A three-dimensional building space navigation system and navigation method based on big data matching

Abstract

The present application relates to the field of stereoscopic navigation technology and provides a three-dimensional navigation system and navigation method for building spaces based on big data matching, including: a three-dimensional scene construction module, configured to collect three-dimensional structural data of a target area through a building information model interface, and to fuse spatial topological relationships to generate a three-dimensional navigation network, wherein the three-dimensional navigation network includes vector relationships between vertically associated channels and horizontal passages on each floor; an environmental feature extraction module, which generates a three-dimensional point cloud model; an intelligent channel planning module, which is configured to receive destination coordinates submitted by a user terminal; and a visual positioning navigation module, which outputs navigation instructions including directional deflection angles and displacement distances. The technical solution disclosed in the present application integrates the spatial topological relationships of building areas through the three-dimensional scene construction module to establish a vector navigation network including three-dimensional traffic elements; and feature matching between real-time images and pre-stored point cloud data to achieve centimeter-level spatial perception accuracy.

Description

Technical Field

[0001] The present application relates to the field of navigation technology, and in particular to a three-dimensional building space navigation system and navigation method based on big data matching. Background Art

[0002] The contents of this section merely provide background information related to this application and may not constitute prior art.

[0003] In large, complex building scenarios such as industrial parks, schools, hospitals, and shopping malls, spatial navigation services generally face technical bottlenecks such as low positioning accuracy, difficulty in determining spatial levels, and insufficient channel planning effectiveness. These technical difficulties are mainly reflected in the following dimensions: First, at the spatial signal level, dense building clusters can cause multipath effects on GNSS satellite signals, and indoor and outdoor cellular base station signals are shielded and attenuated, making it difficult for conventional positioning technologies to break through the meter-level accuracy limit; second, at the spatial dimension level, existing navigation systems mostly rely on a two-dimensional geographic coordinate system and lack the ability to model the three-dimensional spatial topology of a building. This is especially true in high-rise building scenarios, where latitude and longitude coordinates cannot be used to determine the vertical height level of the user; third, at the dynamic environment level, such buildings often have complex three-dimensional spatial structures, with physical isolation or access control between different functional areas, and the dynamic adjustment characteristics of the internal facility layout, making it difficult for traditional navigation systems to build accurate real-time environmental recognition models.

[0004] For non-fixed user groups such as visitors, it is impossible to determine the actual building floor they are on through planar positioning information, and it is also difficult to plan effective channels in complex spaces that include three-dimensional transportation nodes such as elevator corridors and safety passages. This has become a key technical obstacle restricting the practical application of smart building navigation systems. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a three-dimensional building space navigation system based on big data matching. The three-dimensional building space navigation system based on big data matching disclosed in this application can be realized.

[0006] The purpose of this application is achieved through the following technical solutions:

[0007] A three-dimensional building space navigation system based on big data matching, comprising:

[0008] A three-dimensional scene construction module is configured to collect three-dimensional structural data of a target area through a building information model interface, fuse spatial topological relationships to generate a three-dimensional navigation network, wherein the three-dimensional navigation network includes vector relationships between vertical associated channels and horizontal passageways of each floor;

[0009] An environmental feature extraction module collects image information of the three-dimensional navigation network and generates a three-dimensional point cloud model through a stereo vision algorithm;

[0010] an intelligent channel planning module configured to receive the destination coordinates submitted by the user terminal and calculate an optimal route in the three-dimensional navigation network;

[0011] The visual positioning and navigation module is configured to capture the visual data stream of the front scene in real time through the camera of the user terminal, use the feature point matching algorithm to spatially align the real-time visual data with the three-dimensional point cloud model, and output navigation instructions including the direction deflection angle and displacement distance.

[0012] This application's technical solution effectively overcomes the inherent flaws of traditional positioning technology by constructing a three-dimensional spatial navigation system driven by visual features. This application integrates the spatial topology of building areas through a three-dimensional scene construction module, establishing a vector navigation network that includes three-dimensional transportation elements. A mobile laser scanning device is then used to collect millimeter-level precision three-dimensional point cloud data along the navigation path. When a user initiates a navigation request, centimeter-level spatial perception accuracy is achieved by matching the real-time image with the features of the pre-stored point cloud data, ultimately presenting three-dimensional spatial guidance.

[0013] The intersections and passages in buildings are very similar, and when performing visual recognition and visual matching, they are easily confused by similar intersections and passages, making it impossible to accurately determine the current location and thus unable to guide the user to the preset location. To this end, this application provides the following technical solutions:

[0014] In some possible embodiments, the environmental feature extraction module includes:

[0015] The channel information extraction unit performs non-overlapping channel segmentation on the three-dimensional navigation network and generates a channel vector dataset with spatial independence;

[0016] 3D information input unit, which reconstructs the 3D point cloud model of each channel based on multi-view video stream;

[0017] A key feature extraction unit is configured to extract a feature point set with rotation invariance from a three-dimensional point cloud model;

[0018] The channel fingerprint establishment unit obtains the feature point set of each channel and constructs spatial fingerprint feature information containing the channel unique identifier based on the feature point differences of all feature point sets, wherein the spatial fingerprint feature information of each channel is composed of multi-dimensional spatial geodesic distance information between feature points.

