A method, system, device, and medium for registration of lidar point clouds

Abstract

The application provides a laser radar point cloud registration method, system, device and medium, the registration method comprises: obtaining continuous multiple frames of local point clouds of a target object from a laser radar, and converting the local point clouds of each frame into the same local coordinate system to obtain a point cloud sequence; for the point cloud sequence, perform slicing processing according to the time sequence of each frame, and cut the point cloud sequence into multiple time sequence segments; extract the spatial geometric distribution features of each time sequence segment in the local coordinate system; based on the spatial geometric distribution features, calculate the spatial occupation mask of each time sequence segment respectively, and divide the time sequence segments with the same spatial occupation mask into the same group to obtain at least one segment group; based on the segment group, registration is performed to obtain a target aggregated point cloud of the target object. The application can improve the accuracy of point cloud registration.

Description

Technical Field

[0001] This invention application relates to the field of point cloud registration, and more particularly to a registration method, system, device, and medium for lidar point clouds. Background Technology

[0002] In autonomous driving scenarios, vehicle-mounted LiDAR can continuously acquire point clouds of the road environment surrounding the vehicle. However, for dynamic targets such as vehicles on the road, due to factors such as the target's own motion, changes in environmental occlusion, frequent switching of observation angles, and the sparsity of the point cloud itself, a single frame of point cloud can usually only capture a local geometric surface of the target at a certain instant, making it difficult to completely and stably represent the overall geometric shape of the target. In order to obtain high-precision and high-completeness target-level point clouds for downstream tasks, it is usually necessary to aggregate the point cloud observation results of the same target in multiple consecutive frames. Point cloud registration technology is the core means to achieve spatial alignment, accumulation, and fusion of point clouds from multiple frames.

[0003] In existing point cloud aggregation schemes, local point clouds of the same target are typically extracted from multiple consecutive frames of point clouds based on the target tracking bounding box. Using a specific frame or the accumulated point cloud as a reference, frame-by-frame registration and temporal accumulation are performed on the local point clouds of subsequent frames in chronological order. However, in actual driving, dynamic targets present completely different spatial visibility and spatial occupancy patterns to the lidar at different times. The aforementioned existing schemes mechanically understand point cloud aggregation as a single-layer temporal frame-by-frame registration problem, relying entirely on the chronological order of the timeline to stitch point clouds together and perform registration. Because they lack targeted processing of the spatial visibility and spatial geometric distribution relationships of point cloud fragments, when the target produces local fragments with different spatial occupancy patterns due to significant viewpoint rotation or severe occlusion, the system cannot effectively utilize and organize these spatial relationships, blindly registering temporally adjacent fragments. This easily leads to defects such as local geometric feature misalignment and blurring, resulting in low accuracy in point cloud registration. Summary of the Invention

[0004] This invention application provides a method, system, device, and medium for registering lidar point clouds, in order to solve the technical problem of how to improve the accuracy of point cloud registration.

[0005] To address the aforementioned technical problems, this invention provides a registration method for lidar point clouds, comprising: The local point cloud of the target object is obtained from multiple consecutive frames by the lidar, and the local point cloud of each frame is transformed into the same local coordinate system to obtain a point cloud sequence. The point cloud sequence is segmented according to the time order of each frame, and the point cloud sequence is divided into multiple temporal segments. Extract the spatial geometric distribution features of each time segment in the local coordinate system; based on the spatial geometric distribution features, calculate the spatial occupancy mask of each time segment, and divide the time segments with the same spatial occupancy mask into the same group to obtain at least one segment group; Based on the fragment group, registration is performed to obtain the target aggregate point cloud of the target object.

[0006] As a preferred embodiment, the local coordinate system is a three-dimensional coordinate system, and the local coordinate system is divided into eight subspaces according to the three coordinate axes of the three-dimensional coordinate system; The step involves extracting the spatial geometric distribution features of each time segment in the local coordinate system; and calculating the spatial occupancy mask for each time segment based on the spatial geometric distribution features, including: Based on the coordinates of each point in each time segment in the local coordinate system, the spatial geometric distribution features are extracted. Based on the spatial geometric distribution characteristics, determine the occupancy status of each subspace; Based on the determination result of the occupancy status, the mask bits corresponding to the subspaces that constitute effective occupancy are marked as one, and the mask bits corresponding to the subspaces that do not constitute effective occupancy are marked as zero, thus obtaining the space occupancy mask of the timing segment consisting of eight mask bits.

[0007] As a preferred embodiment, the spatial geometric distribution features include the number of points falling into the subspace; The step of determining the occupancy status of each subspace based on the spatial geometric distribution characteristics includes: When the number of points is greater than or equal to a preset number threshold, and the average point spacing of the point cloud within the subspace is less than or equal to a preset density threshold, the subspace is determined to constitute effective occupancy. If the number of points is less than the number threshold, or the average point spacing of the point cloud within the subspace is greater than the density threshold, the subspace is determined not to constitute effective occupancy.

[0008] As a preferred embodiment, the point cloud sequence is segmented according to the temporal order of each frame, dividing the point cloud sequence into multiple temporal segments, including: The earliest frame in the point cloud sequence in chronological order is determined as the starting frame of the current segment; According to the time sequence, for each frame in the point cloud sequence after the starting frame, the current frame is initially registered to the current segment in turn, so as to obtain the initial registration result of the current frame relative to the current segment; Based on the preliminary registration results, it is determined whether the current frame meets the merging condition by using preset gating conditions; When the current frame meets the merging condition, the current frame is merged into the current segment; when the current frame does not meet the merging condition, the current segment is determined as a segmented temporal segment, until the processing of each frame in the point cloud sequence is completed, and multiple temporal segments are obtained.

[0009] As a preferred embodiment, the preset gating conditions include at least one of effective matching ratio gating, registration quality gating, translation amount gating, and attitude change amount gating; The effective matching ratio gate is defined as follows: the proportion of effective points in the preliminary registration result reaches a preset ratio threshold; wherein, the proportion of effective points is the proportion of points in the point cloud of the current frame whose nearest neighbor distance to the point cloud of the current segment does not exceed a preset distance threshold. The registration quality gating is defined as follows: the registration error does not exceed a preset error threshold. The translation amount gate is defined as follows: the translation amount required to align the current frame to the current segment does not exceed a preset translation threshold; wherein, the translation amount is determined based on the translation vector in the transformation matrix corresponding to the preliminary registration result; The attitude change gating is as follows: the attitude change generated by the initial registration of the current frame does not exceed a preset attitude threshold; wherein, the attitude change is obtained by decomposing the rotation data of the transformation matrix; The step of determining whether the current frame meets the merging condition based on the preliminary registration result and using preset gating conditions includes: determining that the current frame meets the merging condition when the preliminary registration result meets all the gating conditions used in the preset gating conditions; otherwise, determining that the current frame does not meet the merging condition.

