A relocalization method based on point cloud and UWB

Abstract

The present invention relates to the field of repositioning technology, specifically disclosing a point cloud and UWB-based repositioning method, comprising the following steps: setting a number of angle combinations, maintaining the depth camera's posture at the center point of a three-dimensional grid to satisfy any angle combination, and collecting point cloud data; obtaining the AR device's 6DoF pose in real time and providing a confidence level; when the confidence level is lower than a preset threshold for five consecutive times, stopping trusting the 6DoF pose provided by the visual SLAM module; obtaining the AR device's three-dimensional spatial coordinates based on UWB, obtaining a point cloud to be matched, and a comparison point cloud; obtaining the degree of matching, and obtaining the AR device's current 6DoF pose based on the maximum degree of matching. The present invention integrates UWB and point clouds in an occluded environment to quickly restore AR precision positioning.

Description

Technical Field

[0001] The present invention relates to the field of repositioning technology, and in particular to a repositioning method based on point cloud and UWB. Background Art

[0002] In augmented reality (AR) systems, relocalization refers to recalculating the camera's 6DoF pose using an established environment map or previously captured reference frames when the device is lost or misaligned during tracking, so that virtual content can be realigned with the real scene.

[0003] A vision-based point cloud relocalization solution is to generate a high-precision three-dimensional point cloud map of the target environment using laser scanning or depth cameras in advance, and then match the images captured by the camera in real time with the visual features in the point cloud map during runtime to calculate the current camera's 6DoF pose.

[0004] However, in large-scale AR projects, when there is a surge in foot traffic, a large number of pedestrians standing or moving in front of the camera's field of view can block key static features such as walls, pillars, and the ground. This prevents the AR glasses from recording the inherent contours and texture information of the environment, and from extracting enough key feature points for map matching. At this point, the visual relocalization algorithm will be unable to return the correct 6DoF pose due to matching failure or low confidence, causing the position of virtual objects to drift and offset. AR content cannot maintain a stable connection with the real scene, and users will see virtual elements floating and misaligned, severely damaging immersion and usability, ultimately giving users a very poor user experience. Summary of the Invention

[0005] The purpose of the present invention is to provide a relocation method based on point cloud and UWB to solve the above technical problems.

[0006] The purpose of the present invention can be achieved through the following technical solutions:

[0007] A point cloud and UWB-based relocalization method includes the following steps:

[0008] Construct a 3D model of the indoor space and divide the 3D model into several 3D grids. The length, width, and height of the 3D grids are all preset values. Set several angle combinations and keep the depth camera's posture at the center point of the 3D grid to meet any angle combination to collect point cloud data.

[0009] The main camera on the AR device is used to capture a depth image of the indoor space. The visual SLAM module is used to obtain the 6DoF pose of the AR device and provide a confidence level. If the confidence level falls below a preset threshold for five consecutive times, the 6DoF pose provided by the visual SLAM module is no longer trusted.

[0010] Acquire the 3D spatial coordinates of the AR device based on UWB, use the point cloud data corresponding to the three-dimensional grid where the 3D spatial coordinates are located as the point cloud to be matched, obtain the depth image captured by the current main camera, and extract the point cloud data as the comparison point cloud;

[0011] Obtain the degree of match between the to-be-matched point cloud and the comparison point cloud, use the to-be-matched point cloud corresponding to the maximum matching degree C as the target point cloud, obtain the angle combination A corresponding to the target point cloud, and obtain the current 6DoF pose of the AR device based on the 3D space coordinates and the angle combination A.

[0012] As a further solution of the present invention: setting the angle combination includes:

[0013] Set the pitch angle range, yaw angle range and roll angle range;

[0014] Starting from the starting points of the pitch angle range, yaw angle range, and roll angle range, a number of pitch angle nodes, yaw angle nodes, and roll angle nodes are set at preset angle intervals;

[0015] A single angle combination includes a pitch angle node, a yaw angle node, and a roll angle node, and different angle combinations include different pitch angle nodes and / or yaw angle nodes and / or roll angle nodes.