[0019] This technical solution introduces a channel-level spatial fingerprint feature construction method, differentiating itself from traditional recognition methods based on isolated feature points or local feature fragments. By establishing a global channel spatial feature association model, it generates unique multidimensional spatial fingerprint feature information for each path. This matrix quantifies the topological differences between feature points, forming essential distinguishing features between channels, effectively suppressing the risk of feature confusion caused by local environmental interference during the path matching process.

[0020] The similarity of each channel in a building is very high. In practice, it is difficult to accurately distinguish the differences between all channels in a building. To this end, this application provides the following technical solutions:

[0021] In some possible embodiments, the channel fingerprint establishing unit includes:

[0022] Feature semantic classifier, obtains the feature points in all feature point sets to count the types of feature points;

[0023] Feature optimization filter, dynamically removes redundant feature points whose feature category distribution entropy value is lower than the preset threshold;

[0024] Feature point allocator, which assigns several types of feature points to each channel. The types of feature points assigned to each channel are different.

[0025] The spatial fingerprint compiler extracts the feature points of the corresponding category for each channel, calculates the distance between each feature point, and generates the spatial fingerprint feature information corresponding to the channel.

[0026] This technical solution implements a differentiated feature category allocation mechanism, assigning a unique feature category combination to each channel. This ensures that the spatial fingerprint matrices of different channels form discretized distinguishing characteristics along the feature type dimension. This allocation strategy, based on set theory, mathematically guarantees that the feature category sets of any two channels have a non-empty difference set, thereby establishing the category distinction boundary between channels.

[0027] When classifying feature points, the classification is generally based on the absolute value of the feature points. However, this classification method has a high overlap rate, and many feature points of different types will be classified into the same type, resulting in low accuracy in feature classification. To address this problem, this application provides the following technical solutions:

[0028] In some possible embodiments, the categories of feature points in the feature semantic classifier are assigned as follows:

[0029] A reference coordinate system for predefined standard eigenvectors in the eigenvector space;

[0030] Calculate the cosine similarity value between the vector of the feature point to be classified and the standard feature vector, and establish the relative position relationship in the feature space;

[0031] The feature vectors are spatially divided according to the preset similarity threshold interval to achieve deterministic attribution of feature point types.

[0032] In the technical solution provided by this application, the classification of feature points is based on their distribution in the reference coordinate system. This not only effectively distinguishes the differences between feature points, but also accurately describes the differences between feature points based on their proximity.

[0033] When selecting feature points of corresponding types for each channel, it is easy to fall into the local optimal error of having the feature points of a certain channel have the highest recognition, resulting in the inability to assign feature point types from a global perspective. To address this problem, this application provides the following technical solutions:

[0034] In some possible embodiments, the manner in which the feature point types are assigned to each channel in the feature point allocator is as follows:

[0035] Pre-set feature requirements, including feature point cardinality threshold and feature point distribution density threshold;

[0036] Based on the feature requirements, an allocation scheme for all feature point types of the corresponding channel is obtained, and each allocation scheme includes at least two types of feature points;

[0037] Obtain the distribution scheme of the feature point types corresponding to all channels and establish a feature point distribution matrix. Each row in the feature point distribution matrix represents the feature points that can be selected by a channel, and each column in the feature point distribution matrix represents a feature point. The elements in the feature point distribution matrix are 0 or 1. When it is 1, it indicates that the feature point meets the feature requirements of the corresponding channel. When it is 0, it indicates that the feature point does not meet the feature requirements of the corresponding channel.

[0038] At least two feature points of different categories are randomly selected from each row of the feature point allocation matrix to assign the category of the feature points to each channel.

[0039] The technical solution provided in this application transforms the feature point classification problem into an optimization problem involving selecting a number of feature points from each row of the feature point allocation matrix by constructing a feature point allocation matrix. This design makes it easier for the processing system to coordinate feature allocation strategies from a global perspective, ensuring that each channel obtains a discriminative feature combination, thereby improving the overall recognition accuracy of the navigation system.

[0040] There are a large number of channels in the entire building. In practice, if the current image information is matched with each channel, the matching time will be long, the navigation delay will be high, and the navigation guidance cannot be released in a timely manner.

[0041] In some possible embodiments, the visual positioning navigation module includes:

[0042] A channel information acquisition unit, which acquires a set of channels that the user can pass through based on the user's historical location and optimal travel route;

[0043] A visual information acquisition unit is used to acquire image information of the scene where the user is located, extract matching extreme value points from the image information, and calculate the distance between each matching extreme value point to generate matching information;

[0044] The position positioning unit compares the matching information with the spatial fingerprint feature information of the channels in the channel set, obtains the spatial fingerprint feature information of the channel with the highest matching degree with the matching information, and uses the channel as the user's location;

[0045] The navigation prompt unit sends navigation instructions to the user based on the user's location and the optimal route.