[0010] As a preferred embodiment, the step of registering based on the fragment group to obtain the target aggregated point cloud of the target object includes: For each of the aforementioned fragment groups, the temporal fragments within the fragment group are registered and fused within the group to obtain the group-level aggregated point cloud corresponding to the fragment group. The group-level aggregated point clouds corresponding to each of the aforementioned fragment groups are registered and fused between groups to obtain a preliminary aggregated point cloud. The time segments that failed to be successfully fused during the intra-group registration and fusion and inter-group registration and fusion processes are registered to the preliminary aggregated point cloud and fused to obtain the target aggregated point cloud of the target object.

[0011] As a preferred embodiment, the intra-group registration and fusion, the inter-group registration and fusion, and the registration and fusion of time segments that were not successfully fused during the intra-group registration and fusion and the inter-group registration and fusion to the preliminary aggregated point cloud are all achieved by performing fine registration between the point cloud to be registered and the preset reference point cloud. The refined registration includes: Based on the point spacing between the point cloud to be registered and the reference point cloud, determine the voxel downsampling scale and the maximum corresponding distance threshold used for the fine registration; Based on the voxel downsampling scale and the maximum corresponding distance threshold, multi-stage point cloud registration is performed on the point cloud to be registered and the reference point cloud. The multi-stage point cloud registration includes initial alignment at a first scale and fine registration at a second scale smaller than the first scale. Each stage prioritizes point-to-surface iterative nearest point registration. When the point-to-surface iterative nearest point registration does not meet the preset registration conditions, point-to-point iterative nearest point registration is used.

[0012] Accordingly, this invention application also provides a registration system for lidar point clouds, including a conversion module, a segmentation module, a grouping module, and a registration module; wherein, The conversion module is used to acquire multiple consecutive frames of local point clouds of the target object from the lidar, and convert the local point clouds of each frame into the same local coordinate system to obtain a point cloud sequence. The segmentation module is used to segment the point cloud sequence into multiple temporal segments according to the temporal order of each frame. The grouping module is used to extract the spatial geometric distribution features of each time segment in the local coordinate system; based on the spatial geometric distribution features, it calculates the spatial occupancy mask of each time segment, and divides time segments with the same spatial occupancy mask into the same group to obtain at least one segment group. The registration module is used to perform registration based on the fragment group to obtain the target aggregate point cloud of the target object.

[0013] As a preferred embodiment, the local coordinate system is a three-dimensional coordinate system, and the local coordinate system is divided into eight subspaces according to the three coordinate axes of the three-dimensional coordinate system; The grouping module extracts the spatial geometric distribution features of each time segment in the local coordinate system; based on the spatial geometric distribution features, it calculates the spatial occupancy mask for each time segment, including: The grouping module extracts the spatial geometric distribution features based on the coordinates of each point in each time segment in the local coordinate system. Based on the spatial geometric distribution characteristics, determine the occupancy status of each subspace; Based on the determination result of the occupancy status, the mask bits corresponding to the subspaces that constitute effective occupancy are marked as one, and the mask bits corresponding to the subspaces that do not constitute effective occupancy are marked as zero, thus obtaining the space occupancy mask of the timing segment consisting of eight mask bits.

[0014] As a preferred embodiment, the spatial geometric distribution features include the number of points falling into the subspace; The grouping module determines the occupancy status of each subspace based on the spatial geometric distribution characteristics, including: When the number of points is greater than or equal to a preset number threshold, and the average point spacing of the point cloud within the subspace is less than or equal to a preset density threshold, the grouping module determines that the subspace constitutes effective occupancy. If the number of points is less than the number threshold, or the average point spacing of the point cloud within the subspace is greater than the density threshold, the subspace is determined not to constitute effective occupancy.

[0015] As a preferred embodiment, the segmentation module performs segmentation processing on the point cloud sequence according to the temporal order of each frame, dividing the point cloud sequence into multiple temporal segments, including: The segmentation module determines the earliest frame in the point cloud sequence in time sequence as the starting frame of the current segment; According to the time sequence, for each frame in the point cloud sequence after the starting frame, the current frame is initially registered to the current segment in turn, so as to obtain the initial registration result of the current frame relative to the current segment; Based on the preliminary registration results, it is determined whether the current frame meets the merging condition by using preset gating conditions; When the current frame meets the merging condition, the current frame is merged into the current segment; when the current frame does not meet the merging condition, the current segment is determined as a segmented temporal segment, until the processing of each frame in the point cloud sequence is completed, and multiple temporal segments are obtained.

[0016] As a preferred embodiment, the preset gating conditions include at least one of effective matching ratio gating, registration quality gating, translation amount gating, and attitude change amount gating; The effective matching ratio gate is defined as follows: the proportion of effective points in the preliminary registration result reaches a preset ratio threshold; wherein, the proportion of effective points is the proportion of points in the point cloud of the current frame whose nearest neighbor distance to the point cloud of the current segment does not exceed a preset distance threshold. The registration quality gating is defined as follows: the registration error does not exceed a preset error threshold. The translation amount gate is defined as follows: the translation amount required to align the current frame to the current segment does not exceed a preset translation threshold; wherein, the translation amount is determined based on the translation vector in the transformation matrix corresponding to the preliminary registration result; The attitude change gating is as follows: the attitude change generated by the initial registration of the current frame does not exceed a preset attitude threshold; wherein, the attitude change is obtained by decomposing the rotation data of the transformation matrix; The segmentation module determines whether the current frame meets the merging condition based on the preliminary registration result and through preset gating conditions, including: if the preliminary registration result meets all the gating conditions used in the preset gating conditions, the segmentation module determines that the current frame meets the merging condition; otherwise, it determines that the current frame does not meet the merging condition.

[0017] As a preferred embodiment, the registration module performs registration based on the fragment group to obtain the target aggregated point cloud of the target object, including: The registration module performs intra-group registration and fusion of each temporal segment within each segment group to obtain a group-level aggregated point cloud corresponding to the segment group. The group-level aggregated point clouds corresponding to each of the aforementioned fragment groups are registered and fused between groups to obtain a preliminary aggregated point cloud. The time segments that failed to be successfully fused during the intra-group registration and fusion and inter-group registration and fusion processes are registered to the preliminary aggregated point cloud and fused to obtain the target aggregated point cloud of the target object.

[0018] As a preferred embodiment, the intra-group registration and fusion, the inter-group registration and fusion, and the registration and fusion of time segments that were not successfully fused during the intra-group registration and fusion and the inter-group registration and fusion to the preliminary aggregated point cloud are all achieved by performing fine registration between the point cloud to be registered and the preset reference point cloud. The refined registration includes: Based on the point spacing between the point cloud to be registered and the reference point cloud, determine the voxel downsampling scale and the maximum corresponding distance threshold used for the fine registration; Based on the voxel downsampling scale and the maximum corresponding distance threshold, multi-stage point cloud registration is performed on the point cloud to be registered and the reference point cloud. The multi-stage point cloud registration includes initial alignment at a first scale and fine registration at a second scale smaller than the first scale. Each stage prioritizes point-to-surface iterative nearest point registration. When the point-to-surface iterative nearest point registration does not meet the preset registration conditions, point-to-point iterative nearest point registration is used.