[0016] As a further solution of the present invention, obtaining the three-dimensional space coordinates of the AR device based on UWB includes:

[0017] Install n UWB base stations at preset locations in the indoor space, where n is a preset number and n>6;

[0018] The three-dimensional spatial coordinates of the AR device are obtained through multi-lateral positioning based on UWB two-way ranging.

[0019] As a further solution of the present invention, obtaining the matching degree between the to-be-matched data and the comparison data includes:

[0020] The point cloud to be matched is converted into the world coordinate system to obtain a first point cloud, and the minimum circumscribed rectangle of the first point cloud is divided into voxel grids to obtain a plurality of sub-grids, which are recorded as first sub-grids and numbered;

[0021] Acquire a second subgrid based on the compared point cloud and number the second subgrid;

[0022] The matching degree is calculated based on the classification labels of the first sub-grid and the second sub-grid with the same number.

[0023] As a further solution of the present invention, obtaining the matching degree between the to-be-matched data and the comparison data further includes:

[0024] Set classification labels, including occupied and idle;

[0025] When the first subgrid contains a single spatial point in the first point cloud, occupied is used as the classification label of the first subgrid; otherwise, free is used as the classification label of the first subgrid;

[0026] Get the classification label of the second subgrid;

[0027] Count the total number of occupied numbers with the same number and classification label B1, count the total number of numbers with the same number but different classification labels B2, and calculate the matching degree P=B1 / B2.

[0028] As a further solution of the present invention: when the maximum matching degree is less than a preset matching degree threshold, the following steps are performed:

[0029] Start the auxiliary camera on the AR device and keep the auxiliary camera's posture to meet any angle combination to collect a depth image of the indoor space, which is recorded as the first image. The point cloud data in the single first image is fused with the comparison point cloud to obtain a new comparison point cloud.

[0030] Get the current 6DoF pose of the AR device based on the new comparison point cloud.

[0031] As a further solution of the present invention: before acquiring the first image, the method further includes:

[0032] The first subgrid and the second subgrid corresponding to the maximum matching degree C are marked as the first grid and the second grid respectively. When the classification label of the first grid numbered i is occupied and the classification label of the second grid numbered i is free, the second grid numbered i is marked as the target grid;

[0033] Cluster the target grids to obtain clusters. Starting from the location of the auxiliary camera, draw a ray to the indoor space. The ray satisfies the angle combination a. If the intersection of the ray and the indoor space is not within the minimum circumscribed sphere of the cluster, the first image corresponding to the angle combination a is not obtained.

[0034] As a further solution of the present invention, obtaining the current 6DoF pose of the AR device based on the three-dimensional space coordinate and angle combination A includes:

[0035] The current 6DoF pose of the AR device is (x1, y1, z1, r1, r2, r3), where x1, y1, and z1 represent the x-axis, y-axis, and z-axis coordinates of the AR device, respectively. r1, r2, and r3 represent the pitch, yaw, and roll angles corresponding to the pitch, yaw, and roll nodes in angle combination A, respectively.

[0036] The beneficial effects of the present invention are as follows:

[0037] 1) This invention couples UWB 3D coordinates with a gridded point cloud library and pre-samples point clouds at the center of each 3D grid through offline multi-angle combinations. This allows the system to instantly narrow the search range based on UWB coarse positioning during runtime, eliminating the overhead of a global search across the entire large-scale map. When the visual SLAM confidence level drops sharply, the system can immediately switch to "UWB + point cloud matching" mode, quickly locating the best candidate point cloud through grid indexing and calculating the 6DoF pose. This maintains the continuity and stability of virtual-real alignment even in scenes with dense crowds and missing textures, significantly improving the reliability and smoothness of the AR experience.