[0046] The technical solution provided by this application analyzes a user's historical trajectory data and the optimal path planning solution to construct a set of candidate channel probabilities, which characterizes the probability distribution characteristics of the user's current location. The system prioritizes matching the real-time visual data with the spatial fingerprint feature information corresponding to the candidate channels, thereby achieving precise positioning while reducing computational complexity.

[0047] In buildings, a lot of information will gradually disappear due to traces of use. For example, tire marks in an underground parking lot will dissipate due to the coverage of other tire marks, or disappear due to cleaning by cleaning staff. In addition, dust on the ground and non-fixed objects may also change position. Therefore, the spatial fingerprint characteristics of each channel will change. To address this problem, this application provides the following technical solutions:

[0048] In some possible embodiments, the position location unit includes:

[0049] The weight setter sets the confidence weight for each feature point based on its location and update time. The higher the height of the feature point from the floor, the higher the confidence weight; the closer the update time of the feature point, the higher the confidence weight.

[0050] A matching calculator calculates the cosine similarity between the matching information and each spatial fingerprint feature information to obtain initial comparison information;

[0051] The comprehensive calculator multiplies the initial comparison information with the confidence weight of each feature point to obtain the ranking information, and takes the channel with the highest ranking information as the user's location.

[0052] In the technical solution provided in this application, a confidence weight is set for each feature point according to its location. If the location of the feature point is high, the possibility of its change is small. Correspondingly, if the update time of the feature point is close, the possibility of its change is also small. In this way, when matching feature points, the influence of environmental changes on matching accuracy can be effectively avoided.

[0053] In some possible embodiments, the big data matching building space three-dimensional navigation system also includes: an environmental feature update module, which obtains image information collected along the way by users who have accurately reached their destination, and updates the image information to a three-dimensional point cloud model.

[0054] In some possible embodiments, feature points are extracted as follows:

[0055] S1: Convert the 3D point cloud model into a 3D voxel model;

[0056] S2: Construct a three-dimensional scale space to generate feature response information at different observation scales;

[0057] S3: For each voxel, check whether the feature response information is the maximum value in the 3D neighborhood. If so, set it as the extreme point; the neighborhood includes the 26 spatial neighbors of the current voxel and the corresponding positions of the upper and lower adjacent scales;

[0058] S4: Assign gradient magnitude and direction to each feature point and use the gradient magnitude and direction as the descriptor of the feature point

[0059] A three-dimensional navigation method for building space based on big data matching adopts the aforementioned three-dimensional navigation system for building space based on big data matching to perform three-dimensional navigation.

[0060] The technical solutions of the embodiments of the present application have at least the following advantages and beneficial effects:

[0061] This technical solution implements a differentiated feature category allocation mechanism, assigning a unique feature category combination to each channel. This ensures that the spatial fingerprint matrices of different channels form discretized distinguishing characteristics along the feature type dimension. This allocation strategy, based on set theory, mathematically guarantees that the feature category sets of any two channels have a non-empty difference, thereby establishing a category-level distinguishing boundary between channels. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a structural diagram of the building space three-dimensional navigation system based on big data matching in this application.

[0063] Figure 2 A partial schematic diagram of the three-dimensional navigation network.

[0064] Figure 3 Assign a matrix to 8×4 feature points. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solutions and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific implementation methods. The same figure marks in the accompanying drawings represent the same components. It should be noted that the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the described embodiments of this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0066] Compared to the embodiments shown in the drawings, feasible embodiments within the scope of protection of the present application may have fewer components, other components not shown in the drawings, different components, differently arranged components, or differently connected components, etc. In addition, two or more components in the drawings may be implemented in a single component, or a single component shown in the drawings may be implemented as multiple separate components.

[0067] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning understood by persons of ordinary skill in the field to which this application belongs. The words "first", "second" and similar terms used in the specification and claims of this application do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "a" or "an" do not necessarily indicate a quantitative limitation. "Up", "down" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0068] refer to Figure 1 : Example 1:

[0069] The first embodiment of the present application discloses a three-dimensional navigation system for building space based on big data matching, including: a three-dimensional scene construction module, an environmental feature extraction module, an intelligent channel planning module, and a visual positioning navigation module.

[0070] The three-dimensional scene construction module is configured to collect three-dimensional structural data of the target area through the building information model interface, fuse the spatial topological relationship to generate a three-dimensional navigation network, and the three-dimensional navigation network includes the vector relationship between the vertical associated channels and the horizontal passage channels of the floors.

[0071] In a building complex consisting of multiple independent or interconnected buildings, spatial nodes are topologically connected via single or multiple channels. This solution simplifies the actual building into a three-dimensional navigation network model of interconnected channels. Within this model, any pair of nodes can be selected as the starting and ending points to generate the corresponding multimodal travel path.