[0019] Accordingly, this application also provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the lidar point cloud registration method.

[0020] Accordingly, this application also provides a computer-readable storage medium, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the lidar point cloud registration method.

[0021] Compared with the prior art, this invention application has the following beneficial effects: This invention provides a method, system, device, and medium for registering lidar point clouds. By extracting the spatial geometric distribution features of each temporal segment and calculating the spatial occupancy mask, temporal segments with the same spatial occupancy mask are grouped together for registration. This method directly overcomes the limitations of existing solutions that rely entirely on the sequential order of the timeline for blind stitching. By grouping segments using spatial occupancy masks, it ensures that segments with similar spatial occupancy can be registered according to the resulting groups. This fundamentally avoids defects such as local geometric feature misalignment and distortion caused by forcibly registering temporally adjacent segments with large differences in spatial occupancy patterns when dynamic targets are subjected to significant viewpoint rotation or severe occlusion. This significantly improves the accuracy of lidar point cloud registration.

[0022] Furthermore, by first segmenting the point cloud sequence according to time sequence to isolate temporally unstable observations, and then regrouping multiple temporal segments spatially using spatial occupancy masks, this technique of first temporally segmenting and then spatially reorganizing allows for the accurate identification and aggregation of scattered segments with the same or similar spatial occupancy patterns even when dynamic targets cause temporal discontinuities or sudden changes in viewpoint in the point cloud observation due to violent movement or severe occupancy on the time axis. This significantly improves the robustness and scene adaptability of the registration method in complex, highly dynamic autonomous driving scenarios. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating an embodiment of the lidar point cloud registration method provided in this invention application.

[0024] Figure 2 This is a schematic diagram of an embodiment of the lidar point cloud registration system provided in this invention application. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1 Please refer to Figure 1 This is a flowchart illustrating an embodiment of the lidar point cloud registration method provided in this invention application.

[0027] Figure 1 The illustrated embodiment includes steps S101 to S104; each step is described in detail below: Step S101: Obtain local point clouds of the target object from multiple consecutive frames using the lidar, and transform the local point clouds of each frame into the same local coordinate system to obtain a point cloud sequence.

[0028] Specifically, this embodiment can be applied to offline or online point cloud processing processes in autonomous driving or assisted driving systems, and its execution entity can be a point cloud processing module in a vehicle-side computing platform, a cloud-based data processing platform, or an offline data production system.

[0029] Its input data can include multiple consecutive frames of LiDAR point clouds, target detection results or target tracking results for each frame, as well as configuration parameters such as sensor extrinsic parameters, region of interest range, registration threshold, and space occupancy threshold.

[0030] When acquiring multiple consecutive frames of local point cloud data of a target object from a lidar and transforming them to the same local coordinate system, this embodiment can be performed as follows: Read a frame of LiDAR point cloud and the target tracking results for that frame; convert the point cloud to a unified working coordinate system based on sensor extrinsic parameters; perform region of interest filtering on the entire frame of point cloud to remove obviously irrelevant areas.

[0031] For a given target object, candidate target points are extracted based on the 3D detection bounding box and / or post-processing detection bounding box. Intensity filtering, timestamp main cluster retention, and voxel deduplication are performed on the candidate target points to reduce noise points and duplicate points. Then, based on the center and orientation of the target bounding box, the point cloud of the target object is transformed into a local coordinate system with the center of the target bounding box as the origin and aligned with the target orientation to obtain the local point cloud of the target object in the frame.

[0032] For each target object, its local point cloud, target category, frame number, transformation from local coordinate system to working coordinate system, and target bounding box size can be saved frame by frame. Thus, the local point clouds of multiple consecutive frames can be organized into a point cloud sequence for the target object according to the target object's number.

[0033] In some embodiments, the target candidate region used to extract candidate target points is not limited to a three-dimensional detection box, but can also be replaced by a two-dimensional detection box, segmentation results or other forms of target candidate regions; the working coordinate system can also be selected as a vehicle coordinate system, sensor coordinate system or other unified coordinate system according to the system definition.

[0034] like Figure 1 As shown, the overall processing flow of this embodiment can be summarized as follows: taking multiple consecutive frames of lidar point clouds and their detection results as input, firstly, the local point clouds of each target object are extracted and organized into a point cloud sequence through the above-mentioned step S101; then, the point cloud sequence is subjected to temporal gating segmentation through step S102 to isolate temporally unstable observations and protect the purity of local segments; then, through steps S103 to S104, each temporal segment is regrouped from the spatial dimension and hierarchical registration is performed, thereby restoring the overall integrity of the target while taking into account the purity of the point cloud.

[0035] Step S102: For the point cloud sequence, perform segmentation processing according to the time order of each frame to divide the point cloud sequence into multiple temporal segments.

[0036] Based on the point cloud sequence obtained in step S101, this embodiment can perform temporal incremental registration and gated segmentation processing on multiple frames of local point cloud sequences of the same target object to divide the point cloud sequence into multiple temporal segments.

[0037] In some preferred embodiments, the step of segmenting the point cloud sequence according to the temporal order of each frame, dividing the point cloud sequence into multiple temporal segments, includes: The earliest frame in the point cloud sequence in chronological order is determined as the starting frame of the current segment; According to the time sequence, for each frame in the point cloud sequence after the starting frame, the current frame is initially registered to the current segment in turn, so as to obtain the initial registration result of the current frame relative to the current segment; Based on the preliminary registration results, it is determined whether the current frame meets the merging condition by using preset gating conditions; When the current frame meets the merging condition, the current frame is merged into the current segment; when the current frame does not meet the merging condition, the current segment is determined as a segmented temporal segment, until the processing of each frame in the point cloud sequence is completed, and multiple temporal segments are obtained.

[0038] Specifically, determining the first frame in the time sequence of the point cloud sequence as the starting frame of the current segment can be done as follows: for the first frame of the target object, directly use it as the starting frame of the first temporal segment, and add the point cloud of this frame to the cumulative point cloud and the voxelized segment target point cloud of the current segment. The starting frame also defines the reference coordinate system of the current segment.

[0039] Furthermore, for each frame after the starting frame, this embodiment sequentially performs preliminary registration of the current frame to the current segment: first, it constructs the initial registration value of the current frame to the current segment coordinate system, and then performs preliminary registration of the current frame to the segment target point cloud of the current segment.