[0038] 2) The similarity calculation between the point cloud to be matched and the comparison point cloud uses the voxel grid occupancy labeling method, which converts complex geometric comparisons into "occupied / free" binary classifications and performs statistics through number alignment, improving robustness to local occlusion, noise, and dynamic targets. This method not only avoids the problem that traditional feature point matching is prone to failure under conditions of scarce texture or changing lighting, but also obtains a precise degree of matching through simple counting ratios, and quickly obtains stable results without deep learning reasoning or high-dimensional feature measurement. This ensures that a usable 6DoF pose can still be output in large spaces with noisy and complex lighting, maintaining the constant position of virtual content and avoiding experience defects such as "floating" and "jittering";

[0039] 3) When the maximum matching degree is insufficient, the present invention does not blindly traverse all angle combinations. Instead, it first clusters the areas where the first subgrid is occupied and the second subgrid with the same number is free. It then guides the auxiliary camera to recapture the depth map only for the clusters with missing point clouds. If the intersection of the ray emitted from the camera and matching angle combination a with the indoor model does not fall within the minimum circumscribed sphere of the cluster, the angle combination is skipped to avoid invalid acquisition. This minimizes the number of acquisition frames and the amount of computation, quickly fills in key geometric information, fuses and refreshes the comparison point clouds in real time, and quickly improves the matching degree to an acceptable range. This targeted incremental update strategy effectively reduces redundant scanning, power consumption, and user waiting time, achieving efficient and seamless positioning self-recovery and steady-state tracking. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The present invention will be further described below with reference to the accompanying drawings.

[0041] Figure 1 It is a flow chart of a point cloud and UWB-based repositioning method of the present invention. DETAILED DESCRIPTION

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0043] See also Figure 1 As shown, the present invention is a relocation method based on point cloud and UWB, comprising the following steps:

[0044] In an open space, a laser scanner or depth camera is first used to slowly move along the main passages and corners of the room, while the inertial navigation odometer is used to synchronously record the position and posture. A complete three-dimensional mesh surface model is generated in a short time using a real-time mapping algorithm. The overall bounding box size is then calculated and the model is evenly divided in three axes according to the preset three-dimensional grid side length, such as one meter. Uniquely numbered mesh bodies are generated in sequence. The center of mass coordinates of each mesh body are extracted, and the depth camera is moved to the center of mass of the current mesh body and rotated to one of the poses on site using a tripod or a robotic arm in combination with the angle combination. The camera is triggered to continuously capture multiple frames of depth maps. The original depth values ​​are subjected to distortion correction, depth filtering, and external parameter calibration, and then converted to a unified world coordinate system in real time and stored as a dense point cloud. All mesh bodies and all angle combinations are repeatedly traversed, and an offline point cloud library indexed by mesh number and pose combination is finally formed in the database.

[0045] It is understandable that the use of grid-by-grid sampling combined with multi-angle poses can ensure that the same spatial position retains high-quality geometric information from different perspectives. At runtime, small-scale point cloud comparison can be completed by simply determining the grid through UWB coarse positioning and calling the corresponding angle sub-library. This not only avoids the computational and storage waste caused by global search, but also reduces the probability of matching failures caused by feature loss when affected by occlusion. This fundamentally lays the foundation for the subsequent rapid recovery of accurate six-degree-of-freedom poses in complex scenes such as densely populated and sparsely textured scenes.

[0046] In a preferred embodiment of the present invention, the angle combination is set to include:

[0047] Engineers determine the available range of pitch angles based on the site height and the user's mirror-holding posture, for example, from negative thirty degrees to thirty degrees, and then determine the available range of yaw angles based on the indoor layout and traditional navigation orientation, for example, from zero degrees to three hundred and sixty degrees. Finally, the available range of roll angles is determined based on the ergonomic comfort limit of the head-mounted device, for example, from negative fifteen degrees to fifteen degrees. These three ranges are then discretized at preset angle intervals. If the interval is set to ten degrees, a node is generated for the pitch angle at negative thirty degrees, negative twenty degrees, and up to thirty degrees, and nodes are generated for the yaw angle at zero degrees, one hundred and eighty degrees, and other positions, respectively. The roll angle generates nodes at negative fifteen degrees, negative five degrees, and other positions. The system traverses all nodes through a triple loop, concatenates "pitch node i, yaw node j, roll node k" into an angle combination and assigns a unique number. It ensures that a new combination is generated only when at least one angle variable is different, until all legal postures are listed and written into the angle lookup table for subsequent grid sampling.

