A point cloud map change real-time detection method based on light ray tracing and registration

Abstract

The application discloses a point cloud map change real-time detection method based on light tracking and registration, and belongs to the technical field of three-dimensional space intelligent perception and point cloud intelligent processing. The method constructs a hierarchical detection architecture of "light tracking pre-screening-double-thread registration precision verification". Firstly, a local point cloud map is dynamically constructed through three-dimensional intelligent grid division, cross-modal depth image mapping is realized in combination with spherical coordinate projection, and a suspected change point set is quickly screened based on an adaptive threshold; then, the efficiency of nearest neighbor search is optimized through hierarchical KD tree indexing, a double-thread parallel mechanism of "new point real-time verification-missing point delayed trigger verification" is adopted, and real change points are accurately identified in combination with multi-frame incremental point cloud accumulation. The application can accurately identify new obstacles, structural missing and other three-dimensional space changes, adapt to long-term point cloud high-precision point cloud map demand scenes such as automatic driving, intelligent robots and underground space intelligent operation and maintenance, and greatly reduce the map update cost and operation and maintenance redundancy.

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional spatial intelligent perception and point cloud intelligent processing technology. Specifically, it relates to a cross-modal point cloud map dynamic change real-time detection method that integrates ray tracing fast screening and KD tree registration precise verification. It is applicable to cutting-edge intelligent systems that have stringent requirements for the accuracy and real-time performance of environmental dynamic perception, such as high-order auxiliary decision-making for autonomous driving, full-domain autonomous navigation of intelligent robots, real-time updating of three-dimensional geographic information, and intelligent operation and maintenance of underground space. Background Technology

[0002] Driven by a new round of technological revolution and industrial transformation, intelligent terminals, represented by autonomous driving and intelligent robots, are evolving towards a higher level of "all-domain perception, real-time decision-making, and autonomous execution." High-precision point cloud maps, as the core digital twin carrier for intelligent systems to perceive three-dimensional physical space, directly determine the upper limit of environmental adaptability and the boundary of operational safety for intelligent systems through their dynamic adaptability. However, the real three-dimensional environment is constantly evolving. Natural erosion, artificial modification, and sudden obstacle intrusion can all cause irreversible changes in the physical space structure, causing static point cloud maps to lose their accurate representation ability in changed areas, resulting in "digital-physical space mismatch," which severely restricts the decision-making reliability and operational safety of intelligent systems.

[0003] Existing point cloud map change detection technologies are caught in a "precision-real-time" dilemma: On the one hand, ray tracing-based detection methods rely on the physical characteristic of the straight-line propagation of LiDAR light to achieve rapid classification through depth field mapping and difference comparison, possessing the potential for millisecond-level response. However, these methods are highly dependent on high-density point cloud maps and are sensitive to scene geometry and point cloud data quality. They are prone to false positives and missed weak changes due to complex scenes such as point cloud noise and small curvature planes, making it difficult to meet the requirements of high-precision perception. On the other hand, point cloud registration-based detection methods achieve refined difference recognition through spatial coordinate alignment and point-to-point distance measurement, with significant advantages in universality and detection accuracy. However, the computational complexity of full-domain point-to-point matching increases exponentially, resulting in a severe computational bottleneck in large-scale point cloud scenarios. This leads to insufficient real-time response capabilities, making it unsuitable for the real-time perception requirements of dynamic scenes.