[0072] like Figure 2 As shown, Figure 2 The local three-dimensional navigation network is shown in Figure 2 Each line segment in represents a channel.

[0073] The three-dimensional navigation network includes two types of topological elements:

[0074] Vertical connecting passages: including elevator shafts, evacuation staircases and other three-dimensional transportation facilities

[0075] Horizontal traffic path: It is composed of corridors, aisles and other flat traffic facilities. After simplifying the building information, the final three-dimensional navigation network can clearly show the relationship between each location.

[0076] The environmental feature extraction module collects image information from the 3D navigation network and generates a 3D point cloud model using a stereo vision algorithm. In this solution, the environmental feature extraction module constructs a 3D point cloud model, essentially a 3D point cloud model of the entire target building area, for subsequent direction identification.

[0077] In actual application scenarios, navigation does not require the full 3D point cloud data of the complete building structure, but focuses on the geometric topological characteristics of the passage. Therefore, this solution implements a passage-level point cloud modeling solution: using spatial masking technology to isolate non-passage areas, only performing 3D reconstruction of passage elements such as elevator shafts and corridors, and generating a low-redundancy passage-level point cloud dataset.

[0078] The specific method is as follows:

[0079] The environmental feature extraction module includes: a channel information extraction unit, a three-dimensional information input unit, a key feature extraction unit, and a channel fingerprint establishment unit.

[0080] The channel information extraction unit performs non-overlapping channel segmentation on the three-dimensional navigation network and generates a channel vector dataset with spatial independence.

[0081] Specifically, the channel information extraction unit extracts a set of paths with clear entrance locations from the three-dimensional navigation network. The channels defined in this solution all have clearly defined entrance locations, allowing the system to accurately determine the spatial coordinates of the channel entrance when the user enters. This design is based on the following navigation requirements: users need to receive steering guidance at the channel entrance node, rather than continuous positioning within the channel.

[0082] Channel division is based on topological node distribution characteristics rather than physical length parameters:

[0083] If a traversable path does not contain an intermediate decision node, the entire path is considered as a single channel entity.

[0084] According to this rule, the channel information extraction unit can extract all channels from the three-dimensional navigation network. The specific navigation path is actually the combination of these channels.

[0085] The above is the extraction method of the channel vector data set extracted by the channel information extraction unit. After the channel vector data set is extracted, the three-dimensional point cloud model of the target building is actually simplified into the three-dimensional point cloud model of the channel in the channel vector data set, which greatly reduces the amount of data calculation.

[0086] The 3D information input unit reconstructs a 3D point cloud model for each channel based on the multi-view video stream. Specifically, the 3D information input unit obtains the multi-view video streams of all channels through a mobile acquisition terminal (including a handheld imaging device or a visual sensor mounted on an AGV vehicle) and performs channel-level 3D reconstruction based on the multi-view video streams. The 3D point cloud modeling process includes:

[0087] (1) Capture multi-view image sequences along the channel axis by moving the acquisition device;

[0088] (2) Apply binocular stereo vision algorithm to convert two-dimensional image sequences into three-dimensional point cloud topology data.

[0089] The key feature extraction unit is configured to extract a feature point set with rotation invariance from the three-dimensional point cloud model.

[0090] When performing feature matching, the key lies in extracting and identifying feature points. The feature point extraction method in this application is as follows:

[0091] S1: Convert the 3D point cloud model into a 3D voxel model.

[0092] Downsample the 3D point cloud model and convert the point cloud data into voxel data A={a i ={x i ,y i , z i}}, where x i ,y i , z i Represents the horizontal coordinate, vertical coordinate and height coordinate of the voxel, a i Represents voxel value, which is a grayscale value.

[0093] S2: Construct a three-dimensional scale space to generate feature response information at different observation scales.

[0094] The SIFT algorithm has a scale-invariant feature, that is, the extracted feature points will not change under different observation scales. The key is to extract feature response information based on different observation scales.

[0095] The specific method is as follows:

[0096] S21: Set the three-dimensional Gaussian kernel G(x, y, z, σ);

[0097] ;

[0098] Where π represents the circumference of a circle, e represents a natural constant, and σ represents the scale parameter of the Gaussian kernel. σ is used to control the degree of smoothing. The larger σ is, the wider the smoothing range is, corresponding to a coarser scale.

[0099] S22: Set the initial scale parameter σ0 and perform three-dimensional Gaussian convolution on the voxel data A to generate the first scale space L0; L0(x, y, z, σ0) = A(x, y, z) * G(x, y, z, σ0), where A(x, y, z) represents the voxel data;

[0100] S23: Generate S scale parameters, each scale parameter generates 1 scale space; σ k =σ0*2 （k / S） , where k represents the index of the scale parameter, S represents the total number of scale parameters, and σ k represents the kth scale parameter, k=0,1,2…S-1;

[0101] ;

[0102] Among them, L k represents the scale space of the kth layer, Indicates intermediate parameters;

[0103] S24: Downsample the voxel data A to obtain a new sampling group, and calculate the scale space at different scales by sampling the new sampling group according to the schemes S22 to S23.