[0040] When constructing the initial registration value, the first frame of the current segment is used as the reference frame, and its pose is set as the unit transformation matrix. For subsequent frames to be merged, if the previous frame has already obtained and recorded the final registration pose, then the final registration pose of the previous frame is used as the initial registration value of the current frame. If the previous frame has not formed a usable pose, then the previous valid pose maintained by the current segment is used as the initial registration value of the current frame. The previous valid pose represents the transformation from the local point cloud of the received frame to the coordinate system of the current segment. Thus, subsequent frames perform local registration optimization based on temporal continuity and historical valid poses within the segment.

[0041] Furthermore, before formal registration, a pre-check of the effective matching ratio can be performed. This involves transforming the current frame to the current segment coordinate system using the initial registration value and calculating the proportion of points in its point cloud that can be found in the vicinity of the segment's target point cloud. If this proportion is too low, it indicates insufficient overlap between the current frame and the current segment, or that the initial registration value is unreliable. Subsequently, preliminary registration is performed on the current frame to obtain the transformation from the current frame to the current segment coordinate system, and the registration quality index is calculated based on this transformation, serving as the basis for subsequent gating judgments.

[0042] Preferably, the preset gating conditions include at least one of effective matching ratio gating, registration quality gating, translation amount gating, and attitude change amount gating; The effective matching ratio gate is defined as follows: the proportion of effective points in the preliminary registration result reaches a preset ratio threshold; wherein, the proportion of effective points is the proportion of points in the point cloud of the current frame whose nearest neighbor distance to the point cloud of the current segment does not exceed a preset distance threshold. The registration quality gating is defined as follows: the registration error does not exceed a preset error threshold. The translation amount gate is defined as follows: the translation amount required to align the current frame to the current segment does not exceed a preset translation threshold; wherein, the translation amount is determined based on the translation vector in the transformation matrix corresponding to the preliminary registration result; The attitude change gating is as follows: the attitude change generated by the initial registration of the current frame does not exceed a preset attitude threshold; wherein, the attitude change is obtained by decomposing the rotation data of the transformation matrix; The step of determining whether the current frame meets the merging condition based on the preliminary registration result and using preset gating conditions includes: determining that the current frame meets the merging condition when the preliminary registration result meets all the gating conditions used in the preset gating conditions; otherwise, determining that the current frame does not meet the merging condition.

[0043] Specifically, the preset gating condition is used to determine whether the current frame is allowed to be merged into the current segment, and is the admission rule for merging the current frame into the current segment. When performing gating judgment, the current segment can be composed of multiple accepted frames. In this embodiment, the point clouds of each accepted frame in the time segment are accumulated in a voxelized manner to form the segment target point cloud, and the segment target point cloud is used as the reference object for the current frame registration and quality evaluation.

[0044] Assuming the local point cloud of the current frame, the target point cloud of the current segment, and the rigid body transformation obtained from the initial registration of the current frame are all determined, let the k-th point p k For example, the coordinates of each point in the current frame after the rigid body transformation can be expressed as: ; In the formula, p k The point before the rigid body transformation. R is the point after rigid body transformation. i For the rotational part of a rigid body transformation, t i Let be the translation vector for the rigid body transformation.

[0045] For each transformed point, its nearest neighbor is found in the target point cloud of the segment, and the corresponding nearest neighbor distance is calculated. Based on the above nearest neighbor distance, the proportion of valid points in the current frame relative to the current segment and the registration error can be calculated.

[0046] In the effective matching ratio gating, the maximum corresponding distance threshold is set as the preset distance threshold. Points whose nearest neighbor distance does not exceed the preset distance threshold constitute effective corresponding points. The proportion of effective points is the ratio of the number of effective corresponding points to the total number of point clouds in the current frame. The proportion of effective points is also the registration adaptation degree in this scheme, used to measure the degree of matching between the current frame and the target point cloud of the current segment. When this proportion reaches the preset proportion threshold, the current frame is determined to have passed the effective matching ratio gating.

[0047] In the registration quality gating, the registration error can be expressed as the nearest neighbor distance d between effective corresponding points. k Root mean square (RMSE) i The representation can be expressed as: ; In the formula, I i This represents the i-th valid corresponding point. Indicates the number of valid corresponding points; The smaller the registration error, the more accurately the current frame is aligned with the current segment. When the registration error does not exceed the preset error threshold, the current frame is determined to have passed the registration quality gate.

[0048] In the translation amount gating, the translation amount trans iBased on the translation vector in the transformation matrix corresponding to the preliminary registration result Determined can be represented as: ; In the formula, t x , t y and t z These represent the translation components that need to be applied in the x, y, and z directions to align the current frame point cloud with the current segment coordinate system.

[0049] When the value does not exceed the preset translation threshold, the current frame is determined to have passed the translation amount gating.

[0050] In the attitude change gating, the attitude change is obtained by decomposing the rotation part of the transformation matrix (i.e., the rotation data mentioned above), for example, by decomposing it into roll angle, pitch angle and yaw angle around three coordinate axes, and comparing them with preset attitude thresholds respectively. When the change does not exceed the preset attitude threshold, the current frame is determined to have passed the attitude change gating.

[0051] In some embodiments, the preset gating condition may further include failure rollback gating, that is, when the current frame registration fails and rollback is performed using the initial registration value, it is further determined whether the pose after rollback is still within a reasonable range, so as to avoid abnormal poses being incorporated into the current segment.

[0052] Furthermore, if the initial registration result satisfies all the gating conditions adopted, the current frame is determined to meet the merging condition. At this time, the current frame is transformed to the current segment coordinate system according to the transformation matrix obtained from the registration, and merged into the current segment. At the same time, the cumulative point cloud, the voxelized segment target point cloud, and the previous valid pose of the current segment are updated. Otherwise, the current frame is determined not to meet the merging condition. At this time, merging is not forced, but the current segment is terminated, and a new temporal segment is initialized with the current frame.

[0053] It should be noted that separating the current frame from the current segment does not mean discarding the frame, but rather isolating it from the current segment as the starting point for a new temporal segment. This avoids contaminating existing temporal segments while preserving the opportunity for re-fusion in subsequent spatial occupancy grouping and higher-level registration. The second stage ultimately outputs multiple temporal segments. Each temporal segment, as it enters subsequent steps, can record its original frames, registration quality metrics, and segmentation markers.

[0054] Step S103: Extract the spatial geometric distribution features of each time segment in the local coordinate system; based on the spatial geometric distribution features, calculate the spatial occupancy mask of each time segment, and divide the time segments with the same spatial occupancy mask into the same group to obtain at least one segment group.

[0055] Based on the multiple time series segments obtained in step S102, this embodiment reorganizes each time series segment from a spatial dimension: firstly, it extracts the spatial geometric distribution features of each time series segment in the local coordinate system, and calculates the spatial occupancy mask of each time series segment accordingly. Then, it divides time series segments with the same spatial occupancy mask into the same group to obtain at least one segment group, thereby allowing time series segments with similar spatial visible areas to be grouped together for priority fusion.