[0004] Furthermore, traditional technical solutions generally rely on high-precision point cloud acquisition and storage across the entire domain, resulting in high time costs, computing power consumption, and manual maintenance costs for data acquisition. They also fail to construct a hierarchical detection architecture of "rapid screening - accurate verification," making it impossible to achieve a coordinated unity of "real-time response" and "high-precision recognition," thus hindering the dynamic environmental perception tasks of high-level intelligent systems. Therefore, developing a point cloud map change detection method that combines real-time response capabilities, centimeter-level detection accuracy, and wide scene adaptability is crucial to overcoming the bottlenecks in intelligent sensing technology. This method can quickly provide surveyors with the location of changing scenes, guiding them in fine scanning of changed areas and avoiding the high manual and time costs of global fine scanning. Summary of the Invention

[0005] Addressing the core pain points of existing technologies, such as "the inability to coordinate and optimize real-time performance and detection accuracy," "poor adaptability to complex scenarios," and "high data processing costs," this invention proposes a real-time point cloud map change detection method based on ray tracing and registration. By constructing a two-stage intelligent detection architecture of "pre-screening - fine verification," it integrates the efficiency of ray tracing with the high-precision advantages of KD tree registration. It innovatively adopts a dual-thread differentiated processing mechanism to achieve ultra-real-time, high-precision, and low-redundancy detection of point cloud map changes, providing reliable three-dimensional spatial dynamic perception support for high-order intelligent systems.

[0006] The technical solution adopted in this invention is as follows:

[0007] Step 1: Perform triaxial extreme value statistics and grid division on the global point cloud map, assign a unique index to each grid, and then select N×N×N neighboring grid point clouds centered on the grid where the current LiDAR is located to construct a local point cloud map; then project and map the real-time point cloud and the local point cloud map into depth images respectively, take the minimum depth value for the same pixel, and compare the depth difference between the local point cloud map and the real-time point cloud for the same pixel by setting a depth error threshold L, and filter out suspected new points and suspected missing points with depth differences exceeding the threshold to form a suspected change point set, and the remaining points are a stable point set; where N≥5;

[0008] Step 2: Construct a hierarchical KD-tree index for the local point cloud map, divided by the median of each dimension and satisfying the minimum point count termination condition; for suspected new points, find the nearest neighbor points in the local point cloud map through the KD-tree, and verify in real time based on the relationship between the distance and the threshold L to distinguish between real new points and non-changed points; for suspected missing points, generate a map of suspected missing points frame by frame and store it by grid. When the radar leaves the N×N×N neighborhood of the corresponding grid, merge the real-time point cloud of subsequent frames to construct an incremental KD-tree, and perform nearest neighbor distance verification on missing points in the same grid. If the distance is greater than the threshold L, it is determined to be a real missing point; otherwise, it is determined to be a non-changed point.

[0009] Furthermore, step 1 specifically includes the following process:

[0010] Step 1.1, Grid Division and Local Point Cloud Map Construction: Obtain the extreme value range of the point cloud map along the X, Y, and Z axes, divide the point cloud map into grids, assign the point cloud to the corresponding grid and assign a unique index ID; then, based on the current spatial location of the LiDAR, calculate the grid ID corresponding to the location of the LiDAR, filter the N×N×N grid point cloud set centered on the grid where the LiDAR is located, and merge them into a local point cloud map; where N≥5;

[0011] Step 1.2, Depth Image Mapping: Traverse the real-time point cloud acquired by the LiDAR, calculate the horizontal angle, elevation angle and depth value of each point, and map them to the corresponding pixel coordinates in the depth image; at the same time, calculate the horizontal angle, elevation angle and depth value of each point in the local point cloud map, and map them to the corresponding pixel coordinates in the depth image; if there are multiple point clouds with the same pixel coordinate, the minimum depth value is selected as the final value.

[0012] Step 1.3, Screening of suspected change points: Set an error threshold L. In the same pixel of the depth image, if the difference between the point depth value in the local point cloud map and the point depth value in the real-time point cloud is greater than L, then the pixel in the real-time point cloud map is determined to have a spatial missing position, and the corresponding point is a suspected missing point; if the difference between the point depth value in the local point cloud map and the point depth value in the real-time point cloud is less than -L, then the pixel in the real-time point cloud map is determined to have a newly added spatial missing position, and the corresponding point is a suspected newly added point; summarize the suspected newly added points and suspected missing points to form a set of suspected change points, and the remaining points are the set of stable points.