[0104] Downsampling the voxel data A mainly involves reducing the voxel data A from high resolution to low resolution. In practice, this can be done by direct sampling, mean pooling, maximum pooling, etc. Downsampling is a well-known technique and will not be further elaborated here. In this way, in step S2, the scale space L at different scales and resolutions can be obtained. k .

[0105] The core goal of scale spaces with varying resolutions is to detect stable feature points across different levels of observation through multi-scale analysis. While the coordinates of the same physical point in the voxel grid change as the resolution decreases (doubling the voxel size), cross-group correspondence is maintained through the inheritance of scale parameters. The initial scale parameter for each group inherits the final scale of the previous group, ensuring scale continuity across groups. For example, if the final scale of the first group is 6.3 cm, the initial scale of the second group will also be 6.3 cm.

[0106] S3: For the voxel L(x, y, z, σ), check whether it is the maximum value in the 3D neighborhood. If so, set it as the extreme point; the neighborhood includes the 26 spatial neighbors of the current voxel (at the same scale) and the corresponding positions of the upper and lower adjacent scales.

[0107] For each pixel point, its scale space at different scales can be obtained, and whether it is an extreme point can be determined based on the scale space.

[0108] Specifically, because it involves the determination of extreme points at different resolutions, the specific solution is as follows:

[0109] S31: In each group, local extreme points are detected by 3D Gaussian pyramid, i.e. L(x, y, z, σ) of the contrast area.

[0110] S32: The extreme points of all groups are converted to a unified physical coordinate system, and then the extreme points of the same physical position and similar scale in different groups are merged, and repeated or redundant feature points are removed.

[0111] S33: Output the physical position, scale, and direction of the feature point. The physical position is the position in the three-dimensional point cloud model.

[0112] In this way, all extreme points can be extracted, which are the required feature points, and then the positions of all feature points can be obtained. After obtaining the positions of the feature points, downsampling the voxel data A is required to obtain the description method of the feature points.

[0113] S4: Assign a gradient magnitude and direction to each feature point, and use the gradient magnitude and direction as the descriptor of the feature point.

[0114] Specifically, S4 includes the following steps:

[0115] S41: For the detected feature points, calculate the gradient components Gx, Gy, and Gz in the x, y, and z directions;

[0116] S42: Calculate the gradient magnitude m and direction θ of the feature point;

[0117] ; , arctan represents the inverse tangent trigonometric function.

[0118] After extracting the gradient amplitude and direction of the feature point, it is used as the descriptor of the feature point, so that the feature point can have the characteristic of scale invariance, and the feature point can also be identified when the shooting angle and distance are changed.

[0119] The channel fingerprint establishment unit obtains the feature point set of each channel and constructs spatial fingerprint feature information containing the channel unique identifier based on the feature point differences of all feature point sets, wherein the spatial fingerprint feature information of each channel is composed of multi-dimensional spatial geodesic distance information between feature points.

[0120] Specifically, the above solution extracts all feature points from each channel. These feature points are scale-invariant and, in practice, remain unchanged by changes in viewing angle or size. Therefore, by identifying these feature points, the channel in which the image is located can be determined.

[0121] In practice, each channel has multiple feature points, and the similarity of these feature points is high, so it is actually impossible to distinguish channels based on one or two feature points. Therefore, it is necessary to use these feature points as a set to distinguish channels.

[0122] Specifically: the channel fingerprint establishment unit includes: feature semantic classifier, feature optimization filter, feature point allocator, and spatial fingerprint compiler.

[0123] The feature semantic classifier is used to distinguish feature points, or in other words, to classify them. In fact, in addition to the feature points themselves, their three-dimensional spatial positions are also used to determine their location. Classifying feature points effectively groups closely related feature points into one category, reducing the number of feature points and thus achieving data dimensionality reduction.

[0124] Specifically: Feature semantic classifier, obtains feature points in all feature point sets to count the types of feature points.

[0125] The types of feature points in the feature semantic classifier are assigned as follows:

[0126] Z1: A reference coordinate system for predefined standard eigenvectors in the eigenvector space.

[0127] The feature points in this scheme have three dimensions, so the feature space is also three-dimensional, and the standard feature vector is a specific point in the three-dimensional space, such as (0, 0, 0).

[0128] Z2: Calculate the cosine similarity between the vector of the feature point to be classified and the standard feature vector, and establish the relative position relationship in the feature space.

[0129] Because the feature points extracted by calculating the cosine similarity between feature points mapped into 3D space and the standard feature vector are 3D information, the feature points can be mapped into a pre-set 3D space. The cosine similarity values ​​can then be used to determine the degree of similarity between the feature points and the standard feature vector. Spatially partitioning the feature vectors according to a preset similarity threshold allows for deterministic classification of feature points. In Z2, the cosine similarity values ​​between the feature to be classified and the standard feature vector are already calculated, so feature point classification can be directly based on this value—for example, feature points with a cosine similarity value between 0 and 0.1 are assigned to category 1. This approach allows for classification of feature points across all channels. While this process mitigates the inherent variability of the feature points, it also reduces the difficulty of identification. Ultimately, the positional relationships between the feature points determine their channel affiliation.

