. A contour steady-state preservation layer, an observation confusion separation layer, and a path observation association coding layer are sequentially set on the output path of the DDRNet backbone network. The current frame image data and inertial sensing data are input into the improved DDRNet network. The DDRNet backbone network performs basic feature extraction and scene semantic integration on the current frame image to generate basic feature results of the current frame building scene. Specifically, the generation of basic feature results of the current frame building scene is as follows: The current frame image data is input into the DDRNet backbone network, and hierarchical convolution is performed on the door frame boundary region, wall corner intersection region, column turning edge region, stair edge region, sign border region, elevator door frame boundary region and handrail end edge region in the current frame image to extract local feature information. The current motion direction information and current attitude change information corresponding to the inertial sensing data are time-aligned with the local feature information to establish a correspondence between the building boundary observation state in the current frame image and the current terminal motion state. Scene semantic integration is performed on the local feature information after time alignment to generate the basic feature result of the current frame building scene, which simultaneously includes building boundary features, regional scene features and terminal motion state features. The basic feature results of the current frame's building scene are input into the contour steady-state preservation layer. Joint constraints are applied to the boundary continuity, boundary transition consistency, and local adjacency integrity of the door frame boundary, wall corner intersection, column turning edge, stair edge, sign frame border, elevator door frame boundary, and handrail end edge. This generates the contour preservation response for each building observation primitive. Specifically, the generation of the contour preservation response for each building observation primitive is as follows: Extract the boundary features corresponding to the door frame boundary, wall corner intersection, column turning edge, stair edge, sign frame, elevator door frame boundary and handrail end edge from the basic feature results of the current frame building scene, and determine the continuous distribution range and boundary turning position of each boundary feature in the current frame; Based on the continuity of each boundary feature within the continuous distribution range, the transition and connection at the boundary turning point, and the adjacency maintenance with adjacent building boundaries, the broken boundary segments, misaligned turning segments, and missing adjacent segments are filled with constraints to generate a continuous boundary profile. The generated continuous boundary contours are associated with the spatial location and component category of the corresponding building observation primitives to form a contour-keeping response for each building observation primitive; The contour-preserving response is input into the observation confusion separation layer. Within a preset spatial range, contour-preserving responses of the same component category and adjacent positions are selected to determine overlapping response regions, common boundary segments, and independent boundary segments. Confusion attribution is then performed using inertial sensor data and the observation embryo competition constraint map. Overlapping response regions are split, and common boundary segments are redistributed to generate candidate observation response clusters. Specifically, the confusion attribution determination using inertial sensor data and the observation embryo competition constraint map involves: Read the current motion direction, turning direction and displacement trend corresponding to the inertial sensing data, and determine the perspective change state of each contour maintaining response under the current frame observation conditions; Retrieve the path condition reversible observation embryo and the observation embryo competition constraint map corresponding to the contour preservation response, and determine the path differentiation relationship, competition exclusion relationship and accompanying observation constraint relationship of contour preservation responses with the same component category and adjacent positions under the current viewpoint change state. Based on path discrimination relationship, competitive exclusion relationship and accompanying observation constraint relationship, the attribution relationship between each contour-preserving response and the overlapping response region is determined one by one. The contour-preserving response that satisfies the corresponding path discrimination relationship under the current viewpoint change state is determined as the priority attribution object of the overlapping response region. The contour-preserving response that does not satisfy the competitive exclusion relationship is removed from the attribution candidate of the overlapping response region, and the corresponding confused attribution determination result is generated. The candidate observation response clusters are input into the path observation association coding layer. The path condition reversible observation embryo and the observation embryo competition constraint map are retrieved. The adjacency arrangement information, visible-hidden switching trajectory information, contour opening and closing trajectory information and the position reorder information of the local accompanying observation unit group of each candidate observation response cluster and the corresponding building observation primitive are associated and coded to form an observation coding result with path differentiation mark and competition exclusion mark. The observation coding results are merged and filtered according to the building observation primitive category, local adjacency relationship, path differentiation relationship and competition and exclusion relationship to generate a candidate set of building observation primitives.