[0048] It is important to note that by binning and discretizing the three-axis angles, we can pre-cover almost all possible camera orientations during the user's indoor movement and natural head turning, avoiding the situation where the target point cloud lacks corresponding viewpoint data during runtime and causes matching failure. At the same time, the triple loop enumeration method exhausts all available combinations at once in the offline stage, and then uses the numbering for fast indexing, thereby reducing real-time addressing overhead and ensuring that the positioning algorithm can quickly find the offline point cloud closest to the current camera pose in complex occlusion scenarios, achieving robust pose solution and virtual-real alignment.

[0049] The depth working mode of the main camera of the headset is called to allow the hardware to continuously output synchronized grayscale frames and depth frames. After the system reads this data, it is handed over to the visual SLAM pipeline to perform feature extraction, inter-frame matching, posture backend optimization and local map update processes to obtain the six-degree-of-freedom pose of the current camera in the world coordinate system; SLAM also calculates a confidence index based on the number of tracked valid feature points, the mean of the reprojection residual and the loop detection status. The program writes this index into a circular queue of length five. After each frame is processed, the queue content is refreshed and checked to see if all the contents in the queue are below the threshold. If five consecutive frames do not meet the standard, the lost flag is immediately set, and the SLAM output pose is marked as unreliable. Subsequently, the incremental update of the map library is suspended and the push of the pose to the back-end navigation module is stopped. At the same time, a degradation signal is sent to the fusion positioning module to switch to the UWB plus point cloud matching process.

[0050] It is worth noting that this can avoid stubbornly using the distorted posture given by visual SLAM when the camera's field of view is blocked by pedestrians or the lighting changes drastically, thereby preventing virtual objects from drifting or jumping in space. The confidence sliding window determines the loss state and can immediately detect the continuous decline in tracking quality and trigger the backup positioning solution, allowing the system to maintain reliable spatial anchoring capabilities in complex dynamic scenes. This leaves sufficient reaction time for the subsequent relocalization algorithm that integrates UWB and offline point clouds, and overall ensures the stable alignment of AR content with the real environment.

[0051] The 3D spatial coordinates of the AR device are obtained through UWB. Based on this, the corresponding grid center and the point cloud obtained by multi-angle combination sampling are loaded from the offline database at one time as the point cloud to be matched. At the same time, the main camera captures a frame of depth image, and its depth value is back-projected using the intrinsic reference to generate a dense point cloud in the camera coordinate system. This point cloud is then radius-filtered and outliers are removed before being directly saved as the comparison point cloud.

[0052] It is important to note that the reason for using UWB to calculate coordinates first and then reading the offline point cloud by grid index is that UWB ranging is almost unaffected by pedestrian occlusion and lighting changes. It can still provide a stable global position when visual tracking fails, thereby limiting the search range to a small grid and significantly reducing matching time. Converting the current depth frame into a comparison point cloud can fully preserve the geometric details of the scene, which is used for voxel-level occupancy comparison with the point cloud to be matched. Ultimately, it helps the algorithm quickly find the reference perspective that is most consistent with the device posture in dynamic and crowded environments, laying a solid foundation for the subsequent output of accurate six-degree-of-freedom posture.

[0053] In another preferred embodiment of the present invention, obtaining the three-dimensional space coordinates of the AR device based on UWB includes:

[0054] During the site survey phase, technicians arrange six or more UWB base stations at equal heights on the ceiling or columns based on the room's shape. They use laser rangefinders and goniometers to perform high-precision measurements of each base station's three-dimensional coordinates and write them into the system configuration file. Then, network cables or wireless synchronization modules are used to align the clocks of all base stations at the nanosecond level. The UWB tags embedded in the AR glasses form a two-way ranging link with the base stations. During operation, the tags transmit pulses to each base station in turn according to the TDMA time slot. Upon receiving the signal, the base stations immediately send back response frames. The system records the round-trip time and subtracts the hardware delay to obtain the distance between each base station and the tag. When distance data comes from at least four base stations simultaneously, the multilateral positioning algorithm uses the prior base station coordinates and solves the tag's precise position in the global coordinate system through nonlinear least squares. For example, the calculated value is (2.36, 5.42, 1.68). This result is written to the shared memory in real time for subsequent grid indexing and fusion positioning modules to call.

