CN-121999002-A - Laser radar point cloud segmentation method and system

CN121999002ACN 121999002 ACN121999002 ACN 121999002ACN-121999002-A

Abstract

The invention relates to the technical field of laser radar point cloud processing, and discloses a laser radar point cloud segmentation method and a laser radar point cloud segmentation system. The system generates a topological graph of the incomplete part by detecting a non-rigid body example with a damage mark, matching a skeleton template, calculates a part motion parameter set, deduces a incomplete part motion assumption based on a kinematic coupling rule, executes layered motion compensation to construct deformation normalized complete point cloud representation, and finally generates a damaged non-rigid body target point cloud segmentation result through a point cloud segmentation network, thereby solving the problem of non-rigid body motion target segmentation identification with incomplete structure in a road scene.

Inventors

ZHENG KANG
LIU LUNXU
WANG BINGWEI
LIU SHIWANG

Assignees

浙江埃科汽车技术服务有限公司

Dates

Publication Date: 20260508
Application Date: 20260409

Claims (10)

1. The laser radar point cloud segmentation method is characterized by comprising the following steps of: acquiring a time sequence point cloud sequence acquired by a laser radar, executing non-rigid body instance detection on the time sequence point cloud sequence, detecting boundary positions generated by geometric discontinuities in each instance, marking the detected boundary positions as potential damage boundaries, and generating a non-rigid body instance set with damage marks; extracting geometric features and spatial positions of visible components from non-rigid body examples with damage marks, matching the geometric features and the spatial positions with a preset skeleton template, and identifying positions and types of missing components according to matching results to generate a topology diagram of the missing components; Calculating inter-frame motion vectors between adjacent frames for the visible part areas, dividing point clouds in the examples into a plurality of motion partitions based on the directional similarity and the amplitude similarity of the inter-frame motion vectors, carrying out semantic association on the motion partitions and visible part nodes in the incomplete part topological graph, and generating a part motion parameter set; Based on the skeleton connection relation in the incomplete part topological graph, extracting the motion parameters of adjacent visible parts from a part motion parameter set, deducing the motion range and track envelope of the missing parts according to the human body motion coupling rule, and generating a incomplete part motion hypothesis; Performing layered motion compensation processing on the visible point cloud by taking the motion parameters of the trunk part as a reference to restore the visible point cloud to a standard posture, generating a virtual occupied point cloud according to the motion assumption of the incomplete part in the missing area, merging the visible point cloud of the standard posture with the virtual occupied point cloud, and constructing a deformation normalized complete point cloud representation; and outputting a point-by-point type prediction result by inputting the deformation normalized whole point cloud representation into a point cloud segmentation network, mapping the point-by-point type prediction result back to an original motion state, and generating a damaged non-rigid body target point cloud segmentation result.
2. The method for partitioning a point cloud of a laser radar according to claim 1, wherein the geometric discontinuity is determined by a geometric continuity indicator, the geometric continuity indicator is obtained by calculating a weighted sum of a normal vector included angle mean value and an inter-point distance standard deviation of a local area of the point cloud, and when the geometric continuity indicator is lower than a preset threshold, the corresponding boundary is marked as a potential damage boundary, and the preset threshold takes a quantile of geometric continuity indicator statistical distribution of a normal target point cloud.
3. The method for partitioning the point cloud of the lidar according to claim 1, wherein the non-rigid body instance detection adopts a method based on motion clustering, performs point cloud registration on adjacent frames in a sequence of time-domain point clouds, calculates a displacement vector of each point between the adjacent frames, clusters the point clouds according to the direction and the amplitude of the displacement vector, clusters the points with similar motion characteristics into one instance, and judges whether the instance is a non-rigid body object according to whether a plurality of sub-regions with different motion directions exist in the instance.
4. The laser radar point cloud segmentation method according to claim 1 is characterized in that the skeleton template matching adopts a partial matching strategy, wherein visible parts are used as anchor points, geometric similarity between a visible part shape descriptor and a corresponding part shape descriptor in the skeleton template is calculated through cosine similarity, when the geometric similarity exceeds a matching threshold, a matching relation is established, and the position and type of a missing part are determined according to the positions of parts in the skeleton template, which are adjacent to the matched parts but have no corresponding areas in the point cloud.