[0026] In this embodiment, performing forward proof by contradiction on the candidate matching chain to obtain a stable matching result includes: Each candidate building observation primitive in the current frame building observation primitive candidate set is associated with the historical keyframe building observation primitive, the corresponding path condition reversible observation embryo, and the corresponding local accompanying observation unit group to generate multiple candidate matching chains. Each candidate matching chain includes the current frame building observation primitive, the historical keyframe building observation primitive, the corresponding path condition reversible observation embryo, and the corresponding local accompanying observation unit group. For each candidate matching chain, the forward-looking observation projection sequence is constructed along the preset short-range forward-looking path by combining the movement direction, turning state and displacement change represented by the current frame inertial sensing data. The forward-looking observation projection sequence is used to record the switching order of the main building observation elements, the opening and closing order of the outline, the exposure order of the accompanying observation units and the position reordering order of the accompanying observation units in the continuous forward-looking positions of the candidate matching chain. According to the forward-looking observation projection sequence, corresponding observations are extracted from the continuously acquired verification frame images, and the actual observation results in the verification frame images are compared with the forward-looking observation projection sequence position by position to generate the counter-verification records corresponding to each candidate matching chain. The counter-verification records include conflict records of the order of appearance and concealment, conflict records of the opening and closing of the contour, conflict records of the appearance of the accompanying observation, and conflict records of the position reordering. Based on the counter-evidence verification records and observation embryo competition constraint maps corresponding to each candidate matching chain, a competition exclusion check is performed on each candidate matching chain. Candidate matching chains that call the same accompanying observation unit, occupy the same path distinction relationship, or violate the competition exclusion relationship during forward-looking observation are identified. The corresponding conflict records are written into the forward-looking counter-evidence results of the candidate matching chains, forming the forward-looking counter-evidence judgment results for each candidate matching chain. Specifically, the competition exclusion check for each candidate matching chain is performed as follows: Retrieve the local accompanying observation unit group identifier, preset movement path identifier, and path condition reversible observation embryo identifier corresponding to each candidate matching chain, and read the counter-proof verification record corresponding to each candidate matching chain. Based on the observation embryo competition constraint map, determine whether there are overlapping relationships of accompanying observation units, path differentiation relationship conflicts, and competitive exclusion relationship conflicts among each candidate matching chain; For candidate matching chains with overlapping accompanying observation units, compare the exposure order, occlusion order and position reordering results of the corresponding accompanying observation units during the forward observation process, and determine the candidate matching chains that do not meet the constraints of the reversible observation embryo of the corresponding path conditions as accompanying observation conflict chains. For candidate matching chains with path distinction relationship conflicts or competitive exclusion relationship conflicts, compare the results of the visible / hidden switching, contour opening / closing, and accompanying observation unit position reordering of the corresponding candidate matching chains during the forward observation process. Candidate matching chains that are inconsistent with the competitive constraint map of the observed embryo are identified as competitive exclusion conflict chains. Write the conflict type, conflict location and conflict-related object corresponding to the accompanying observation conflict chain and the competitive exclusion conflict chain into the forward proof result of the corresponding candidate matching chain to form the competitive exclusion verification result of each candidate matching chain. The forward-looking proof results of all candidate matching chains are merged, and candidate matching chains with forward-looking proof conflicts are removed. The candidate matching chains that continuously satisfy the following conditions within the preset short-range forward-looking path are retained as stable matching results.