[0056] In some preferred embodiments, the local coordinate system is a three-dimensional coordinate system, and the local coordinate system is divided into eight subspaces according to the three coordinate axes of the three-dimensional coordinate system; The step involves extracting the spatial geometric distribution features of each time segment in the local coordinate system; and calculating the spatial occupancy mask for each time segment based on the spatial geometric distribution features, including: Based on the coordinates of each point in each time segment in the local coordinate system, the spatial geometric distribution features are extracted. Based on the spatial geometric distribution characteristics, determine the occupancy status of each subspace; Based on the determination result of the occupancy status, the mask bits corresponding to the subspaces that constitute effective occupancy are marked as one, and the mask bits corresponding to the subspaces that do not constitute effective occupancy are marked as zero, thus obtaining the space occupancy mask of the timing segment consisting of eight mask bits.

[0057] Specifically, the local coordinate system has been established in step S101: the center of the target box is taken as the origin of the local coordinate system, and the target point cloud is rotated to a representation consistent with the target orientation according to the heading angle of the target box, so that the point clouds of the same target object in different frames or different time segments can be compared in spatial distribution under this local coordinate system.

[0058] When calculating the space occupancy mask, this embodiment uses the three coordinate planes x=0, y=0, and z=0 in the local coordinate system as boundaries. Based on the positive and negative relationship of the x, y, and z coordinates of each point relative to the local coordinate origin, the space surrounding the target is divided into eight subspaces, and each point is assigned to the corresponding subspace. This division method can be understood as a 2×2×2 space division formed along the three coordinate axes. The judgment is based on the positive and negative relationship of the point coordinates relative to the local coordinate origin, without requiring the target's outer space to be constructed as a strict cube.

[0059] Subsequently, the spatial geometric distribution features are extracted based on the coordinates of each point in the local coordinate system in each time segment, and the occupancy status of each subspace is determined accordingly. For subspaces that constitute effective occupancy, their corresponding mask bits are marked as one, and for subspaces that do not constitute effective occupancy, their corresponding mask bits are marked as zero. Thus, each time segment obtains a spatial occupancy mask consisting of eight mask bits, which is used to describe the visible area distribution of the time segment in the local coordinate system.

[0060] In some preferred embodiments, the spatial geometric distribution features include the number of points falling into the subspace; The step of determining the occupancy status of each subspace based on the spatial geometric distribution characteristics includes: When the number of points is greater than or equal to a preset number threshold, and the average point spacing of the point cloud within the subspace is less than or equal to a preset density threshold, the subspace is determined to constitute effective occupancy. If the number of points is less than the number threshold, or the average point spacing of the point cloud within the subspace is greater than the density threshold, the subspace is determined not to constitute effective occupancy.

[0061] Specifically, the quantity threshold can take several forms: for example, it can be set as a preset proportion relative to the total number of points in the current segment, and if the number of points in a subspace does not reach the number corresponding to the preset proportion, it is considered that the number of points is insufficient; it can also be set as a preset minimum number of points, and if the number of points in a subspace is lower than the minimum number of points, it is considered that the number of points is insufficient; or both the above proportion and the minimum number of points can be used as constraints. Provided that the number of points meets the requirements, the typical point spacing of the point cloud in the subspace is further calculated as the average point spacing, and if it does not exceed a preset density threshold, the subspace is determined to be effectively occupied. Thus, the determination of effective occupancy is not simply based on the presence or absence of points in the subspace, but rather on a comprehensive assessment of the point proportion, the minimum number of points, and the point cloud density, thereby avoiding the misjudgment of a small number of sparse or discrete points as effectively occupied.

[0062] Step S104: Register the fragment group to obtain the target aggregated point cloud of the target object.

[0063] In this step, hierarchical registration can be performed based on at least one group of segments to obtain the target aggregate point cloud of the target object.

[0064] In some preferred embodiments, the step of registering based on the fragment group to obtain the target aggregated point cloud of the target object includes: For each of the aforementioned fragment groups, the temporal fragments within the fragment group are registered and fused within the group to obtain the group-level aggregated point cloud corresponding to the fragment group. The group-level aggregated point clouds corresponding to each of the aforementioned fragment groups are registered and fused between groups to obtain a preliminary aggregated point cloud. The time segments that failed to be successfully fused during the intra-group registration and fusion and inter-group registration and fusion processes are registered to the preliminary aggregated point cloud and fused to obtain the target aggregated point cloud of the target object.

[0065] Specifically, before performing registration and fusion within a group, several key groups can be selected from each fragment group to participate in the fusion process first.

[0066] The selection of the primary group can be based on engineering rules such as spatial coverage richness and the number of segments. For example, segments occupying more subspaces are given priority; segments with more time-series segments within a group are given priority; and segments that meet the minimum number of time-series segments or minimum space occupancy requirements are given priority. Time-series segments that are not selected as primary groups are not discarded, but are retained for subsequent supplementary registration stages to continue to be attempted for incorporation.

[0067] Furthermore, for each of the aforementioned fragment groups, intra-group registration and fusion are performed on each temporal fragment within the fragment group to obtain the group-level aggregated point cloud corresponding to the fragment group. Specifically, this may include: selecting a temporal fragment within the group as the initial group target point cloud; sequentially using other temporal fragments within the group as point clouds to be registered, and performing fine-grained registration between the point clouds to be registered and the current group target point cloud; if the registration result meets the gating condition, then the point cloud to be registered is merged into the current group target point cloud according to the transformation obtained from the registration; if the registration fails or the gating condition is not met, then the temporal fragment and its registration quality index are recorded and reserved for processing in the final supplementary registration stage.

[0068] Subsequently, the group-level aggregated point clouds corresponding to each segment group are registered and fused between groups to obtain a preliminary aggregated point cloud. Inter-group registration and fusion can be performed incrementally in a sequential manner: select one group-level aggregated point cloud as the initial target point cloud, and register other group-level aggregated point clouds sequentially onto this target point cloud; if registration is successful, it is merged; if registration fails, its status is recorded for subsequent processing. The difference between inter-group registration and fusion and intra-group registration and fusion is that the object to be registered changes from a single temporal segment to a group-level aggregated point cloud, and the reference object changes from the group target point cloud to the already formed preliminary aggregated point cloud.

[0069] Finally, the time segments that failed to merge during the intra-group registration and fusion and inter-group registration and fusion processes, as well as the time segments not selected as the main group, are registered to the preliminary aggregated point cloud and fused to obtain the target aggregated point cloud of the target object. This final supplementary registration uses the already formed preliminary aggregated point cloud as a reference, and attempts to register and fuse the remaining time segments one by one. This allows the time segments separated due to gating in the second stage to still have a chance to be pulled back into the final target aggregated point cloud at the spatial level, improving the integrity of the final aggregation result without compromising the purity of the existing point cloud.

