Geographic space knowledge base soft indexing method and system based on 6DOF and high-dimensional Gaussian distribution

Abstract

The invention discloses a geospatial knowledge base soft indexing method and system based on 6DOF and high-dimensional Gaussian distribution, and belongs to the field of geospatial information processing and knowledge base retrieval. The method comprises the following steps: modeling through mu (optimal 6DOF attitude) and sigma (attitude sensitivity) of high-dimensional Gaussian, breaking through the limitation of single-point binding, and reducing missed recall and error recall; a continuous probability score is used for replacing a fixed threshold value, the recall priority can be dynamically adjusted, and the robustness of a multi-knowledge-unit overlapping scene is improved; a sight line influence factor is introduced to solve the problem of wrong recall of back knowledge; the position component is highlighted through the weight matrix, so that the sensitivity of the index to the position change meets the actual demand, and the response timeliness is improved. And through a matching relationship between a high-dimensional Gaussian distribution modeling knowledge unit and a user 6DOF attitude, a continuous index score is calculated to realize dynamic recall, so that the problems of inaccurate spatial representation, rigid index mechanism, neglected view angle and weight imbalance existing in the existing geographic space knowledge base index technology are solved.

Description

Technical Field

[0001] This invention belongs to the field of geospatial information processing and knowledge base retrieval, specifically involving a geospatial knowledge base soft indexing method and system based on 6DOF (six degrees of freedom) and high-dimensional Gaussian distribution. Background Technology

[0002] Geospatial knowledge base indexing establishes the association between knowledge units and geospatial information through technical means, realizing rapid knowledge retrieval based on user location and posture, and is widely used in scenarios such as AR navigation, intelligent navigation, and spatial decision support. Existing technologies mainly include three categories: (1) 2D image feature matching and recognition. Based on image texture features such as SIFT and ORB, similarity matching is performed with a preset image database to retrieve associated knowledge. This technology relies on the clarity of image texture and is not robust enough in scenarios with changes in lighting or lack of texture. (2) Simple 3D coordinate retrieval. The user's 3D coordinates are obtained through GPS and SLAM, and the difference between the Euclidean distance / Manhattan distance and the coordinates of the knowledge unit is calculated. A fixed threshold (such as 5 meters) is set to determine whether to retrieve the knowledge unit. This method only focuses on the coordinates of a single point and ignores the spatial distribution characteristics of the knowledge unit. (3) Semantic retrieval. The semantics of the user's text are parsed through natural language processing and matched with the semantic tags of the knowledge unit. This technology relies on the user's active input and cannot achieve passive real-time retrieval based on spatial posture.

[0003] Existing technologies suffer from four key defects, leading to insufficient retrieval accuracy and robustness: (1) Limited spatial representation. Binding knowledge units to single-point coordinates (latitude, longitude, and altitude) fails to represent the actual distribution range and ambiguous boundaries, easily resulting in edge-based missed recalls and adjacent false recalls (such as knowledge confusion between adjacent display cases in a museum exhibition area). (2) Rigid indexing mechanism. Using a "yes / no" hard index with fixed threshold judgments, it is impossible to dynamically adjust the recall priority based on knowledge importance and user preferences. For example, core exhibits and auxiliary descriptions in adjacent display cases cannot achieve differentiated recall. (3) Viewpoint indifference. Without considering the user's line of sight, it is easy to recall knowledge that is out of sight (such as recalling exhibit B behind the user when the user is facing exhibit A), reducing the interactive experience and even causing safety hazards (such as false recall of road condition information outside the field of vision in intelligent driving scenarios). (4) Unreasonable weight allocation. Assigning the same weight to the 6DOF position (x, y, z) and rotation (rx, ry, rz) ignores the core requirement that position has a much greater impact on indexing than orientation in actual applications, leading to indexing accuracy deviations. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a soft indexing method and system for geospatial knowledge bases based on 6DOF and high-dimensional Gaussian distribution, thereby solving the problems of inaccurate spatial representation, rigid indexing mechanism, neglect of perspective, and weight imbalance in existing geospatial knowledge base indexing technologies.

[0005] In this specification, "soft index" specifically refers to a mechanism that calculates continuous index scores and dynamically adjusts recall priorities through probability density, which is different from "yes / no" hard indexes. It can solve problems such as inaccurate spatial representation, rigid mechanism, neglect of perspective, and weight imbalance.