[0027] In this embodiment, the construction of the pose debt chain, upon detecting a building observation primitive corresponding to an unstable spatial anchor point, involves performing closure correction on the set of stable spatial anchor points, including: Based on the stable matching results, each stable matching building observation primitive is bound to its corresponding spatial location to generate a stable spatial anchor point set, and each stable spatial anchor point is associated with the corresponding path condition reversible observation embryo identifier, local accompanying observation unit group identifier, and guide information identifier. Visibility detection and observation consistency detection are performed on the set of stable spatial anchor points. When a stable spatial anchor point is not identified or the observation result is inconsistent with the path condition reversible observation embryo, the stable spatial anchor point is marked as an unstable spatial anchor point. Specifically, the visibility detection and observation consistency detection for the set of stable spatial anchor points are performed as follows: Read the candidate set of building observation primitives in the current frame image and the path condition reversible observation embryos corresponding to each stable spatial anchor point, and determine the exposed area and the occluded area of ​​each stable spatial anchor point under the current frame view conditions. The actual observation results in the current frame image are compared with the areas to be exposed corresponding to each stable spatial anchor point to determine whether the corresponding building observation primitives are identified in the current frame, and the visibility detection results of each stable spatial anchor point are generated. For the identified building observation primitives, further compare their visibility switching state, outline opening and closing state, and accompanying observation unit position reordering state with the constraint information in the corresponding path condition reversible observation embryo to generate observation consistency detection results for each stable spatial anchor point. The visibility detection results and the observation consistency detection results are jointly judged, and stable spatial anchor points that are not identified or fail the observation consistency detection are identified as unstable spatial anchor points. During the period when the unstable spatial anchor point is in an unstable state, the pose change process is segmented based on inertial sensor data and the navigation path direction, generating pose debt segments arranged in chronological order. The cumulative pose change and corresponding path interval identifier are recorded for each pose debt segment. Specifically, the segmentation of the pose change process based on inertial sensor data and the navigation path direction is as follows: Read the continuous inertial sensing data and corresponding navigation path direction information after the unstable spatial anchor point enters the unstable state, and determine the displacement change direction, turning change state and travel path change state of the terminal device during the instability period; Following the principles of maintaining consistent displacement direction, consistent turning state, and continuous navigation path direction, the continuous pose change process during instability is segmented into multiple pose change segments that are connected end to end. For each pose change segment, record the start time, end time, cumulative pose change within the segment, and corresponding navigation path interval identifier. Arrange the pose change segments in chronological order to generate pose debt segments arranged in chronological order. For each pose debt segment, determine the front stable spatial anchor point and the rear stable spatial anchor point that are spatially continuous with the pose debt segment. Then, connect each pose debt segment sequentially according to the connection relationship between the front stable spatial anchor point, the pose debt segment, and the rear stable spatial anchor point to form a pose debt chain. When a building observation primitive corresponding to an unstable spatial anchor point is detected, the current observation pose is obtained and compared with the predicted pose at the end of the pose debt chain. Based on the pose deviation, reverse write-back correction is performed on the pose debt segment, and closure correction is performed on the set of stable spatial anchor points.

[0028] In this embodiment, the output of stable building AR navigation results includes: Based on the set of stable spatial anchor points after closure correction, the guide information identifier, spatial position and attitude direction corresponding to each stable spatial anchor point are read, and each stable spatial anchor point is bound to the corresponding guide information to generate an AR guide content display set. Based on the current pose of the terminal device, the spatial position and orientation of each stable spatial anchor point, spatial projection positioning processing is performed on each guide content in the AR guide content display set to determine the display position, display orientation and display range of each guide content in the current view. Based on the building boundary information and scene area information in the current frame image data, the spatial occlusion relationship between each guide content and the building observation primitive is determined. Occlusion suppression processing is performed on the guide content within the occlusion area, and visible display processing is performed on the guide content not within the occlusion area, generating the AR guide display result of the current frame. Based on the changes in the terminal device's pose, the changes in the stable spatial anchor point's position, and the current frame's AR navigation display results, the display position, orientation, and status of each navigation content are continuously updated to output stable architectural AR navigation results.