[0013] Furthermore, step 2 specifically includes the following process:

[0014] Step 2.1, Construct KD tree index: Taking the local point cloud map as the processing object, recursively select the median point of each dimension as the partition node, divide the point set into left and right subsets until the number of points in the subset is less than the preset threshold, and construct a hierarchical binary space index structure KD tree.

[0015] Step 2.2, Real-time verification thread for suspected new points: Traverse each point in the set of suspected new points, search for its nearest neighbor in the local point cloud map using a KD tree, and calculate the distance F between the two points; if F is greater than the error threshold L, it is determined to be a real new point, otherwise it is a non-changed point;

[0016] Step 2.3, Delayed Verification Thread for Suspected Missing Points: Suspected missing points from each frame are superimposed to form a suspected missing point map and assigned to corresponding grids; the grid ID of the LiDAR is recorded as the grid to be detected. When the LiDAR moves out of the boundary of the N×N×N grid set centered on the grid to be detected, the detection process is initiated: Based on the fusion of multiple frames of real-time point clouds subsequently acquired by the LiDAR, an incremental point cloud is formed. An incremental point cloud KD tree is constructed. Points with the same ID as the grid to be detected in the suspected missing point map are traversed, and their nearest neighbors in the incremental point cloud are searched and the distance F is calculated. If F is greater than L, it is determined to be a real missing point; otherwise, it is a non-changed point.

[0017] Furthermore, in step 1.3, the error threshold is twice the point cloud map density.

[0018] The advantages of this invention compared to the prior art are:

[0019] 1. Breaking through the bottleneck of "accuracy-real-time" coordination: This invention achieves rapid removal of over 90% of stable points through a ray tracing pre-screening module, transforming global detection into local focused detection, resulting in a significant reduction in data processing scale; combined with a 5×5×5 grid sliding window to dynamically construct a local map and a KD tree hierarchical index, the efficiency of nearest neighbor search is improved by more than 2 times; in the dual-thread parallel processing architecture, real-time point verification and delayed triggering of missing points are added, achieving a total detection latency of less than 100ms per frame, meeting the real-time processing requirements of 10Hz high-frequency point cloud data, with accuracy comparable to traditional high-precision registration methods, achieving a coordinated unification of "ultra-real-time response" and "centimeter-level accuracy".

[0020] 2. Robustness in complex scenarios: The ray tracing pre-screening stage of this invention adopts a conservative threshold judgment strategy to effectively avoid missing the detection of real change points; the registration and fine verification stage uses KD tree high-precision nearest neighbor matching and multi-frame incremental point cloud accumulation verification to significantly reduce false change point misjudgment caused by noise, viewpoint occlusion, and measurement errors. It can be adapted to complex indoor and outdoor topology scenes, low signal-to-noise ratio point cloud environments, and dynamic interference scenes, and its robustness far exceeds that of traditional single detection methods.

[0021] 3. Constructing a low-cost and high-efficiency detection system: This invention eliminates the need for repeated acquisition and storage of high-precision point clouds across the entire area. By dynamically constructing local point clouds and fusing incremental data, it significantly reduces the time cost and computing power consumption of data acquisition. The detection results can accurately guide staff to conduct fine-grained scanning and map updates only for the changed areas, avoiding redundant workload of repeated mapping across the entire area, improving map update efficiency, reducing operation and maintenance costs, and possessing extremely high engineering application value.

[0022] 4. Expanding adaptability to high-end intelligent scenarios: The method of this invention can be seamlessly integrated into high-end intelligent terminals such as autonomous driving, intelligent robots, and 3D geographic information systems. It can accurately identify various types of 3D spatial changes such as new obstacles, structural defects, and environmental modifications, providing core technical support for obstacle avoidance decisions in advanced assisted driving, autonomous navigation of intelligent robots, and real-time updates of 3D geographic information, and promoting the upgrade of intelligent perception technology towards "full-domain, real-time, and precise". Attached Figure Description

[0023] Figure 1 This is a schematic diagram illustrating the cross-modal depth mapping principle of the ray tracing pre-screening module of the present invention.