[0055] It is understandable that such deployment and use of UWB can ensure a stable global position source in the absence of direct light, pedestrian occlusion, and scarce textures, and the three-dimensional coordinates are directly in the world coordinate system consistent with the offline point cloud, reducing the complexity of subsequent coordinate conversion; through base station clock synchronization and two-way ranging, the one-way propagation delay error is eliminated, so that the multi-lateral positioning output accuracy meets the requirements of the same order of magnitude as the offline grid side length, providing key support for rapid retrieval of the point cloud to be matched, reducing the matching search space, and improving the overall re-positioning speed and reliability, realizing positioning degradation and rapid recovery when visual failure occurs;

[0056] Obtain the degree of match between the to-be-matched point cloud and the comparison point cloud. Use the to-be-matched point cloud corresponding to the maximum matching degree C as the target point cloud. Obtain the angle combination A corresponding to the target point cloud. Determine the current 6DoF pose of the AR device based on the 3D spatial coordinates and angle combination A.

[0057] In a specific embodiment, obtaining the matching degree between the to-be-matched data and the comparison data includes:

[0058] After reading the offline point cloud to be matched, a four-dimensional homogeneous transformation is performed on each three-dimensional coordinate according to the external parameter matrix recorded when it was saved, and unified to the world coordinate system. Then, the extreme values ​​of all points are calculated to obtain the axis-aligned bounding box. According to the set voxel side length, for example, five centimeters, it is divided in three axes and numbered to obtain the three-dimensional grid index. The (i, j, k) coordinates of each voxel are mapped to a one-dimensional integer using a three-dimensional array or hash table to achieve fast access. Then, for each voxel in the bounding box, it is determined whether it contains at least one point. If so, an "occupied" mark is written in the first sub-grid table. Otherwise, an "occupied" mark is written. The bounding box of the comparison point cloud collected in real time can be directly obtained without converting the coordinate system. The second sub-grid table is generated according to the same origin, voxel size and numbering strategy as the first point cloud and filled with occupied or idle labels. The algorithm traverses all voxel numbers that appear in the two tables. When a number is occupied in both tables, the B1 count is increased by one. When the labels in the two tables are inconsistent, the B2 count is increased by one. After the traversal is completed, the matching degree is calculated according to the formula P=B1 / B2, and the current maximum P value and its index are maintained in the candidate set for subsequent pose output.

[0059] It should be noted that the method of coordinate unification plus voxel occupancy ratio can simplify point clouds of arbitrary resolution and density into binary grids with the same structure, so that the comparison process does not rely on fine point-level alignment but only focuses on voxel-level overlap, which not only reduces the computational complexity but also enhances the robustness to noise and local occlusion; the matching degree is defined as the ratio of the number of simultaneously occupied voxels to the number of conflicting voxels. Its principle is similar to the intersection ratio in set theory. The more intersections and the fewer differences, the larger P is, which can intuitively reflect the degree of consistency between the two sets of geometries in space. At the same time, using B2 as the denominator instead of a simple intersection-union ratio can offset the visible area differences caused by changes in viewing angles (B2 only counts the voxels that actually cause contradictions: one side is occupied and the other side is idle; both sides are idle or occupied and will not enter the denominator. In this way, attention is paid to The focus is on the "difference" itself, while excluding "common defaults." If the change in perspective simply causes the camera to miss a piece of wall, that area will be marked as idle in both the offline and real-time point clouds and will not fall into B2. What actually increases B2 is when the online point cloud says that the voxel has a point while the offline point cloud says it doesn't (or vice versa), which represents inconsistency in geometric information. Therefore, the level of the B1 / B2 metric is mainly controlled by the balance between "overlapping occupancy" and "conflicting occupancy," and will not be diluted by a large number of "empty-empty" voxels caused by field of view cropping or occlusion. This allows for a fairer evaluation of the geometric fit of the two point clouds within the observed area, and is more resistant to differences in visible range caused by different perspectives, providing a reliable basis for the subsequent direct assembly of the six-degree-of-freedom pose by combining UWB coordinates.