[0070] In some embodiments, the selection strategy for the main group can also include indicators such as the number of points, density, overlap, and historical registration success rate; in addition to adopting the sequential incremental method, the inter-group registration fusion can also be implemented by methods such as overlap-based sorting, mask distance-based clustering, or graph optimization.

[0071] Preferably, the intra-group registration fusion, the inter-group registration fusion, and the registration of time segments that were not successfully fused during the intra-group registration fusion and the inter-group registration fusion to the preliminary aggregated point cloud and then fusion are all achieved by performing fine registration between the point cloud to be registered and the preset reference point cloud. The refined registration includes: Based on the point spacing between the point cloud to be registered and the reference point cloud, determine the voxel downsampling scale and the maximum corresponding distance threshold used for the fine registration; Based on the voxel downsampling scale and the maximum corresponding distance threshold, multi-stage point cloud registration is performed on the point cloud to be registered and the reference point cloud. The multi-stage point cloud registration includes initial alignment at a first scale and fine registration at a second scale smaller than the first scale. Each stage prioritizes point-to-surface iterative nearest point registration. When the point-to-surface iterative nearest point registration does not meet the preset registration conditions, point-to-point iterative nearest point registration is used.

[0072] Specifically, the reference point cloud in the above-mentioned fine registration is the target point cloud of the current group in intra-group registration and fusion, the pre-formed aggregated point cloud in inter-group registration and fusion, and the target-level pre-aggregated point cloud in supplementary registration.

[0073] Before determining the voxel downsampling scale and the maximum corresponding distance threshold, the fine-grained registration can first preprocess the point cloud to be registered and the reference point cloud. The preprocessing includes point cloud downsampling, normal vector estimation, point size control, and / or outlier filtering. Then, based on the point spacing between the point cloud to be registered and the reference point cloud, the voxel downsampling scale and the maximum corresponding distance threshold used for fine-grained registration are adaptively determined. Finally, based on the voxel downsampling scale and the maximum corresponding distance threshold, the multi-stage point cloud registration is performed on the point cloud to be registered and the reference point cloud.

[0074] In the multi-stage point cloud registration, initial alignment is first performed at a first scale, followed by fine registration at a second scale smaller than the first scale. Each stage preferentially employs point-to-surface iterative nearest-point registration. If point-to-surface iterative nearest-point registration fails to meet preset registration conditions or encounters an execution anomaly, point-to-point iterative nearest-point registration is used. In some embodiments, for point clouds with insufficient geometric constraints, four-degree-of-freedom registration including heading angle and translation can be used as a supplement.

[0075] After the multi-stage point cloud registration is completed, the registration fit, registration error, translation amount and / or attitude change amount can be calculated, and the reliability of the registration result can be judged according to the preset threshold. If the registration result meets the gating condition, the point cloud to be registered is transformed to the coordinate system of the reference point cloud according to the transformation matrix obtained by registration and then incorporated. After incorporation, the fused point cloud can be downsampled again to control the point cloud size and maintain the efficiency of subsequent registration. If the registration fails or does not meet the gating condition, it is not incorporated for the time being, but its failure status, registration quality index and corresponding transformation information are recorded for subsequent retry or final supplementary registration stage.

[0076] It should be noted that the point cloud registration in this embodiment can be implemented using the iterative nearest point algorithm or other rigid body point cloud registration algorithms. To improve engineering stability, mechanisms such as multi-scale registration, point-to-surface and point-to-point registration backoff, voxel downsampling, registration quality threshold, and pose change threshold can be used in conjunction. The core of this embodiment lies in combining the extraction of the target local point cloud, temporal gating segmentation, spatial occupancy grouping, intra-group registration, inter-group registration, and final supplementary registration into a complete dynamic target point cloud aggregation link, rather than improving the mathematical model of the iterative nearest point algorithm itself. In some embodiments, the spatial occupancy encoding is not limited to an eight-bit spatial occupancy mask, but can also be implemented using more spatial sub-regions, three-dimensional meshes, directional encoding, or other visibility features.

[0077] Based on the above steps S101 to S104, this embodiment can finally output the target aggregated point cloud of the target object, and can also output the point cloud corresponding to each time segment, the correspondence between each time segment and the original frame, the registration transformation relationship from time segment to segment group and from segment group to target, the registration quality index of each stage, and intermediate results that can be used by downstream tasks such as subsequent dynamic target mask generation, target reconstruction or local map generation.

[0078] Accordingly, such as Figure 2 As shown, this invention application also provides a registration system 200 for lidar point clouds, including a conversion module 201, a segmentation module 202, a grouping module 203, and a registration module 204; wherein, The conversion module 201 is used to acquire local point clouds of the target object from the lidar for multiple consecutive frames, and convert the local point clouds of each frame into the same local coordinate system to obtain a point cloud sequence. The segmentation module 202 is used to segment the point cloud sequence according to the time order of each frame, and to segment the point cloud sequence into multiple temporal segments. The grouping module 203 is used to extract the spatial geometric distribution features of each time segment in the local coordinate system; based on the spatial geometric distribution features, it calculates the spatial occupancy mask of each time segment, and divides the time segments with the same spatial occupancy mask into the same group to obtain at least one segment group. The registration module 204 is used to perform registration based on the fragment group to obtain the target aggregate point cloud of the target object.

[0079] As a preferred embodiment, the local coordinate system is a three-dimensional coordinate system, and the local coordinate system is divided into eight subspaces according to the three coordinate axes of the three-dimensional coordinate system; The grouping module 203 extracts the spatial geometric distribution features of each time segment in the local coordinate system; based on the spatial geometric distribution features, it calculates the spatial occupancy mask of each time segment, including: The grouping module 203 extracts the spatial geometric distribution features based on the coordinates of each point in each time segment in the local coordinate system. Based on the spatial geometric distribution characteristics, determine the occupancy status of each subspace; Based on the determination result of the occupancy status, the mask bits corresponding to the subspaces that constitute effective occupancy are marked as one, and the mask bits corresponding to the subspaces that do not constitute effective occupancy are marked as zero, thus obtaining the space occupancy mask of the timing segment consisting of eight mask bits.

[0080] As a preferred embodiment, the spatial geometric distribution features include the number of points falling into the subspace; The grouping module 203 determines the occupancy status of each subspace based on the spatial geometric distribution characteristics, including: When the number of points is greater than or equal to a preset number threshold and the average point spacing of the point cloud in the subspace is less than or equal to a preset density threshold, the grouping module 203 determines that the subspace constitutes effective occupation. If the number of points is less than the number threshold, or the average point spacing of the point cloud within the subspace is greater than the density threshold, the subspace is determined not to constitute effective occupancy.