[0130] The feature optimization filter dynamically removes redundant feature points whose feature category distribution entropy falls below a preset threshold. In the feature semantic classifier, all feature points are classified, but some feature points are rare or mostly concentrated within a specific, narrow range, so these feature points are removed. Therefore, the entropy value includes both the number of feature points and their distribution range, where the distribution range can be described by a probability density function. The specific calculation method is not detailed here.

[0131] The feature point allocator assigns a number of different feature points to each channel. This helps reduce the similarity between channels. For example, if there are 10 channels and 20 types of feature points, each channel can be assigned two types of feature points. This allows for channel identification based on the type of feature points. This reduces computational time during subsequent matching.

[0132] Specifically:

[0133] The way to assign feature point types to each channel in the feature point allocator is as follows:

[0134] Step 1: Pre-set feature requirements, which include feature point cardinality threshold and feature point distribution density threshold.

[0135] The feature point cardinality threshold is the minimum number of feature points, and the feature point distribution density threshold is the number of feature points per square.

[0136] Step 2: Based on the feature requirements, obtain the allocation scheme of all feature point types of the corresponding channel, and each allocation scheme includes at least two types of feature points.

[0137] Feature requirements essentially mean that the feature points selected in a channel must meet a minimum feature point cardinality threshold and a feature point distribution density threshold. For example, if a certain type of feature point is selected as a channel feature, then the feature points of that type must meet the feature requirements in that channel. This ensures that the feature points of that type are capable of describing the channel.

[0138] Step 3: Obtain the distribution plan of the feature point types corresponding to all channels and establish a feature point distribution matrix. Each row in the feature point distribution matrix represents a feature point that can be selected by a channel, and each column in the feature point distribution matrix represents a feature point. The elements in the feature point distribution matrix are 0 or 1. When it is 1, it indicates that the feature point meets the feature requirements of the corresponding channel. When it is 0, it indicates that the feature point does not meet the feature requirements of the corresponding channel.

[0139] Step 4: Randomly select at least two feature points of different types from each row of the feature point assignment matrix to assign the type of feature points to each channel.

[0140] like Figure 3 As shown, Figure 3 An 8×4 feature point allocation matrix is ​​given in , in which two elements are selected in each row to form a feature point type allocation scheme.

[0141] In this way, the feature point allocator can allocate at least two feature points to each channel, and the channel can be described by these two feature points.

[0142] The specific description of the channel is as follows:

[0143] The spatial fingerprint compiler extracts the feature points of the corresponding category for each channel, calculates the distance between each feature point, and generates the spatial fingerprint feature information corresponding to the channel.

[0144] During the feature point dimensionality reduction process, the amount of geometric information carried by a single feature point will inevitably decay, resulting in limited independent representation capabilities. To break through the dimensionality constraint, this solution innovatively introduces the topological relationship between feature points as incremental information for spatial feature fingerprints, specifically including the following two dimensions:

[0145] 1) Relative distance ratio: By quantifying the distance ratio of each pair of feature points (such as the ratio of the distance between B1-B2 and B2-B3), a scale-invariant geometric constraint is constructed.

[0146] 2) Spatial orientation coding: The polar coordinate system is used to describe the relative orientation relationship between feature points, forming a topological descriptor containing angle information.

[0147] Taking the trifocal system as an example, when the three feature points B1, B2, and B3 are detected, the system not only records the coordinates of each point but also extracts the B1B2 / B2B3 spacing ratio coefficient and the B1-B2-B3 triangular configuration orientation parameters. This design has two technical advantages: first, it breaks through the distance measurement limitations of traditional monocular vision systems and achieves virtual depth perception through geometric proportional relationships; second, during the feature matching stage, feature comparison can be completed directly based on 2D image data, without relying on binocular vision systems or depth perception equipment such as LiDAR, significantly reducing system hardware complexity.

[0148] By adopting the above solution, a unique spatial fingerprint feature information can be generated for all channels. During subsequent positioning, the current position can be determined by simply identifying the spatial fingerprint feature information.

[0149] The intelligent channel planning module is configured to receive the destination coordinates submitted by the user terminal and calculate the optimal route within the three-dimensional navigation network. The optimal communication route generation scheme is prior art and will not be elaborated here. The key to this application lies in how to use image information for positioning.

[0150] The specific positioning solution is undertaken by the visual positioning and navigation module, specifically:

[0151] The visual positioning and navigation module is configured to capture the visual data stream of the front scene in real time through the camera of the user terminal, use the feature point matching algorithm to spatially align the real-time visual data with the three-dimensional point cloud model, and output navigation instructions including the direction deflection angle and displacement distance.

[0152] The visual positioning navigation module includes a channel information acquisition unit, a visual information acquisition unit, a position positioning unit, and a navigation prompt unit. Among them, the channel information acquisition unit obtains a set of channels that the user can pass through based on the user's historical location and the optimal route.