[0006] The geospatial knowledge base soft indexing method based on 6DOF and high-dimensional Gaussian distribution provided by this invention models the matching relationship between knowledge units and user 6DOF poses through high-dimensional Gaussian distribution, and calculates continuous index scores to achieve dynamic recall. The specific steps are as follows: Step 1: Constructing a geospatial knowledge base Each knowledge unit is associated with a high-dimensional Gaussian distribution parameter whose dimension matches the 6DOF pose. The random variable is the user's 6DOF pose, and the orientation is described using Euler angles. The specific steps are as follows: (1) Knowledge unit division: Divide into independent knowledge units according to application scenarios (such as museums, urban blocks). Each unit contains basic information (name, content), semantic tags and spatial attributes (such as the museum's "Bronze Ware Exhibition Area" and "Shang Dynasty Bronze Ding").

[0007] (2) Associate high-dimensional Gaussian distribution parameters with each knowledge unit, with the parameter dimensions matching 6DOF, as specifically defined below: (2-1) Mean vector μ: 6-dimensional vector This matches the user's 6DOF pose in three spatial dimensions (x-axis, y-axis, z-axis) and three orientation dimensions (Eulerian angles rx (r stands for rotation, rx refers to rotation around x), ry, and rz). For example, 1.5m directly in front of the exhibit. The "optimal orientation" refers to the rotation angle when facing the exhibit. Euler angles need to be phase-normalized, with the value range limited to [-π, π]. Periodicity issues are resolved by adjusting ±2π. This mean vector as a whole represents the 6DOF pose of the user and knowledge unit with optimal matching.

[0008] (2-2) Covariance matrix Σ: 6-dimensional diagonal matrix The diagonal elements are the attitude deviation attenuation parameters. The smaller the variance, the faster the probability of the user's attitude deviating from the mean vector μ decreases, which means higher sensitivity. The specific value rules are as follows: Key Exhibits and Exhibition Areas: Location Variance σ x =σ y =0.5m (corresponding to a variance of 0.25m²), orientation variance σ rx =5° (corresponding to 0.087 rad, variance 0.0076 rad²); Open exhibition area: Location variance σ x =σ y =2m (corresponding to a variance of 4.0m²), orientation variance σ rx =30° (corresponding to 0.52 rad, variance 0.52 rad²); z-axis (height): Since the user's height changes little and the sensitivity requirement for height is low, the variance is uniformly set to 0.01m².

[0009] (3) Knowledge base storage: The knowledge unit basic information, semantic tags and high-dimensional Gaussian distribution parameters (μ, Σ) are stored in association, and the parameters are dynamically updated (such as updating μ after the exhibit is moved).

[0010] Step 2: Obtain user 6DOF information and calculate line-of-sight vectors to achieve real-time acquisition and processing of user spatial status. The steps are as follows: (1) 6DOF Information Acquisition: Sensors are adapted to different scenarios; outdoor systems use a combination of GPS, IMU, and electronic compass, while indoor systems use SLAM devices; the acquired data is represented as follows: P= ( x,y,z,rx,ry,rz), ​​which middle( x,y,z () represents the user's location, which is consistent with the coordinate system of the knowledge base (those skilled in the art can establish a unified coordinate system according to the actual scenario to ensure the consistency of location data), ( ) represents the head rotation angle, used to characterize the direction of the gaze; (2) Data preprocessing: Kalman filtering is used to eliminate random noise (such as GPS meter-level error) and improve data accuracy; (3) Calculation of gaze vector: Taking the z-axis of the user's head coordinate system as the positive direction of the gaze, through... The corresponding rotation matrix will change the reference vector Transform to the scene coordinate system to obtain the view vector. ; (4) Rotation matrix transformation formula: The rotation order is specified as follows ,in (5) Universal lock avoidance: When the pitch angle is close to When using quaternions to represent the attitude, the Euler angle gimbal lock problem is avoided.

[0011] Step 3: Calculate the knowledge unit index score The continuous index score is obtained by weighted fusion of positional similarity and line-of-sight influence factor. The specific steps are as follows: (1) The positional similarity calculation is based on a high-dimensional Gaussian probability density model, combined with a weight matrix to highlight the positional components. The formula is as follows: (1-1) Calculate the attitude difference vector The position difference represents the spatial deviation between the user and the optimal position of the knowledge unit, and the orientation difference represents the angular deviation of the line of sight, providing basic data for subsequent similarity calculation.