[0081] As a preferred embodiment, the segmentation module 202 performs segmentation processing on the point cloud sequence according to the temporal order of each frame, dividing the point cloud sequence into multiple temporal segments, including: The segmentation module 202 determines the earliest frame in the point cloud sequence in time sequence as the starting frame of the current segment; According to the time sequence, for each frame in the point cloud sequence after the starting frame, the current frame is initially registered to the current segment in turn, so as to obtain the initial registration result of the current frame relative to the current segment; Based on the preliminary registration results, it is determined whether the current frame meets the merging condition by using preset gating conditions; When the current frame meets the merging condition, the current frame is merged into the current segment; when the current frame does not meet the merging condition, the current segment is determined as a segmented temporal segment, until the processing of each frame in the point cloud sequence is completed, and multiple temporal segments are obtained.

[0082] As a preferred embodiment, the preset gating conditions include at least one of effective matching ratio gating, registration quality gating, translation amount gating, and attitude change amount gating; The effective matching ratio gate is defined as follows: the proportion of effective points in the preliminary registration result reaches a preset ratio threshold; wherein, the proportion of effective points is the proportion of points in the point cloud of the current frame whose nearest neighbor distance to the point cloud of the current segment does not exceed a preset distance threshold. The registration quality gating is defined as follows: the registration error does not exceed a preset error threshold. The translation amount gate is defined as follows: the translation amount required to align the current frame to the current segment does not exceed a preset translation threshold; wherein, the translation amount is determined based on the translation vector in the transformation matrix corresponding to the preliminary registration result; The attitude change gating is as follows: the attitude change generated by the initial registration of the current frame does not exceed a preset attitude threshold; wherein, the attitude change is obtained by decomposing the rotation data of the transformation matrix; The segmentation module 202 determines whether the current frame meets the merging condition based on the preliminary registration result and through preset gating conditions, including: when the preliminary registration result meets all the gating conditions used in the preset gating conditions, the segmentation module 202 determines that the current frame meets the merging condition; otherwise, it determines that the current frame does not meet the merging condition.

[0083] As a preferred embodiment, the registration module 204 performs registration based on the fragment group to obtain the target aggregated point cloud of the target object, including: The registration module 204 performs intra-group registration and fusion of each temporal segment within each segment group to obtain the group-level aggregated point cloud corresponding to the segment group. The group-level aggregated point clouds corresponding to each of the aforementioned fragment groups are registered and fused between groups to obtain a preliminary aggregated point cloud. The time segments that failed to be successfully fused during the intra-group registration and fusion and inter-group registration and fusion processes are registered to the preliminary aggregated point cloud and fused to obtain the target aggregated point cloud of the target object.

[0084] As a preferred embodiment, the intra-group registration and fusion, the inter-group registration and fusion, and the registration and fusion of time segments that were not successfully fused during the intra-group registration and fusion and the inter-group registration and fusion to the preliminary aggregated point cloud are all achieved by performing fine registration between the point cloud to be registered and the preset reference point cloud. The refined registration includes: Based on the point spacing between the point cloud to be registered and the reference point cloud, determine the voxel downsampling scale and the maximum corresponding distance threshold used for the fine registration; Based on the voxel downsampling scale and the maximum corresponding distance threshold, multi-stage point cloud registration is performed on the point cloud to be registered and the reference point cloud. The multi-stage point cloud registration includes initial alignment at a first scale and fine registration at a second scale smaller than the first scale. Each stage prioritizes point-to-surface iterative nearest point registration. When the point-to-surface iterative nearest point registration does not meet the preset registration conditions, point-to-point iterative nearest point registration is used.

[0085] Accordingly, this application also provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the lidar point cloud registration method.

[0086] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal, connecting various parts of the terminal via various interfaces and lines.

[0087] The memory can be used to store the computer program. The processor implements various functions of the terminal by running or executing the computer program stored in the memory and calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0088] Accordingly, this application also provides a computer-readable storage medium, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the lidar point cloud registration method.

[0089] Wherein, if the modules of the device / terminal equipment / system integration are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when the computer program is executed by a processor, it can implement the steps of the various method embodiments described above. Wherein, the computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0090] Compared with the prior art, this invention application has the following beneficial effects: This invention provides a method, system, device, and medium for registering lidar point clouds. By extracting the spatial geometric distribution features of each temporal segment and calculating the spatial occupancy mask, temporal segments with the same spatial occupancy mask are grouped together for registration. This method directly overcomes the limitations of existing solutions that rely entirely on the sequential order of the timeline for blind stitching. By grouping segments using spatial occupancy masks, it ensures that segments with similar spatial occupancy can be registered according to the resulting groups. This fundamentally avoids defects such as local geometric feature misalignment and distortion caused by forcibly registering temporally adjacent segments with large differences in spatial occupancy patterns when dynamic targets are subjected to significant viewpoint rotation or severe occlusion. This significantly improves the accuracy of lidar point cloud registration.

[0091] Furthermore, by first segmenting the point cloud sequence according to time sequence to isolate temporally unstable observations, and then regrouping multiple temporal segments spatially using spatial occupancy masks, this technique of first temporally segmenting and then spatially reorganizing allows for the accurate identification and aggregation of scattered segments with the same or similar spatial occupancy patterns even when dynamic targets cause temporal discontinuities or sudden changes in viewpoint in the point cloud observation due to violent movement or severe occupancy on the time axis. This significantly improves the robustness and scene adaptability of the registration method in complex, highly dynamic autonomous driving scenarios.

[0092] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A registration method for lidar point clouds, characterized in that, include: The local point cloud of the target object is obtained from multiple consecutive frames by the lidar, and the local point cloud of each frame is transformed into the same local coordinate system to obtain a point cloud sequence. The point cloud sequence is segmented according to the time order of each frame, and the point cloud sequence is divided into multiple temporal segments. Extract the spatial geometric distribution features of each of the time segments in the local coordinate system; Based on the spatial geometric distribution characteristics, the spatial occupancy mask of each time segment is calculated, and time segments with the same spatial occupancy mask are grouped into the same group to obtain at least one segment group. Based on the fragment group, registration is performed to obtain the target aggregate point cloud of the target object; The process of registering the target object based on the fragment group to obtain the target aggregate point cloud includes: For each of the aforementioned fragment groups, the temporal fragments within the fragment group are registered and fused within the group to obtain the group-level aggregated point cloud corresponding to the fragment group. The group-level aggregated point clouds corresponding to each of the aforementioned fragment groups are registered and fused between groups to obtain a preliminary aggregated point cloud. The time segments that failed to be successfully fused during the intra-group registration and fusion and inter-group registration and fusion processes are registered to the preliminary aggregated point cloud and fused to obtain the target aggregated point cloud of the target object.