[0153] The channel set is actually the channel that the user is most likely to pass through. In practice, the channel set is the channel obtained by expanding the user's last calculated location. Of course, in practice, the channel set can also be ranked by the distance between each channel and the last location, with the closer to the last location, the higher the ranking.

[0154] The visual information acquisition unit is used to acquire image information of the scene where the user is located, extract matching extreme points from the image information, and calculate the distance between each matching extreme point to generate matching information.

[0155] The matching extreme points are actually feature points. The method of obtaining matching extreme points is the same as that of obtaining feature points, so it will not be repeated here. The matching information includes the relative distance ratio relationship and spatial orientation encoding.

[0156] The position positioning unit compares the matching information with the spatial fingerprint feature information of the channels in the channel set, obtains the spatial fingerprint feature information of the channel with the highest matching degree with the matching information, and uses the channel as the user's location.

[0157] The location unit compares the matching information with the channels in the channel set. If the contrast exceeds a preset threshold, it is used as the user's location. The navigation prompt unit sends navigation instructions to the user based on the user's location and the optimal route.

[0158] Once the user's location and desired route are known, navigation instructions can be sent to the user. The navigation instructions can be voice notifications or VR-enabled guidance information.

[0159] Example 2: Example 2 provides a position positioning unit based on Example 1, which can measure the impact of environmental updates on path matching in the building space. Specifically:

[0160] The location unit includes a weight setter, a matching calculator, and a comprehensive calculator. The weight setter assigns a confidence weight to each feature point based on its location and update time. The higher the height of the feature point from the floor, the higher the confidence weight. The closer the feature point was to the update time, the higher the confidence weight.

[0161] The confidence weight actually describes the probability of a feature point changing. Generally speaking, feature points closer to the ground are more susceptible to environmental interference (for example, dust covering the ground can cause feature points to change), while feature points higher above the ground are less susceptible to human influence and therefore have a lower probability of changing. Similarly, if certain feature points were recently entered and updated, their status is less likely to change, so the system will assign them a higher confidence weight.

[0162] The matching calculator calculates the cosine similarity between the matching information and each spatial fingerprint feature information to obtain initial comparison information. The calculation method of the initial comparison information is the same as the comparison method in Example 1. In Example 1, the comparison result is directly used as the judgment basis, but in this embodiment, the result is calculated.

[0163] The comprehensive calculator multiplies the initial comparison information with the confidence weight of each feature point to obtain the ranking information, and takes the channel with the highest ranking information as the user's location.

[0164] In other words, the comprehensive calculator weights the matching information and the values ​​calculated for each feature point in the spatial feature fingerprint information, and then uses the weighted sum as sorting information, which is then used as the final screening method. This screening method is more accurate in practice and can avoid the impact of manual cleaning and sweeping on channel identification.

[0165] Example 3: Example 2 provides an environment feature update module based on Example 1. The environment update module can update the three-dimensional point cloud model according to the image information entered by the user during the navigation process.

[0166] Specifically: The building space three-dimensional navigation system based on big data matching also includes an environmental feature update module to obtain image information collected along the way to accurately reach the destination, and update the image information to a three-dimensional point cloud model.

[0167] The update steps are as follows:

[0168] Step 1: Obtain the user's movement path from the starting point to the end point and obtain the corresponding channel;

[0169] Step 2: Get the image information when passing through each channel;

[0170] Step 3: Extract feature points from the image information and update the feature points to the spatial feature fingerprint information of the corresponding channel.

[0171] When the user is moving, he will continuously collect the position information of each channel, so it is only necessary to update the collected feature points to the spatial feature fingerprint information.

[0172] For example, the spatial feature fingerprint information of a channel includes 5 feature points. One of the 5 feature points has changed and cannot be queried during the last navigation. Then, the newly recorded feature point of the same type can be added to the spatial feature fingerprint information of the channel.

[0173] Example 4: A three-dimensional building space navigation method based on big data matching, using the three-dimensional building space navigation system based on big data matching described in Example 1 to perform three-dimensional navigation.

[0174] The above are merely preferred embodiments of the present application and are not intended to limit the present application. Those skilled in the art will readily appreciate that various modifications and variations are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