[0012] (1-2) Construct the weight matrix The six parameters within the parentheses represent the weights in the three directions of position and the three directions of rotation, respectively. (Can be slightly adjusted according to the scenario, must meet the following requirements) The rotation component weight is set to 0, which highlights the dominant role of the position in the index through weight allocation, and fits the actual application requirements.

[0013] (1-3) Probability density calculation (core formula) ,in It is the inverse of the covariance matrix. The closer the user's posture is to μ, The closer it is to 1, the more continuously the matching degree is represented.

[0014] (2) Calculation of line of sight influence factor The formula for quantifying the match between the user's gaze and the knowledge unit is as follows:

[0015] The parameters in the formula are defined as follows: The position difference vector represents the spatial distance between the user and the optimal location of the knowledge unit; ,in The vector pointing from the user to the center of the knowledge unit, where θ≤90° indicates the line of sight is towards the knowledge unit; k is the distance attenuation coefficient (default k=0.1, m -1 (It can be optimized according to the scenario). As a positive constraint, it ensures that a positive effect is introduced only when the gaze is directed toward the knowledge unit, avoiding false recall when the gaze is directed toward the knowledge unit.

[0016] (3) Calculation of comprehensive index score Weighted fusion of positional similarity and line-of-sight influence factors:

[0017] in α ∈[0.7 , 0.9] (default) α =0. 8 This ensures that the location is dominant while also optimizing the line of sight.

[0018] Step 4: Dynamic Priority Recall Process (1) Score ranking The overall score output by calling the index calculation module All knowledge units are sorted from highest to lowest score to form a preliminary recall sequence, ensuring that highly matched knowledge units are presented first.

[0019] (2) Double screening A dual mechanism of "Top-K filtering + threshold filtering" is adopted: first, the top-K units in the ranking are selected (such as in AR navigation scenarios). K =3, balancing display space and recall completeness), then filtering scores below the threshold. t (default τ= Units with a low matching degree of 0.1 are removed.

[0020] (3) Output and update of results: The filtered knowledge units are output to the user terminal in descending order of scores (such as AR glasses overlay display, APP graphic display). At the same time, as the user's posture changes, the process of steps two to four is repeated in real time to realize the dynamic update of the index results and adapt to the needs of real-time interactive scenarios.

[0021] In the above method, the indoor SLAM device mentioned in step two (1) is a combination of AR glasses depth camera and IMU.

[0022] In the above method, further, a、k、t、 Parameters such as ∑ can be optimized according to the scenario using the controlled variable method, without imposing specific numerical limits.

[0023] This invention also provides a soft indexing system for geospatial knowledge bases based on 6DOF and high-dimensional Gaussian distribution. This system implements the aforementioned soft indexing method for geospatial knowledge bases based on 6DOF and high-dimensional Gaussian distribution, serving as the hardware and software carrier of the method. The system adopts a modular design and specifically includes: The knowledge base construction module, as the core carrier of the knowledge base, realizes the creation, storage and updating of geospatial knowledge base. Its main function is to configure high-dimensional Gaussian distribution parameters (μ, Σ) for each knowledge unit and establish the correlation between knowledge information and spatial attitude.

[0024] The knowledge base construction module adopts a "client-side editing + server-side storage" architecture: the client provides a visual operation interface, supporting users to input basic knowledge information, mark the best related positions, and select high / medium / low sensitivity (the system automatically matches the preset Σ value); at the storage level, MySQL is used to store structured knowledge data (basic information, semantic tags), and Redis is used to cache frequently accessed μ and Σ parameters, thereby improving retrieval response efficiency through caching optimization.

[0025] The 6DOF information input module is used to realize real-time acquisition of user 6DOF data, noise removal and other preprocessing operations. At the same time, it calculates the line-of-sight vector based on the rotation matrix to provide high-quality user space state data for the index calculation module.

[0026] The 6DOF information input module selects sensing devices according to scene requirements at the hardware level: GPS module, IMU module and electronic compass are integrated for outdoor scenes, and SLAM device (such as AR glasses depth camera + IMU) is used for indoor scenes; at the software level, existing known device drivers are integrated, random noise (such as GPS meter-level error) is eliminated by Kalman filtering algorithm, line-of-sight vector calculation is completed based on OpenGL rotation matrix library, and finally the preprocessed data is output to the index calculation module.