[0060] It should be noted that the method for obtaining the second subgrid is similar to that for the first subgrid, and the method for obtaining the classification label of the second subgrid is also similar to that for the first subgrid, so it will not be repeated here.

[0061] It should be noted that the current 6DoF pose of the AR device based on the 3D space coordinate and angle combination A includes:

[0062] The current 6DoF pose of the AR device is (x1, y1, z1, r1, r2, r3), where x1, y1, and z1 represent the x-axis, y-axis, and z-axis coordinates of the AR device, respectively. r1, r2, and r3 represent the pitch, yaw, and roll angles corresponding to the pitch, yaw, and roll nodes in angle combination A, respectively.

[0063] In a specific embodiment, when the maximum matching degree is less than a preset matching degree threshold, the following steps are performed:

[0064] After detecting that the maximum matching degree is lower than the threshold, the system first reads the offline first sub-grid table and the real-time second sub-grid table that have just been compared, and compares the numbers of the two one by one. All numbers marked as occupied offline and free in real time are added to the target grid set; the program sends the centroid coordinates of these target grids to DBSCAN for clustering, obtains several clusters and calculates the center and radius of the minimum circumscribed sphere for each cluster; then, with the current world coordinates of the auxiliary camera as the starting point of the ray, the unused angle combination (a represents any one) is taken out from the angle lookup table in sequence, and the unit direction vector is obtained based on the pitch, yaw, and roll angles and the 3D indoor ray is obtained. Ray casting is performed in the model. If the closest intersection of the ray and the geometric body falls inside the circumscribed sphere of any cluster, it means that the line of sight can cover the missing area. The system immediately locks the posture and opens the shutter to capture a frame of depth image. The depth frame is converted into a point cloud after distortion correction, bilateral filtering and voxel downsampling. The new point cloud is hashed with the original comparison point cloud using the voxel center as the key. If the same key already has data, the average depth is taken. Otherwise, it is directly inserted to generate an updated comparison point cloud and refresh the occupancy flag table. Finally, the voxel occupancy matching algorithm is called again to calculate the new matching degree and the best offline point cloud angle combination, and the new six-degree-of-freedom pose is output to the rendering thread.

[0065] It is worth noting that by first finding the locations where the offline and real-time grid labels are inconsistent, the missing geometric information of the current perspective can be accurately located to avoid blind rescanning; using density clustering to merge discrete target grids into several clusters, and then using the intersection of rays and cluster circumscribed spheres to determine the necessity of shooting, only the perspectives that really need to be supplemented can be collected without wasting time and energy; incrementally merging the new point cloud voxel level into the comparison point cloud and then rematching it immediately, the system can quickly improve the matching degree and restore reliable positioning with minimal data collection and computing costs, thereby ensuring that the AR content continues to be stably bound to the real environment and maintain the user's immersive experience.