[0024] Figure 2 This is a flowchart illustrating the hierarchical detection process of the overall method of this invention.

[0025] Figure 3 This is a schematic diagram of the spatial boundary triggering mechanism for delayed verification of suspected missing points in this invention.

[0026] Figure 4 This is an illustration of the application effect of the present invention in the intelligent operation and maintenance scenario of underground space. Detailed Implementation

[0027] The invention will be further explained below with reference to the accompanying drawings. A method for real-time detection of point cloud map changes based on ray tracing and registration, such as... Figure 2 The hierarchical detection flowchart of the overall method is shown, including the following steps:

[0028] Step 1: Ray tracing pre-screening to generate a set of suspected change points;

[0029] The global point cloud map is subjected to triaxial extreme value statistics and grid division. A unique index is assigned to each grid. Then, with the grid where the current LiDAR is located as the center, N×N×N neighboring grid point clouds are selected to construct a local point cloud map. The real-time point cloud and the local point cloud map are then projected and mapped into depth images respectively. The minimum depth value of the same pixel is taken. By setting a depth error threshold L, the depth difference between the local point cloud map and the real-time point cloud at the same pixel is compared. Suspected new points and suspected missing points with depth differences exceeding the threshold are selected to form a set of suspected changed points. The remaining points are a set of stable points. Wherein, N≥5.

[0030] The specific process is as follows:

[0031] Step 1.1, Intelligent Grid Division and Dynamic Construction of Local Point Cloud Map: First, the extreme value range of the prior point cloud map on the X, Y, and Z axes is obtained through a 3D spatial boundary perception algorithm. , , )~( , , Based on the adaptive raster partitioning formula, the global point cloud map is discretized into 3D raster cells and divided into fixed grids. A grid.

[0032]

[0033] In the formula, K is the maximum effective measurement range of the lidar, and floor() is the floor function. For each point cloud in the global map, its corresponding unique two-dimensional spatial index ID is calculated and assigned to the grid cell, and then the points are clustered in the grid cell with that ID. This achieves spatial structured management of the point cloud and stores it in a two-dimensional linear array for easy and fast retrieval.

[0034]

[0035] In the formula, Let be the coordinates of any point in the point cloud map. Then, based on the current spatial location of the LiDAR, calculate the grid ID corresponding to the LiDAR's location. The calculation formula is the same as the point division formula, except that the input is the three-dimensional spatial coordinates of the LiDAR's location.

[0036] Based on the ID of the grid where the LiDAR is located, a set of 5×5×5 three-dimensional grid cluster IDs centered on the current LiDAR grid is dynamically filtered, and the point cloud data within the cluster is fused and stitched together to construct a lightweight, highly correlated local point cloud map, achieving "localization of global problems" and significantly reducing the scale of data processing.

[0037] .

[0038] Step 1.2, Cross-modal depth image mapping: as follows Figure 1 As shown in the schematic diagram of the cross-modal depth mapping principle of the ray tracing pre-screening module of this invention, the LiDAR point cloud can be projected into a two-dimensional depth image through a spherical coordinate projection algorithm. The specific steps are: traversing the real-time point cloud data stream collected by the LiDAR, calculating the horizontal angle θ, elevation angle φ, and three-dimensional spatial depth value d for each point, and projecting it to the corresponding pixel coordinates of the two-dimensional depth image based on coordinate mapping rules; for scenarios where multiple point clouds exist at the same pixel location, a minimum depth value priority strategy is adopted to ensure the physical authenticity of the depth image; similarly, all point cloud data in the local point cloud map are mapped to the two-dimensional depth image through the same projection rules, forming a dual depth map comparison system of "real-time point cloud depth map - local map depth map". The calculation formulas for the horizontal angle, elevation angle, and depth value are:

[0039] Horizontal angle: ,

[0040] Angle of elevation: ,

[0041] Depth value:

[0042] In the formula, arctan2() is the four-quadrant arctangent function, which can accurately represent the azimuth information of the point cloud in three-dimensional space, ensuring the uniqueness and accuracy of depth image mapping.