2. The registration method for lidar point clouds as described in claim 1, characterized in that, The local coordinate system is a three-dimensional coordinate system, and the local coordinate system is divided into eight subspaces according to the three coordinate axes of the three-dimensional coordinate system; The spatial geometric distribution features of each of the time segments in the local coordinate system are extracted; Based on the aforementioned spatial geometric distribution characteristics, the spatial occupancy mask for each of the time segments is calculated, including: Based on the coordinates of each point in each time segment in the local coordinate system, the spatial geometric distribution features are extracted. Based on the spatial geometric distribution characteristics, determine the occupancy status of each subspace; Based on the determination result of the occupancy status, the mask bits corresponding to the subspaces that constitute effective occupancy are marked as one, and the mask bits corresponding to the subspaces that do not constitute effective occupancy are marked as zero, thus obtaining the space occupancy mask of the timing segment consisting of eight mask bits.

3. The registration method for lidar point clouds as described in claim 2, characterized in that, The spatial geometric distribution characteristics include the number of points falling into the subspace; The step of determining the occupancy status of each subspace based on the spatial geometric distribution characteristics includes: When the number of points is greater than or equal to a preset number threshold, and the average point spacing of the point cloud within the subspace is less than or equal to a preset density threshold, the subspace is determined to constitute effective occupancy. If the number of points is less than the number threshold, or the average point spacing of the point cloud within the subspace is greater than the density threshold, the subspace is determined not to constitute effective occupancy.

4. The registration method for lidar point clouds as described in claim 1, characterized in that, The point cloud sequence is segmented according to the temporal order of each frame, dividing the point cloud sequence into multiple temporal segments, including: The earliest frame in the point cloud sequence in chronological order is determined as the starting frame of the current segment; According to the time sequence, for each frame in the point cloud sequence after the starting frame, the current frame is initially registered to the current segment in turn, so as to obtain the initial registration result of the current frame relative to the current segment; Based on the preliminary registration results, it is determined whether the current frame meets the merging condition by using preset gating conditions; When the current frame meets the merging condition, the current frame is merged into the current segment; when the current frame does not meet the merging condition, the current segment is determined as a segmented temporal segment, until the processing of each frame in the point cloud sequence is completed, and multiple temporal segments are obtained.

5. The registration method for lidar point clouds as described in claim 4, characterized in that, The preset gating conditions include at least one of effective matching ratio gating, registration quality gating, translation amount gating, and attitude change amount gating. The effective matching ratio gate is defined as follows: the proportion of effective points in the preliminary registration result reaches a preset ratio threshold; wherein, the proportion of effective points is the proportion of points in the point cloud of the current frame whose nearest neighbor distance to the point cloud of the current segment does not exceed a preset distance threshold. The registration quality gating is: the registration error does not exceed a preset error threshold; The translation amount gate is defined as follows: the translation amount required to align the current frame to the current segment does not exceed a preset translation threshold; wherein, the translation amount is determined based on the translation vector in the transformation matrix corresponding to the preliminary registration result; The attitude change gating is as follows: the attitude change generated by the initial registration of the current frame does not exceed a preset attitude threshold; wherein, the attitude change is obtained by decomposing the rotation data of the transformation matrix; The step of determining whether the current frame meets the merging condition based on the preliminary registration result and using preset gating conditions includes: determining that the current frame meets the merging condition when the preliminary registration result meets all the gating conditions used in the preset gating conditions; otherwise, determining that the current frame does not meet the merging condition.

6. The registration method for lidar point clouds as described in claim 1, characterized in that, The intra-group registration and fusion, the inter-group registration and fusion, and the registration and fusion of time segments that were not successfully fused during the intra-group registration and fusion and the inter-group registration and fusion to the preliminary aggregated point cloud are all achieved by performing fine registration between the point cloud to be registered and the preset reference point cloud. The refined registration includes: Based on the point spacing between the point cloud to be registered and the reference point cloud, determine the voxel downsampling scale and the maximum corresponding distance threshold used for the fine registration; Based on the voxel downsampling scale and the maximum corresponding distance threshold, multi-stage point cloud registration is performed on the point cloud to be registered and the reference point cloud. The multi-stage point cloud registration includes initial alignment at a first scale and fine registration at a second scale smaller than the first scale. Each stage prioritizes point-to-surface iterative nearest point registration. When the point-to-surface iterative nearest point registration does not meet the preset registration conditions, point-to-point iterative nearest point registration is used.

7. A registration system for lidar point clouds, characterized in that, It includes a conversion module, a segmentation module, a grouping module, and a registration module; among which, The conversion module is used to acquire multiple consecutive frames of local point clouds of the target object from the lidar, and convert the local point clouds of each frame into the same local coordinate system to obtain a point cloud sequence. The segmentation module is used to segment the point cloud sequence into multiple temporal segments according to the temporal order of each frame. The grouping module is used to extract the spatial geometric distribution features of each time segment in the local coordinate system; based on the spatial geometric distribution features, it calculates the spatial occupancy mask of each time segment, and divides time segments with the same spatial occupancy mask into the same group to obtain at least one segment group. The registration module is used to perform registration based on the fragment group to obtain the target aggregate point cloud of the target object; The registration module performs registration based on the fragment group to obtain the target aggregated point cloud of the target object, including: The registration module performs intra-group registration and fusion of each temporal segment within each segment group to obtain a group-level aggregated point cloud corresponding to the segment group. The group-level aggregated point clouds corresponding to each of the aforementioned fragment groups are registered and fused between groups to obtain a preliminary aggregated point cloud. The time segments that failed to be successfully fused during the intra-group registration and fusion and inter-group registration and fusion processes are registered to the preliminary aggregated point cloud and fused to obtain the target aggregated point cloud of the target object.

8. The registration system for lidar point clouds as described in claim 7, characterized in that, The intra-group registration and fusion, the inter-group registration and fusion, and the registration and fusion of time segments that were not successfully fused during the intra-group registration and fusion and the inter-group registration and fusion to the preliminary aggregated point cloud are all achieved by performing fine registration between the point cloud to be registered and the preset reference point cloud. The refined registration includes: Based on the point spacing between the point cloud to be registered and the reference point cloud, determine the voxel downsampling scale and the maximum corresponding distance threshold used for the fine registration; Based on the voxel downsampling scale and the maximum corresponding distance threshold, multi-stage point cloud registration is performed on the point cloud to be registered and the reference point cloud. The multi-stage point cloud registration includes initial alignment at a first scale and fine registration at a second scale smaller than the first scale. Each stage prioritizes point-to-surface iterative nearest point registration. When the point-to-surface iterative nearest point registration does not meet the preset registration conditions, point-to-point iterative nearest point registration is used.

9. A terminal device, characterized in that, The system includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the registration method for lidar point clouds as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the registration method for lidar point clouds as described in any one of claims 1 to 6.

Images