[0029] refer to Figure 2 A visual recognition-based AR navigation system for buildings includes: The pre-acquisition modeling module is used to acquire continuous image data and inertial sensing data of the target building area and generate a set of building observation primitives; The observation embryo construction module is used to construct path-condition reversible observation embryos based on the set of building observation primitives and to establish a competitive constraint graph of observation embryos. The observation candidate generation module is used to collect the current frame image data and inertial sensor data and input them into the improved DDRNet network, and generate a candidate set of building observation primitives by combining the observation embryo competition constraint map; The matching and counter-proof module is used to generate candidate matching chains based on the candidate set of building observation primitives, path-condition reversible observation embryos, and observation embryo competition constraint graphs, and obtain stable matching results. The anchor point correction module is used to establish a set of stable spatial anchor points based on the stable matching results and map navigation information. When the spatial anchor points become unstable, it constructs a pose debt chain and performs closure correction. The navigation display module is used to project and display AR navigation content based on a corrected set of stable spatial anchor points and dynamically update the navigation results.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to a theme park featuring Manchu-style architecture. The park covers approximately 38,000 square meters and includes areas such as Manchu-style residences, an ice sculpture exhibition hall, snow-covered streets, and a cultural exhibition corridor. The park's architecture employs numerous repetitive wooden door frame designs, continuous columns, symmetrical window lattice structures, and uniformly styled ice sculpture entrance components. Especially in winter, when the buildings are covered in snow and ice, the surface texture becomes monotonous, reducing visual contrast. This causes traditional visual recognition-based AR navigation systems to easily encounter problems such as door frame confusion, column misjudgment, and path marker drift during the recognition process, severely impacting the visitor experience.

[0031] In this embodiment, an AR navigation system is deployed in the park. First, before the park opens, pre-data collection is performed on the simulated Manchu architectural complex, lasting approximately 2.5 hours and covering the main street, the entrance to the ice sculpture exhibition hall, and the corridor area, obtaining approximately 110,000 frames of continuous image data. Inertial sensor data is also collected simultaneously. Based on the collected data, a set of architectural observation primitives is generated, focusing on extracting structural information characteristic of Manchu architecture, such as door frame outlines, window frame boundaries, column edges, and eaves transitions. For each architectural observation primitive, a path-condition reversible observation embryo is constructed. By simulating visitors walking, turning, and closely observing in the snow, the visibility and outline changes of each observation primitive under different movement paths are recorded, and a competitive constraint map of the observation embryos is established to effectively distinguish similar-looking ice and snow components.

[0032] During system operation, visitors activate the AR navigation function after entering the park via their mobile devices. The system collects real-time image data and inertial sensor data of the current frame and inputs it into an improved DDRNet network to perform architectural observation confusion separation processing. In the area of ​​continuous door frames in Manchu style, the system can distinguish multiple door frames with the same appearance and, combined with path observation association coding, match the current observation results with a pre-established observation embryo competition constraint map.

[0033] As visitors walk along the snow trail, the system verifies continuous observation results through forward-looking verification to ensure stable identification. When visitors enter the ice sculpture exhibition hall or pass through areas with dense columns, if the anchor point becomes temporarily unstable due to obstruction or changes in lighting, the system automatically constructs a pose debt chain and performs closure correction after re-identifying the corresponding architectural observation primitive, ensuring that the guide signs quickly return to the correct position.

[0034] In actual operation testing, a four-day test was conducted in the park during the peak season for ice and snow tourism, collecting approximately 3,600 pieces of visitor guidance data. The test time covered the period from 9:00 AM to 4:00 PM, with the ambient temperature between -18°C and -10°C.

[0035] Table 1. AR navigation performance test data in a simulated Manchu-style ice and snow architecture scene.

[0036] As shown in Table 1, the traditional method exhibits a high false recognition rate across all test areas, with an overall average of 9.06%, particularly noticeable in doorway blocks and areas with dense colonnades. The method of this invention stably controls the false recognition rate between 3.0% and 3.4%, with an average of 3.20%. This indicates that the introduction of path-conditional reversible observation embryos and the observation embryo competition constraint map improves the ability to distinguish repetitive building components and reduces the problem of identification confusion.

[0037] Regarding continuous recognition performance, traditional methods achieve a success rate of approximately 86.7%, but recognition is prone to interruption in corner and occluded scenes. The method of this invention improves this metric to 94.7%, exhibiting high consistency across all regions. This demonstrates that by employing obfuscation separation processing and a forward-looking counter-evidence mechanism, effective constraints can be imposed on multi-frame observation results, enhancing recognition stability under cross-viewpoint conditions.