[0043] Step 1.3, Intelligent Threshold Determination and Suspected Change Point Screening: An error threshold L (twice the map density) is adaptively calculated based on the point cloud map density. This threshold has been verified through extensive experiments and can effectively accommodate point cloud sampling errors and environmental noise interference. A depth difference intelligent determination model is constructed: In the same pixel of the depth image, if the difference between the point depth value in the local point cloud map and the point depth value in the real-time point cloud is greater than L, the pixel in the real-time point cloud map is determined to have a spatial missing position, and the corresponding point is a suspected missing point. In the same pixel of the depth image, if the difference between the point depth value in the local point cloud map and the point depth value in the real-time point cloud is less than -L, the pixel in the real-time point cloud map is determined to have a newly added spatial missing position, and the corresponding point is a suspected newly added point. This model enables rapid separation of suspected change points and stable points. Stable point sets are directly determined as non-change points and do not require subsequent processing, achieving "redundancy reduction" in data processing.

[0044] Step 2: Dual-thread registration and fine verification to identify real change points;

[0045] A hierarchical KD-tree index is constructed for the local point cloud map, divided by the median of each dimension and satisfying the minimum point count termination condition. For suspected new points, the nearest neighbor point in the local point cloud map is found through the KD-tree, and the distance is verified in real time based on the relationship between the distance and the threshold L to distinguish between real new points and non-changed points. For suspected missing points, a map of suspected missing points is generated by overlaying frames and stored by grid. When the radar leaves the N×N×N neighborhood of the corresponding grid, the subsequent real-time point cloud of multiple frames is fused to construct an incremental KD-tree. The nearest neighbor distance of missing points in the same grid is verified. If it is greater than the threshold L, it is determined to be a real missing point; otherwise, it is determined to be a non-changed point.

[0046] The specific process is as follows:

[0047] Step 2.1, Hierarchical KD-tree index construction: Taking the local point cloud map as the processing object, a multidimensional spatial median partitioning algorithm is used to recursively select the median point of each dimension as the partitioning node, adaptively dividing the point cloud set into left and right subsets until the number of points in the subsets is less than a preset threshold (30~80). A hierarchical binary spatial index structure KD-tree is then constructed. This index structure can reduce the time complexity of nearest neighbor search from O(N) to O(…). This lays the foundation for rapid registration in the future.

[0048] Step 2.2, Real-time verification thread for suspected new points: Based on the constructed KD tree index, a real-time verification thread is started to perform a millisecond-level nearest neighbor search for each point in the suspected new point set, and calculate the spatial Euclidean distance F between the point and the nearest neighbor in the local point cloud map; if F is greater than the error threshold L, it is determined to be a real new point (i.e., a new obstacle or structure appearing in the environment); otherwise, it is determined to be a non-changed point (caused by noise or measurement error), thus achieving real-time and accurate identification of new changes.

[0049] Step 2.3, Delayed Verification Thread for Suspected Missing Points: Employing a multi-frame data accumulation and spatial boundary triggering mechanism, suspected missing points detected in each frame are superimposed to form a dynamically updated map of suspected missing points. These maps are then assigned to corresponding grids according to the grid division rules in Step 1.1. The grid ID of the LiDAR is recorded in real-time as the grid to be detected. As shown in Figure 3, which illustrates the spatial boundary triggering mechanism for delayed verification of suspected missing points, when the LiDAR moves outside the boundary of the 5×5×5 grid set centered on the grid to be detected, the delayed verification process is triggered: Based on the fusion of multiple real-time point clouds subsequently acquired by the LiDAR, an incremental point cloud is formed. An incremental point cloud KD tree index is constructed. Points matching the grid ID in the suspected missing point map are traversed, and their nearest neighbors in the incremental point cloud are searched, with the spatial Euclidean distance F calculated. If F is greater than the error threshold L, it is determined to be a true missing point (i.e., the absence or removal of existing structures in the environment); otherwise, it is determined to be a non-changed point, thus resolving the misjudgment problem caused by LiDAR viewpoint occlusion.