[0027] The index calculation module, as the core calculation unit of the system, calls the knowledge parameters (μ, Σ) from the knowledge base module and the user data from the 6DOF information input module to calculate the position similarity in sequence. Factors affecting vision and composite index score .

[0028] The index calculation module uses C++ to write the core calculation code and optimizes matrix operations based on the Eigen math library (such as...). To improve computational efficiency, a caching mechanism was designed to temporarily store the most recent 5 frames of user data and frequently accessed knowledge parameters, avoiding repeated reading of server data. Communication was established with the knowledge base module (for reading parameters) and the 6DOF input module (for reading user data) through data interfaces to ensure smooth data flow.

[0029] The knowledge retrieval module receives the comprehensive score output by the index calculation module, sorts and filters it to obtain the final retrieval result and outputs it to the user terminal. At the same time, it collects user interaction feedback for knowledge base parameter optimization.

[0030] Supports user configuration of Top-K and t The parameters are adapted to different scenarios; the output layer provides multi-terminal interfaces to adapt to various output formats such as augmented reality overlay display and application graphic display; the feedback layer collects user interaction data (such as operation data of users manually selecting knowledge units) through the interface, and transmits this data back to the knowledge base module for dynamic optimization of μ and Σ parameters.

[0031] Furthermore, in the aforementioned system, the terminal interface of the knowledge retrieval module includes an AR glasses HDMI interface, a mobile APP HTTP interface, and a computer Socket interface.

[0032] In the above system, the knowledge base module provides knowledge parameters to the index calculation module; the 6DOF input module provides user data to the index calculation module; the index calculation module outputs score data to the knowledge retrieval module; the knowledge retrieval module outputs the final result to the user terminal; at the same time, the interactive feedback data generated by the user is transmitted to the knowledge base module for optimization and adjustment of the knowledge base parameters.

[0033] This invention also provides a computer storage medium storing a computer program; when the program is invoked by a processor, it executes all the steps of the aforementioned geospatial knowledge base soft indexing method based on 6DOF and high-dimensional Gaussian distribution (knowledge base construction, 6DOF information processing, index score calculation, and knowledge retrieval). This storage medium can be integrated into a system server / client or provided as a standalone device.

[0034] Compared with existing technologies, the present invention has the following beneficial effects: (1) Precise pose matching: By modeling μ (best 6DOF pose) and Σ (pose sensitivity) of high-dimensional Gaussian, the limitation of "single point binding" is overcome, and the missed recall and false recall are reduced.

[0035] (2) Flexible soft indexing mechanism: Continuous probability scores are used instead of fixed thresholds, which can dynamically adjust the recall priority and improve the robustness of scenarios with multiple knowledge units overlapping.

[0036] (3) Perspective-based retrieval: Introducing the line of sight influence factor to achieve "what you see is what you get" and solve the problem of false recall of knowledge from the opposite direction.

[0037] (4) Reasonable weight allocation: By highlighting the position component through the weight matrix, the sensitivity of the index to position changes is made to meet the actual needs, thereby improving the timeliness of response. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the overall process of the geospatial knowledge base soft indexing method based on 6DOF and high-dimensional Gaussian distribution of this invention.

[0039] The diagram illustrates the closed-loop logic of "attitude acquisition - score calculation - dynamic recall", clearly defining the input data (user 6DOF data, knowledge base parameters) and output results (recall results) for each step.

[0040] Figure 2 This is a data flow diagram of the geospatial knowledge base soft indexing system module based on 6DOF and high-dimensional Gaussian distribution as described in this invention.

[0041] This diagram visually illustrates the interaction between the modules, reflecting the hierarchical architecture of "data storage (knowledge base construction module) - data acquisition (6DOF information input module) - data calculation (index calculation module) - result output (knowledge retrieval module)", and clearly defining the functional boundaries of each module. Detailed Implementation

[0042] Example 1 This example demonstrates a museum AR guided tour scenario, with the main steps being: knowledge base construction → 6DOF information acquisition → index score calculation → knowledge retrieval.

[0043] The implementation scenarios and preliminary preparations are as follows: Scene parameters: The bronze exhibition hall of a history museum was selected as the implementation scene. The exhibition hall is 20m×30m×3m in size and divided into 3 knowledge units: A (bronze ware exhibition area, low sensitivity), B (Shang Dynasty bronze ding, high sensitivity), and C (casting process description, medium sensitivity). The user terminal uses HoloLens 2 AR glasses, which supports SLAM positioning and AR overlay display.