[0038] Regarding navigation stability, traditional methods have an average recovery time of approximately 2.1 seconds and an average drift error of approximately 0.72 meters, which easily leads to navigation deviation. The method of this invention, while maintaining a similar recovery time, controls the drift error to a lower level, resulting in a more stable navigation process. This demonstrates that the pose debt chain and closure correction mechanism can effectively suppress error accumulation and improve the overall reliability of AR navigation.

[0039] The above description is only a preferred 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 of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A visual recognition-based AR navigation method for buildings, characterized in that, include: Pre-collection is performed on the target building area to acquire continuous image data and inertial sensor data, generating a set of building observation primitives; Based on the set of building observation primitives, a path-condition reversible observation embryo is constructed for each building observation primitive. The observation embryo competition constraint map is established based on the difference in the observation response of each path-condition reversible observation embryo under the preset movement path conditions. During the guided tour, the current frame image data and inertial sensor data of the collected terminal device are input into the improved DDRNet network to perform building observation confusion separation processing and path observation association coding processing, and generate a candidate set of building observation primitives by combining the observation primitive competition constraint map; The candidate set of building observation primitives in the current frame is associated with the building observation primitives, path-condition reversible observation embryos and observation embryo competition constraint maps of historical key frames to generate candidate matching chains. The candidate matching chains are then subjected to forward proof by contradiction to obtain stable matching results. Based on the stable matching results, a set of stable spatial anchor points is established and guided information is mapped. When the spatial anchor points become unstable, pose debt units are generated and pose debt chains are constructed. When the building observation primitive corresponding to the unstable spatial anchor point is detected, a closure correction is performed on the set of stable spatial anchor points. Based on a stable set of spatial anchor points after closure correction, the AR navigation content is spatially projected and displayed. The AR navigation content is dynamically updated according to changes in the current viewpoint and the building occlusion relationship, and a stable building AR navigation result is output.

2. The AR navigation method for buildings based on visual recognition according to claim 1, characterized in that, The continuous image data includes a sequence of multiple frames of building scenes acquired in chronological order, timestamp information corresponding to each frame, and pixel resolution information corresponding to each frame. The inertial sensing data includes triaxial acceleration data, triaxial angular velocity data, and terminal attitude angle data calculated from the triaxial acceleration data and triaxial angular velocity data corresponding to the timestamps of each frame.

3. The AR navigation method for buildings based on visual recognition according to claim 1, characterized in that, The generated set of building observation primitives includes Boundary detection and geometric structure analysis are performed on continuous image data to extract boundary features corresponding to door frame boundaries, wall corner intersections, column turning edges, stair edges, sign borders, elevator door frame boundaries, and handrail end edges. Based on the boundary features, the intersection positions, connection relationships, and relative orientation relationships between each boundary are calculated to determine candidate structural units with stable spatial indication significance. The candidate structural units are then labeled with type and spatial location, forming a set of building observation primitives containing structural type identifiers, spatial coordinate information, and adjacency relationship information.

4. The AR navigation method for buildings based on visual recognition according to claim 1, characterized in that, The process of establishing a competition constraint map of observed embryos based on the differences in observed responses of reversibly observed embryos under preset movement path conditions includes: For each building observation primitive, a local observation domain corresponding to the building observation primitive is determined. Auxiliary building observation primitives with stable spatial correlation with the building observation primitive are extracted within the local observation domain, and a local accompanying observation unit group corresponding to the building observation primitive is generated. For each local accompanying observation unit group, a set of preset movement paths corresponding to the guided walking mode is set. According to each preset movement path, the order of appearance, order of occlusion, order of contour opening and closing, and order of relative position change of the main building observation unit and each auxiliary building observation unit in continuous image data are recorded to generate a path response sequence. Based on the path response sequence, a path-condition reversible observation embryo is constructed for each building observation element. The path-condition reversible observation embryo includes the main building observation element's body identification information, the adjacency arrangement information of the local accompanying observation unit group, the display-concealment switching trajectory information under each preset movement path, the outline opening and closing trajectory information, and the position reordering information of the accompanying observation units. Competitive trigger analysis is performed on the reversible observation embryos under each path condition. Building observation primitive pairs with the same component category and overlapping features in local morphology under static view are selected. The differences in the display and concealment switching order, the speed of contour opening and closing, and the changes in the position of the accompanying observation unit are compared under each preset movement path. Competitive trigger markers for the corresponding building observation primitive pairs are generated. Based on each competition trigger marker, a competition exclusion relationship is established between building observation primitives. Each competition exclusion relationship is associated with the corresponding path condition reversible observation embryo to generate building observation primitive-level competition constraint units. All building observation primitive-level competitive constraint units are summarized, and an observation embryo competitive constraint map is established according to the correlation between the building observation primitive pairs that constitute the building observation primitive-level competitive constraint units, the corresponding preset movement paths, and the corresponding observation response differences.