[0050] The error threshold L is twice the density of the point cloud map. This value has been verified through a large number of experiments and can effectively cover point cloud sampling errors, lidar measurement noise and ambient light interference, taking into account both detection accuracy and robustness.

[0051] The principle behind this invention is as follows:

[0052] 1. Preprocessing stage: Obtain the prior point cloud map of the target area, and extract the extreme value range of its X, Y, and Z axes using a 3D spatial boundary detection algorithm. , , )~( , , Set the grid side length K, and divide the prior point cloud map into three-dimensional grids according to the grid division formula. Assign a unique index ID to each grid.

[0053] 2. Ray tracing pre-screening implementation.

[0054] 3. Based on the real-time positioning information of the lidar (obtained by registering the lidar point cloud with the prior point cloud map to ensure that the coordinate system of the real-time collected point cloud data is consistent with that of the prior point cloud map), dynamically select 5×5×5 grid clusters centered on the grid where the current lidar is located, and fuse them to form a local point cloud map.

[0055] Based on the formulas for calculating horizontal angle, elevation angle and depth value, the real-time point cloud and local point cloud map are mapped to a two-dimensional depth image respectively, and the minimum depth value of the same pixel is selected as the final value.

[0056] Based on the point cloud map density (average distance E between adjacent points), an error threshold of F=2E is set; suspected new points and suspected missing points are screened through a depth difference intelligent judgment model to form a set of suspected change points. The number of suspected change points in a single frame is much smaller than the total amount of real-time point cloud.

[0057] 4. Dual-thread registration and fine-tuning implementation:

[0058] Construct a KD-tree index using a local point cloud map as the object;

[0059] Start a real-time verification thread for suspected new points. Perform nearest neighbor search on each suspected new point using a KD tree and calculate the spatial Euclidean distance F. If F > L, it is determined to be a real new point.

[0060] Start a delayed verification thread for suspected missing points, and accumulate multiple frames of suspected missing points to form a map of suspected missing points; after the LiDAR moves out of the 5×5×5 grid boundary of the grid to be detected, construct an incremental point cloud KD tree; traverse the suspected missing points corresponding to the grid to be detected, search for the nearest neighbor point and calculate the distance F. If F>L, it is determined to be a real missing point.

[0061] 5. Results Output and Application: Summarize the 3D coordinates of the actual newly added points and the actual missing points detected by the LiDAR in each frame. Finally, obtain and display all missing and newly added points on the map in real time. For example... Figure 4 This is an illustration of the application effect of the present invention in the intelligent operation and maintenance scenario of underground space.

Claims

1. A real-time change detection method for point cloud maps based on ray tracing and registration, characterized in that, Includes the following steps: Step 1: Perform triaxial extreme value statistics and grid division on the global point cloud map, assign a unique index to each grid, and then select N×N×N neighboring grid point clouds centered on the grid where the current LiDAR is located to construct a local point cloud map; then project and map the real-time point cloud and the local point cloud map into depth images respectively, take the minimum depth value for the same pixel, and compare the depth difference between the local point cloud map and the real-time point cloud for the same pixel by setting a depth error threshold L, and filter out suspected new points and suspected missing points with depth differences exceeding the threshold to form a suspected change point set, and the remaining points are a stable point set; where N≥5; Step 2: Construct a hierarchical KD-tree index for the local point cloud map, divided by the median of each dimension and satisfying the minimum point count termination condition; for suspected new points, find the nearest neighbor points in the local point cloud map through the KD-tree, and verify in real time based on the relationship between the distance and the threshold L to distinguish between real new points and non-changed points; for suspected missing points, generate a map of suspected missing points frame by frame and store it by grid. When the radar leaves the N×N×N neighborhood of the corresponding grid, merge the real-time point cloud of subsequent frames to construct an incremental KD-tree, and perform nearest neighbor distance verification on missing points in the same grid. If the distance is greater than the threshold L, it is determined to be a real missing point; otherwise, it is determined to be a non-changed point.