[0044] Equipment deployment: The server is configured with an Intel Xeon E3 processor, 16GB of memory, and a 500GB SSD to meet the requirements for data storage and real-time computing; for positioning calibration, four AprilTag visual markers are deployed in the four corners of the exhibition hall with coordinates (0,0,0), (20,0,0), (0,30,0), and (20,30,0) respectively, for SLAM positioning accuracy calibration to ensure the accuracy of user location data.

[0045] Preset parameters: Establish a coordinate system for the exhibition hall, with the origin set at the midpoint of the ground at the entrance; x-axis range 0→20m, y-axis range 0→30m, z-axis range 0→3m; the core algorithm parameter is preset as follows: fusion coefficient. Distance attenuation coefficient Recall threshold Top-K value Weight matrix .

[0046] Step 1: Building a Knowledge Base for the Bronze Exhibition Hall The construction process follows the core logic of "unit division - parameter configuration - associated storage". The specific configurations of the three knowledge units are as follows: (1) Knowledge Unit A (Low sensitivity, exhibition area level) Positioned as knowledge at the level of bronze ware exhibition area, the basic content covers the overview of bronze ware exhibition area from Shang Dynasty to Western Zhou Dynasty, and the semantic tags are set as "bronze ware" and "exhibition area".

[0047] Core high-dimensional Gaussian parameters: Mean vector:

[0048] Covariance matrix:

[0049] Parameter characteristics: The sensitivity to the location and orientation of the open exhibition area is relatively large, and it is suitable for the large-scale coverage of the exhibition area.

[0050] (2) Knowledge Unit B (High Sensitivity, Core Exhibits) Positioned as a core exhibit, it focuses on bronze ding vessels from the late Shang Dynasty. The basic content includes core information such as the vessel's overall height of 1.2m and an analysis of the taotie pattern. The semantic tags are set as "Shang Dynasty" and "core exhibit".

[0051] Core high-dimensional Gaussian parameters: Mean vector:

[0052] Covariance matrix:

[0053] Parameter characteristics: Both position and orientation variances are extremely small, ensuring accurate matching when the user's posture is close to the optimal associated position, adapting to the high-precision retrieval needs of core exhibits.

[0054] (3) Knowledge Unit C (Medium sensitivity, process description type) Positioned as process description knowledge, the basic content is a flowchart of the mold casting process, and the semantic tag is set to "casting process".

[0055] Core high-dimensional Gaussian parameters: Mean vector:

[0056] Covariance matrix:

[0057] Parameter characteristics: The variance of position and orientation is at an intermediate level, balancing search accuracy and range flexibility, and adapting to the auxiliary search needs of process specifications.

[0058] Storage method: MySQL stores the full information, and Redis caches 3 units of μ and Σ (valid for 24 hours).

[0059] Step 2: Obtain user 6DOF information and gaze vector (1) Data collection The user stands 1 meter diagonally in front of Knowledge Unit B (Shang Dynasty bronze ding), and uses the SLAM and IMU modules of the AR glasses to collect raw 6DOF data. The raw data is represented as follows: The data contains random noise related to position and rotation angle.

[0060] (2) Data preprocessing The Kalman filter algorithm is used to denoise the original data, eliminating random errors, and the resulting accurate 6DOF attitude data is obtained. The -90° rotation angle has been converted to a radian value of -1.57 rad.

[0061] (3) Calculation of line of sight vector Based on user head rotation angle (Rotate -90° around the y-axis) via a rotation matrix Transform the z-axis reference vector (0, 0, 1) of the head coordinate system to the exhibition hall coordinate system, and calculate the gaze vector. That is, it points in the positive direction of the x-axis (the direction where knowledge unit B is located).

[0062] (4) Data transmission AR glasses transmit pre-processed data via WiFi network. and The data is transmitted to the server in real time at a frequency of 10Hz to meet the real-time requirements of dynamic updates.

[0063] Step 3: Calculate the knowledge unit index score This step uses the user data obtained in step two as the core basis and calculates the index score of each knowledge unit by combining preset parameters. The specific process is as follows: (1) Basis for calculation The core user data is 6DOF posture. With line of sight vector (Pointing to knowledge unit B); Preset key parameters include the weight matrix. Fusion coefficient Distance attenuation coefficient Recall threshold The core parameters of each knowledge unit (mean vector μ, covariance matrix Σ) are taken from the knowledge base constructed in step 1.