[0066] The above is a detailed description of an embodiment of the present invention. However, the content is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A relocation method based on point cloud and UWB, characterized in that: The following steps are involved: Construct a 3D model of the indoor space and divide the 3D model into several 3D grids. The length, width, and height of the 3D grids are all preset values. Set several angle combinations and keep the depth camera's posture at the center point of the 3D grid to meet any angle combination to collect point cloud data. The main camera on the AR device is used to capture a depth image of the indoor space. The visual SLAM module is used to obtain the 6DoF pose of the AR device and provide a confidence level. If the confidence level falls below a preset threshold for five consecutive times, the 6DoF pose provided by the visual SLAM module is no longer trusted. Acquire the 3D spatial coordinates of the AR device based on UWB, use the point cloud data corresponding to the three-dimensional grid where the 3D spatial coordinates are located as the point cloud to be matched, obtain the depth image captured by the current main camera, and extract the point cloud data as the comparison point cloud; Obtain the degree of match between the to-be-matched point cloud and the comparison point cloud. Use the to-be-matched point cloud corresponding to the maximum matching degree C as the target point cloud. Obtain the angle combination A corresponding to the target point cloud. Determine the current 6DoF pose of the AR device based on the 3D spatial coordinates and angle combination A. The angle settings include: Set the pitch angle range, yaw angle range and roll angle range; Starting from the starting points of the pitch angle range, yaw angle range, and roll angle range, a number of pitch angle nodes, yaw angle nodes, and roll angle nodes are set at preset angle intervals; A single angle combination includes a pitch angle node, a yaw angle node, and a roll angle node, and different angle combinations include different pitch angle nodes and / or yaw angle nodes and / or roll angle nodes.

2. The point cloud and UWB-based relocation method according to claim 1, characterized in that: Obtaining the 3D spatial coordinates of an AR device based on UWB includes: Install n UWB base stations at preset locations in the indoor space, where n is a preset number and n>6; The three-dimensional spatial coordinates of the AR device are obtained through multi-lateral positioning based on UWB two-way ranging.

3. The point cloud and UWB-based relocation method according to claim 1, characterized in that: Obtaining the matching degree between the data to be matched and the comparison data includes: The point cloud to be matched is converted into the world coordinate system to obtain a first point cloud, and the minimum circumscribed rectangle of the first point cloud is divided into voxel grids to obtain a plurality of sub-grids, which are recorded as first sub-grids and numbered; Acquire a second subgrid based on the compared point cloud and number the second subgrid; The matching degree is calculated based on the classification labels of the first sub-grid and the second sub-grid with the same number.

4. The point cloud and UWB-based relocation method according to claim 3, characterized in that: Obtaining the matching degree between the data to be matched and the comparison data also includes: Set classification labels, including occupied and idle; When the first subgrid contains a single spatial point in the first point cloud, occupied is used as the classification label of the first subgrid; otherwise, free is used as the classification label of the first subgrid; Get the classification label of the second subgrid; Count the total number of occupied numbers with the same number and classification label B1, count the total number of numbers with the same number but different classification labels B2, and calculate the matching degree P=B1 / B2.

5. The point cloud and UWB-based relocation method according to claim 4, characterized in that: When the maximum matching degree is less than the preset matching degree threshold, the following steps are performed: Start the auxiliary camera on the AR device and keep the auxiliary camera's posture to meet any angle combination to collect a depth image of the indoor space, which is recorded as the first image. The point cloud data in the single first image is fused with the comparison point cloud to obtain a new comparison point cloud. Get the current 6DoF pose of the AR device based on the new comparison point cloud.

6. The point cloud and UWB-based relocation method according to claim 5, characterized in that: Before acquiring the first image, the following steps are also included: The first subgrid and the second subgrid corresponding to the maximum matching degree C are marked as the first grid and the second grid respectively. When the classification label of the first grid numbered i is occupied and the classification label of the second grid numbered i is free, the second grid numbered i is marked as the target grid; Cluster the target grids to obtain clusters. Starting from the location of the auxiliary camera, draw a ray to the indoor space. The ray satisfies the angle combination a. If the intersection of the ray and the indoor space is not within the minimum circumscribed sphere of the cluster, the first image corresponding to the angle combination a is not obtained.

7. The point cloud and UWB-based relocation method according to claim 1, characterized in that: The current 6DoF pose of the AR device based on the 3D space coordinate and angle combination A includes: The current 6DoF pose of the AR device is (x1, y1, z1, r1, r2, r3), where x1, y1, and z1 represent the x-axis, y-axis, and z-axis coordinates of the AR device, respectively. r1, r2, and r3 represent the pitch, yaw, and roll angles corresponding to the pitch, yaw, and roll nodes in angle combination A, respectively.

Images