2. The method for real-time detection of point cloud map changes based on ray tracing and registration according to claim 1, characterized in that, The specific process of step 1 includes: Step 1.1, Grid Division and Local Point Cloud Map Construction: Obtain the extreme value range of the point cloud map along the X, Y, and Z axes, divide the point cloud map into grids, assign the point cloud to the corresponding grid and assign a unique index ID; then, based on the current spatial location of the LiDAR, calculate the grid ID corresponding to the location of the LiDAR, filter the N×N×N grid point cloud set centered on the grid where the LiDAR is located, and merge them into a local point cloud map; where N≥5; Step 1.2, Depth Image Mapping: Traverse the real-time point cloud acquired by the LiDAR, calculate the horizontal angle, elevation angle and depth value of each point, and map them to the corresponding pixel coordinates in the depth image; at the same time, calculate the horizontal angle, elevation angle and depth value of each point in the local point cloud map, and map them to the corresponding pixel coordinates in the depth image; if there are multiple point clouds with the same pixel coordinate, the minimum depth value is selected as the final value. Step 1.3, Screening of suspected change points: Set an error threshold L. In the same pixel of the depth image, if the difference between the point depth value in the local point cloud map and the point depth value in the real-time point cloud is greater than L, then the pixel in the real-time point cloud map is determined to have a spatial missing position, and the corresponding point is a suspected missing point; if the difference between the point depth value in the local point cloud map and the point depth value in the real-time point cloud is less than -L, then the pixel in the real-time point cloud map is determined to have a newly added spatial missing position, and the corresponding point is a suspected newly added point; summarize the suspected newly added points and suspected missing points to form a set of suspected change points, and the remaining points are the set of stable points.

3. The method for real-time detection of point cloud map changes based on ray tracing and registration according to claim 2, characterized in that, The specific process of step 2 includes: Step 2.1, Construct KD tree index: Taking the local point cloud map as the processing object, recursively select the median point of each dimension as the partition node, divide the point set into left and right subsets until the number of points in the subset is less than the preset threshold, and construct a hierarchical binary space index structure KD tree. Step 2.2, Real-time verification thread for suspected new points: Traverse each point in the set of suspected new points, search for its nearest neighbor in the local point cloud map using a KD tree, and calculate the distance F between the two points; if F is greater than the error threshold L, it is determined to be a real new point, otherwise it is a non-changed point; Step 2.3, Delayed Verification Thread for Suspected Missing Points: Suspected missing points from each frame are superimposed to form a suspected missing point map and assigned to corresponding grids; the grid ID of the LiDAR is recorded as the grid to be detected. When the LiDAR moves out of the boundary of the N×N×N grid set centered on the grid to be detected, the detection process is initiated: Based on the fusion of multiple frames of real-time point clouds subsequently acquired by the LiDAR, an incremental point cloud is formed. An incremental point cloud KD tree is constructed. Points with the same ID as the grid to be detected in the suspected missing point map are traversed, and their nearest neighbors in the incremental point cloud are searched and the distance F is calculated. If F is greater than L, it is determined to be a real missing point; otherwise, it is a non-changed point.

4. The method for real-time detection of point cloud map changes based on ray tracing and registration according to claim 2, characterized in that, In step 1.3, the error threshold is twice the density of the point cloud map.

Images