Claims

1. A building space three-dimensional navigation system based on big data matching, characterized in that: include: A three-dimensional scene construction module is configured to collect three-dimensional structural data of a target area through a building information model interface, fuse spatial topological relationships to generate a three-dimensional navigation network, wherein the three-dimensional navigation network includes vector relationships between vertical associated channels and horizontal passageways of each floor; An environmental feature extraction module collects image information of the three-dimensional navigation network and generates a three-dimensional point cloud model through a stereo vision algorithm; an intelligent channel planning module configured to receive the destination coordinates submitted by the user terminal and calculate an optimal route in the three-dimensional navigation network; a visual positioning and navigation module configured to capture a visual data stream of a front scene in real time through a camera of a user terminal, spatially align the real-time visual data with the three-dimensional point cloud model using a feature point matching algorithm, and output a navigation instruction including a direction deflection angle and a displacement distance; The channel information extraction unit performs non-overlapping channel segmentation on the three-dimensional navigation network and generates a channel vector dataset with spatial independence; 3D information input unit, which reconstructs the 3D point cloud model of each channel based on multi-view video stream; A key feature extraction unit is configured to extract a feature point set with rotation invariance from a three-dimensional point cloud model; A channel fingerprint establishment unit obtains a feature point set of each channel and constructs spatial fingerprint feature information containing a unique channel identifier based on the feature point differences of all feature point sets, wherein the spatial fingerprint feature information of each channel is composed of multi-dimensional spatial geodesic distance information between feature points; The visual positioning and navigation module includes: A channel information acquisition unit, which acquires a set of channels that the user can pass through based on the user's historical location and optimal travel route; A visual information acquisition unit is used to acquire image information of the scene where the user is located, extract matching extreme value points from the image information, and calculate the distance between each matching extreme value point to generate matching information; The position positioning unit compares the matching information with the spatial fingerprint feature information of the channels in the channel set, obtains the spatial fingerprint feature information of the channel with the highest matching degree with the matching information, and uses the channel as the user's location; The navigation prompt unit sends navigation instructions to the user based on the user's location and the optimal route.

2. The building space three-dimensional navigation system based on big data matching according to claim 1 is characterized in that: The channel fingerprint establishment unit includes: Feature semantic classifier, obtains the feature points in all feature point sets to count the types of feature points; Feature optimization filter, dynamically removes redundant feature points whose feature category distribution entropy value is lower than the preset threshold; Feature point allocator, which assigns several types of feature points to each channel. The types of feature points assigned to each channel are different. The spatial fingerprint compiler extracts the feature points of the corresponding category for each channel, calculates the distance between each feature point, and generates the spatial fingerprint feature information corresponding to the channel.

3. The building space three-dimensional navigation system based on big data matching according to claim 1 is characterized in that: The types of feature points in the feature semantic classifier are assigned as follows: A reference coordinate system for predefined standard eigenvectors in the eigenvector space; Calculate the cosine similarity value between the vector of the feature point to be classified and the standard feature vector, and establish the relative position relationship in the feature space; The feature vectors are spatially divided according to the preset similarity threshold interval to achieve deterministic attribution of feature point types.

4. The building space three-dimensional navigation system based on big data matching according to claim 2 is characterized in that: The way to assign feature point types to each channel in the feature point allocator is as follows: Pre-set feature requirements, feature requirements Including feature point cardinality threshold and feature point distribution density threshold; Based on the feature requirements, an allocation scheme for all feature point types of the corresponding channel is obtained, and each allocation scheme includes at least two types of feature points; Obtain the distribution scheme of the feature point types corresponding to all channels and establish a feature point distribution matrix. Each row in the feature point distribution matrix represents the feature points that can be selected by a channel, and each column in the feature point distribution matrix represents a feature point. The elements in the feature point distribution matrix are 0 or 1. When it is 1, it indicates that the feature point meets the feature requirements of the corresponding channel. When it is 0, it indicates that the feature point does not meet the feature requirements of the corresponding channel. At least two feature points of different categories are randomly selected from each row of the feature point allocation matrix to assign the category of the feature points to each channel.

5. The building space three-dimensional navigation system based on big data matching according to claim 1 is characterized in that: The position positioning unit includes: The weight setter sets the confidence weight for each feature point based on its location and update time. The higher the height of the feature point from the floor, the higher the confidence weight; the closer the update time of the feature point, the higher the confidence weight. A matching calculator calculates the cosine similarity between the matching information and each spatial fingerprint feature information to obtain initial comparison information; The comprehensive calculator multiplies the initial comparison information with the confidence weight of each feature point to obtain the ranking information, and takes the channel with the highest ranking information as the user's location.

6. The building space three-dimensional navigation system based on big data matching according to claim 1 is characterized in that: It also includes an environmental feature update module to obtain image information collected along the way by users who have accurately reached their destination, and update the image information to a three-dimensional point cloud model.

7. The building space three-dimensional navigation system based on big data matching according to claim 1 is characterized in that: The feature points are extracted as follows: S1: Convert the 3D point cloud model into a 3D voxel model; S2: Construct a three-dimensional scale space to generate feature response information at different observation scales; S3: For each voxel, check whether the feature response information is the maximum value in the 3D neighborhood. If so, set it as the extreme point; the neighborhood includes the 26 spatial neighbors of the current voxel and the corresponding positions of the upper and lower adjacent scales; S4: Assign a gradient magnitude and direction to each feature point, and use the gradient magnitude and direction as the descriptor of the feature point.

8. A three-dimensional navigation method for building space based on big data matching, characterized in that: Three-dimensional navigation is performed using the building space three-dimensional navigation system based on big data matching as described in any one of claims 1 to 7.

Images