5. The AR navigation method for buildings based on visual recognition according to claim 1, characterized in that, The process of generating a candidate set of building observation primitives by combining the competitive constraint map of observation embryos includes: An improved DDRNet network is constructed using the DDRNet backbone network as the basic feature extraction network. A contour steady-state preservation layer, an observation confusion separation layer, and a path observation association coding layer are sequentially set on the output path of the DDRNet backbone network. The current frame image data and inertial sensing data are input into the improved DDRNet network. The DDRNet backbone network performs basic feature extraction and scene semantic integration on the current frame image to generate basic feature results of the current frame building scene. The basic feature results of the current frame building scene are input into the contour steady-state preservation layer. Joint constraints are applied to the boundary continuity, boundary transition consistency and local adjacency integrity of the door frame boundary, wall corner intersection, column turning edge, stair edge, sign frame, elevator door frame boundary and handrail end edge to generate the contour preservation response corresponding to each building observation primitive. The contour-preserving response is input into the observation confusion separation layer. Contour-preserving responses with the same component category and adjacent positions are selected within a preset spatial range. Overlapping response regions, common boundary segments, and independent boundary segments are determined. Confusion attribution is performed by combining inertial sensing data and observation embryo competition constraint map. Overlapping response regions are split and common boundary segments are redistributed to generate candidate observation response clusters. The candidate observation response clusters are input into the path observation association coding layer. The path condition reversible observation embryo and the observation embryo competition constraint map are retrieved. The adjacency arrangement information, visible-hidden switching trajectory information, contour opening and closing trajectory information and the position reorder information of the local accompanying observation unit group of each candidate observation response cluster and the corresponding building observation primitive are associated and coded to form an observation coding result with path differentiation mark and competition exclusion mark. The observation coding results are merged and filtered according to the building observation primitive category, local adjacency relationship, path differentiation relationship and competition and exclusion relationship to generate a candidate set of building observation primitives.

6. The AR navigation method for buildings based on visual recognition according to claim 1, characterized in that, The step of performing forward proof by contradiction on the candidate matching chain to obtain a stable matching result includes: Each candidate building observation primitive in the current frame building observation primitive candidate set is associated with the historical keyframe building observation primitive, the corresponding path condition reversible observation embryo, and the corresponding local accompanying observation unit group to generate multiple candidate matching chains. For each candidate matching chain, a forward-looking observation projection sequence is constructed along a preset short-range forward-looking path, combining the movement direction, turning state, and displacement changes represented by the current frame inertial sensing data. According to the forward-looking observation projection sequence, corresponding observations are extracted from the continuously acquired verification frame images, and the actual observation results in the verification frame images are compared with the forward-looking observation projection sequence position by position to generate the counter-proof verification record for each candidate matching chain. Based on the counter-evidence verification records and observation embryo competition constraint map corresponding to each candidate matching chain, competition exclusion verification is performed on each candidate matching chain to identify candidate matching chains that call the same accompanying observation unit, occupy the same path distinction relationship, or violate the competition exclusion relationship during the forward observation process. The corresponding conflict records are written into the forward counter-evidence results of the candidate matching chains to form the forward counter-evidence judgment results corresponding to each candidate matching chain. The forward-looking proof results of all candidate matching chains are merged, and candidate matching chains with forward-looking proof conflicts are removed. The candidate matching chains that continuously satisfy the following conditions within the preset short-range forward-looking path are retained as stable matching results.