[0064] (2) Core calculation formula The calculation logic employs a weighted fusion of "positional similarity + line-of-sight influence factor." The core formula and explanation are as follows: First, positional similarity is calculated based on a high-dimensional Gaussian probability density model, with the formula being... ,in The first factor is the posture difference, where only the position component is involved in the calculation; the second is the line-of-sight influence factor, which quantifies the matching degree between the line of sight and the knowledge unit, and the formula is... , The cosine of the angle between the line of sight and the unit's direction (positive direction effective). The first factor is spatial distance; the second is the comprehensive score, which is achieved through weighted fusion to realize location dominance and line-of-sight optimization. The formula is: .

[0065] (3) Calculation process and results for each knowledge unit (3-1) Knowledge Unit A (Low Sensitivity, Exhibition Area Level) The core parameters are Location variance Calculation process: Position components of attitude difference Because it is far from the optimal location and has low sensitivity attenuation, ; line of sight away from the unit ( Therefore Result: Overall score The conditions for recall are not met.

[0066] (3-2) Knowledge Unit B (Gao Min, core exhibit) The core parameters are Location variance Calculation process: Position components of attitude difference High sensitivity characteristics make ; Line of sight unit ( ), after distance decay Result: Overall score If the recall conditions are met, the score is dominated by positive vision.

[0067] (3-3) Knowledge Unit C (Zhongmin, Process Description) The core parameters are Location variance Calculation process: Position components of attitude difference The medium sensitivity characteristic makes ; line of sight away from the unit ( Therefore Result: Overall score The conditions for recall are not met.

[0068] Step 4: Knowledge Retrieval and Dynamic Update Based on the comprehensive score results from step three, a closed-loop process of "sorting-filtering-outputting-dynamic updating" is executed, as follows: (1) Recall screening process First, sort the knowledge units from highest to lowest based on their overall scores. The sorting result is as follows: Then, a dual screening mechanism is implemented: first, the top 3 units (covering all knowledge units) are selected, and then those with scores below the threshold are filtered out. Of the units, only knowledge unit B ultimately met the recall criteria. The AR glasses overlaid detailed information on the bronze ding, including its dimensions (1.2m in height), an analysis of the taotie pattern, and the historical background of the late Shang Dynasty.

[0069] (2) Dynamic update verification (scenario-based verification) As user posture changes, the system repeats steps 2-4 (6DOF information acquisition - score calculation - recall) in real time to verify the dynamic recall effect. The key scenario verification results are as follows: (2-1) Scenario 1: Turning to the sensitive unit C The user moves to the best relevant location in Knowledge Unit C. line-of-sight vector (Looking at C directly). Calculation logic: position difference. Make position similarity Positive gaze matching reduces the gaze influence factor. Overall score The AR terminal automatically switches to display the casting process diagram, enabling precise and dynamic recall.

[0070] (2-2) Scenario 2: Regression of low-sensitivity unit A Users move to the center of the exhibition area. (The optimal relevance of knowledge unit A), with no specific direction of sight. Calculation logic: position difference. Make position similarity A wide field of vision coverage reduces the influencing factor of vision. Overall score The AR terminal switches to display an overview of the bronze ware exhibition area.

[0071] (3) Advantages of dynamic updates After the user's posture changes, the system can switch knowledge during the user's movement by transmitting data at 10Hz and performing real-time calculations. This avoids the problem of "information lag after location change" caused by fixed indexes and adapts to the real-time interactive needs of AR navigation.

Claims

1. A soft indexing method for geospatial knowledge bases based on 6DOF and high-dimensional Gaussian distribution, characterized in that... The steps include the following: Step 1: Constructing a geospatial knowledge base For each knowledge unit, a high-dimensional Gaussian distribution parameter is associated, and a 6DOF pose is matched. The random variable is the user's 6DOF pose, and the orientation is described by Euler angles. The specific steps are as follows: (1) Knowledge unit division: Independent knowledge units are divided according to application scenarios. Each unit contains basic information, semantic tags and spatial attributes; (2) Associate high-dimensional Gaussian distribution parameters with each knowledge unit, with the parameter dimensions matching 6DOF, as specifically defined below: (2-1) Mean vector μ: 6-dimensional vector These are used to match the user's 6DOF pose in three spatial dimensions (x-axis, y-axis, z-axis) and three orientation dimensions (Eulerian angles rx, ry, rz), respectively. For "best associated position", ( The "optimal orientation" refers to the rotation angle when facing the exhibit; the Euler angles need to be phase-normalized and their range is limited to [-π, π]. The periodicity problem is solved by adjusting ±2π. The mean vector as a whole represents the 6DOF posture of the optimal match between the user and the knowledge unit. (2-2) Covariance matrix Σ: 6-dimensional diagonal matrix is a diagonal covariance matrix adapted to 6DOF user space pose, with elements representing pose deviation decay parameters. The smaller the variance, the faster the probability decays when the user pose deviates from the mean vector μ, which means higher sensitivity. The specific value rules are as follows: Key Exhibits and Exhibition Areas: Location Variance σ x =σ y =0.5m, orientation variance σ rx =5°; Open exhibition area: Location variance σ x =σ y =2m, orientation variance σ rx =30°; z-axis: Since the user's height changes little and the sensitivity to height is low, the variance is uniformly set to 0.01m². (3) Knowledge base storage: The knowledge unit basic information, semantic tags and high-dimensional Gaussian distribution parameters μ and Σ are stored in association, and the parameters are dynamically updated; Step 2: Obtain user 6DOF information and calculate line-of-sight vectors to achieve real-time acquisition and processing of user spatial status. The steps are as follows: (1) 6DOF Information Acquisition: Sensors are adapted to different scenarios; outdoor systems use a combination of GPS, IMU, and electronic compass, while indoor systems use SLAM devices; the acquired data is represented as follows: P= ( x,y,z,rx,ry,rz). in( x,y,z () represents the user's location, which is consistent with the knowledge base coordinate system. ( ) represents the head rotation angle, used to characterize the direction of the gaze; (2) Data preprocessing: Kalman filtering is used to eliminate random noise and improve data accuracy; (3) Calculation of gaze vector: Taking the z-axis of the user's head coordinate system as the positive direction of the gaze, through... The corresponding rotation matrix will rotate the reference vector (0) , 1 , 1) Transform to the scene coordinate system to obtain the view vector. ; (4) Rotation matrix transformation formula: The rotation order is specified as follows ,in , (5) Universal lock avoidance: When the pitch angle is close to At this time, quaternions are used to represent the attitude to avoid the Euler angle gimbal lock problem; Step 3: Calculate the knowledge unit index score The continuous index score is obtained by weighted fusion of positional similarity and line-of-sight influence factor. The specific steps are as follows: (1) Position similarity calculation is based on a high-dimensional Gaussian probability density model, combined with a weight matrix to highlight the positional components. The formula is as follows: (1-1) Calculate the attitude difference vector The position difference represents the spatial deviation between the user and the optimal position of the knowledge unit, and the orientation difference represents the angular deviation of the line of sight, providing basic data for subsequent similarity calculation. (1-2) Construct the weight matrix ,in These represent the weights in the three directions of the position, with the rotation component weight set to 0. This weight allocation highlights the dominant role of the position in the index, aligning with practical application needs. (1-3) Probability density calculation ,in It is the inverse of the covariance matrix. The closer the user's posture is to μ, The closer it is to 1, the more continuously the matching degree is represented by quantitative characteristics; (2) Calculation of line of sight influence factor The formula for quantifying the match between the user's gaze and the knowledge unit is as follows: , The parameters in the formula are defined as follows: The position difference vector represents the spatial distance between the user and the optimal location of the knowledge unit; ,in A vector pointing the user to the center of the knowledge unit. θ ≤90° indicates that the line of sight is directed towards the knowledge unit; k This is the distance attenuation coefficient; As a positive constraint, it ensures that a positive effect is introduced only when the gaze is directed toward the knowledge unit, avoiding false recall when the knowledge is reversed; (3) Calculation of comprehensive index score Weighted fusion of positional similarity and line-of-sight influence factors: , Where α∈[0.7,0.9], ensuring position dominance while also considering line-of-sight optimization; Step 4: Dynamic Priority Recall Process (1) Score ranking The overall score output by calling the index calculation module All knowledge units are sorted from highest to lowest score to form a preliminary recall sequence, ensuring that highly matched knowledge units are presented first. (2) Double screening A dual mechanism of "Top-K filtering + threshold filtering" is adopted: first, the top-K units in the ranking are selected, and then those with scores below the threshold are filtered out. τ (default τ= Units with a low matching degree of 0.1 are removed; (3) Output and update of results: The filtered knowledge units are output to the user terminal in descending order of scores. At the same time, as the user's posture changes, the process of steps two to four is repeated in real time to realize the dynamic update of the index results and adapt to the needs of real-time interactive scenarios.