7. The AR navigation method for buildings based on visual recognition according to claim 1, characterized in that, The construction of the pose debt chain involves performing closure correction on the set of stable spatial anchor points when a building observation primitive corresponding to an unstable spatial anchor point is detected, including: Based on the stable matching results, each stable matching building observation primitive is bound to its corresponding spatial location to generate a stable spatial anchor point set, and each stable spatial anchor point is associated with the corresponding path condition reversible observation embryo identifier, local accompanying observation unit group identifier, and guide information identifier. Visibility detection and observation consistency detection are performed on the set of stable spatial anchor points. When a stable spatial anchor point is not identified or the observation result is inconsistent with the path condition reversible observation embryo, the stable spatial anchor point is marked as an unstable spatial anchor point. During the period when the unstable spatial anchor point is in an unstable state, the pose change process is divided into segments based on inertial sensing data and the direction of the navigation path, generating pose debt segments arranged in chronological order, and recording the cumulative pose change and corresponding path interval identifier for each pose debt segment; For each pose debt segment, determine the front stable spatial anchor point and the rear stable spatial anchor point that are spatially continuous with the pose debt segment. Then, connect each pose debt segment sequentially according to the connection relationship between the front stable spatial anchor point, the pose debt segment, and the rear stable spatial anchor point to form a pose debt chain. When a building observation primitive corresponding to an unstable spatial anchor point is detected, the current observation pose is obtained and compared with the predicted pose at the end of the pose debt chain. Based on the pose deviation, reverse write-back correction is performed on the pose debt segment, and closure correction is performed on the set of stable spatial anchor points.

8. The AR navigation method for buildings based on visual recognition according to claim 1, characterized in that, The output of stable building AR navigation results includes: Based on the set of stable spatial anchor points after closure correction, the guide information identifier, spatial position and attitude direction corresponding to each stable spatial anchor point are read, and each stable spatial anchor point is bound to the corresponding guide information to generate an AR guide content display set. Based on the current pose of the terminal device, the spatial position and orientation of each stable spatial anchor point, spatial projection positioning processing is performed on each guide content in the AR guide content display set to determine the display position, display orientation and display range of each guide content in the current view. Based on the building boundary information and scene area information in the current frame image data, the spatial occlusion relationship between each guide content and the building observation primitive is determined. Occlusion suppression processing is performed on the guide content within the occlusion area, and visible display processing is performed on the guide content not within the occlusion area, generating the AR guide display result of the current frame. Based on the changes in the terminal device's pose, the changes in the stable spatial anchor point's position, and the current frame's AR navigation display results, the display position, orientation, and status of each navigation content are continuously updated to output stable architectural AR navigation results.

9. A visual recognition-based AR navigation system for buildings, comprising the visual recognition-based AR navigation method for buildings as described in any one of claims 1 to 8, characterized in that, include: The pre-acquisition modeling module is used to acquire continuous image data and inertial sensing data of the target building area and generate a set of building observation primitives; The observation embryo construction module is used to construct path-condition reversible observation embryos based on the set of building observation primitives and to establish a competitive constraint graph of observation embryos. The observation candidate generation module is used to collect the current frame image data and inertial sensor data and input them into the improved DDRNet network, and generate a candidate set of building observation primitives by combining the observation embryo competition constraint map; The matching and counter-proof module is used to generate candidate matching chains based on the candidate set of building observation primitives, path-condition reversible observation embryos, and observation embryo competition constraint graphs, and obtain stable matching results. The anchor point correction module is used to establish a set of stable spatial anchor points based on the stable matching results and map navigation information. When the spatial anchor points become unstable, it constructs a pose debt chain and performs closure correction. The navigation display module is used to project and display AR navigation content based on a corrected set of stable spatial anchor points and dynamically update the navigation results.

Images