2. The method according to claim 1, characterized in that, The indoor SLAM device described in step two (1) is a combination of an AR glasses depth camera and an IMU.

3. The method according to claim 1, characterized in that, parameter α, k, τ ∑ is optimized according to the scenario using the method of controlling variables.

4. A geospatial knowledge base soft indexing system based on 6DOF and high-dimensional Gaussian distribution, characterized in that, The system executes the geospatial knowledge base soft indexing method based on 6DOF and high-dimensional Gaussian distribution as described in any one of claims 1 to 3, and serves as the hardware and software carrier of the method.

5. The method according to claim 4, characterized in that, Includes the following structural modules, The knowledge base construction module configures high-dimensional Gaussian distribution parameters μ and Σ for each knowledge unit, establishes the association between knowledge information and spatial attitude, and realizes the creation, storage and updating of geospatial knowledge base; The 6DOF information input module is used to realize the real-time acquisition of user 6DOF data and noise removal. At the same time, it calculates the line-of-sight vector based on the rotation matrix, providing high-quality user space state data for the index calculation module. The index calculation module, as the core calculation unit of the system, calls the knowledge parameters μ and Σ from the knowledge base module and the user data from the 6DOF information input module to calculate the position similarity in sequence. Visual Influence Factors and composite index score ; The knowledge retrieval module receives the comprehensive score output by the index calculation module, sorts and filters it to obtain the final retrieval result and outputs it to the user terminal. At the same time, it collects user interaction feedback for knowledge base parameter optimization. The knowledge base construction module provides knowledge parameters to the index calculation module; the 6DOF input module provides user data to the index calculation module; the index calculation module outputs score data to the knowledge retrieval module; the knowledge retrieval module outputs the final result to the user terminal; at the same time, the interactive feedback data generated by the user is transmitted to the knowledge base module for optimization and adjustment of the knowledge base parameters.

6. The module according to claim 5, characterized in that, The knowledge base construction module adopts an architecture of "client-side editing + server-side storage": the client provides a visual operation interface, supporting users to input basic knowledge information, mark the best related positions, and select high / medium / low sensitivity; at the storage level, MySQL is used to store structured knowledge data, and Redis is used to cache frequently accessed μ and Σ parameters, thereby improving retrieval response efficiency through caching optimization.

7. The module according to claim 5, characterized in that, The 6DOF information input module selects sensing devices according to the scene requirements at the hardware level: GPS module, IMU module and electronic compass are integrated for outdoor scenes, and SLAM device is used for indoor scenes; at the software level, it integrates existing known device drivers, eliminates random noise through Kalman filtering algorithm, completes line-of-sight vector calculation based on OpenGL rotation matrix library, and finally outputs the preprocessed data to the index calculation module.

8. The module according to claim 5, characterized in that, The index calculation module uses C++ to write the core calculation code, optimizes matrix operations based on the Eigen math library to improve calculation efficiency, designs a caching mechanism to temporarily store the most recent 5 frames of user data and frequently accessed knowledge parameters to avoid repeatedly reading server data, and establishes communication with the knowledge base module and 6DOF input module through data interfaces to ensure smooth data flow.

9. The module according to claim 5, characterized in that, The knowledge retrieval module supports user configuration of Top-K and... τ The parameters are adapted to different scenarios; the output layer provides multi-terminal interfaces to adapt to various output formats such as augmented reality overlay display and application graphic display; the feedback layer collects user interaction data through the interface and transmits the data back to the knowledge base module for dynamic optimization of μ and Σ parameters; the terminal interfaces of the knowledge retrieval module include AR glasses HDMI interface, mobile APP HTTP interface and computer Socket interface.

10. A computer storage medium storing a computer program; when the program is invoked by a processor, it executes all the steps of the above-described geospatial knowledge base soft indexing method based on 6DOF and high-dimensional Gaussian distribution, wherein the storage medium is integrated into a system server or client, or provided as a standalone carrier.

Images