5. The method for partitioning the point cloud of the laser radar according to claim 4, wherein the extraction of the visible component is achieved by partitioning a non-rigid body example point cloud, the example point cloud is partitioned into a plurality of connected areas according to the spatial continuity and the normal vector consistency of the point cloud, each connected area corresponds to one visible component, the shape descriptor is calculated by adopting a point characteristic histogram method, a combined characteristic of a normal vector included angle, a distance between points and a relative direction between each point and a neighborhood point of each point in the point cloud of the component is calculated, and the combined characteristic is counted in a characteristic space to generate a multi-dimensional histogram.
6. The method for partitioning the point cloud of the laser radar according to claim 1, wherein the calculation of the inter-frame motion vector adopts a nearest point matching method, each point in a current frame searches a point closest to a point cloud of a previous frame for serving as a corresponding point, the difference between the coordinates of the point of the current frame and the coordinates of the corresponding point serves as the inter-frame motion vector of the point, the partitioning of the motion partition adopts a clustering method based on graph cut, the points in an instance are constructed into a graph structure, the weight of an edge is calculated according to the similarity of the motion vector between the two points, and the graph structure is partitioned by adopting a normalized graph cut algorithm to obtain the motion partition.
7. The laser radar point cloud segmentation method according to claim 1, wherein the inference of the motion assumption of the incomplete component adopts a constraint propagation method, wherein the motion parameters of adjacent visible components are used as boundary conditions, joint freedom degree constraints defined according to a skeleton connection relation are used for calculating motion parameter value ranges of the incomplete component in respective freedom degree directions, a central value of the motion parameter value ranges is used as an estimated motion parameter, the size of the motion parameter value ranges is converted into confidence degrees through linear mapping, and the confidence degrees are higher when the motion parameter value ranges are smaller.
8. The laser radar point cloud segmentation method according to claim 1 is characterized in that the generation of the virtual occupation point cloud is achieved by adopting a motion constraint sampling method, wherein the spatial distribution density of virtual points is determined according to geometric parameters of corresponding parts in a skeleton template in a motion range defined by motion assumption of incomplete parts, the spatial distribution density is calculated according to volume information in the geometric parameters and preset unit volume points, and the virtual occupation points are generated by evenly sampling in the motion range according to the spatial distribution density.
9. The method for partitioning the point cloud of the laser radar according to claim 1, further comprising the step of complementing the geometry of the fracture primitive, wherein after the non-rigid body instance set with the damage marks is generated, the geometric attribute continuity of the point cloud areas at the two sides of the potential damage boundary is analyzed, the geometric primitive parameters of the point cloud areas at the two sides of the potential damage boundary are calculated through the RANSAC algorithm in a fitting mode, when the geometric primitives obtained through fitting at the two sides of the potential damage boundary are identical in type and the difference of the geometric primitive parameters is smaller than a preset threshold, the fracture primitive is judged to be broken, virtual extension estimation is performed on the area judged to be broken based on the geometric primitive parameters to generate fracture primitive complementing parameters, and the fracture primitive complementing parameters are used for assisting the generation of the virtual occupied point cloud.
10. A lidar point cloud segmentation system for performing the lidar point cloud segmentation method of any of claims 1 to 9, comprising: The damage example detection module is used for acquiring a time sequence point cloud sequence acquired by the laser radar, executing non-rigid example detection on the time sequence point cloud sequence, detecting boundary positions generated by geometric discontinuities in each example and marking the boundary positions as potential damage boundaries, and generating a non-rigid example set with damage marks; The incomplete topology generation module is used for extracting geometric features and spatial positions of the visible components from the non-rigid body examples with the damage marks, matching the geometric features and the spatial positions with a preset skeleton template, and identifying the positions and types of the missing components according to the matching results to generate an incomplete component topology diagram; The motion parameter calculation module is used for calculating an inter-frame motion vector between adjacent frames for the visible part area, dividing an instance inner point cloud into a plurality of motion partitions based on the inter-frame motion vector, carrying out semantic association on the motion partitions and visible part nodes in the incomplete part topological graph, and generating a part motion parameter set; The motion assumption deducing module is used for extracting the motion parameters of adjacent visible components from the component motion parameter set based on the skeleton connection relation in the incomplete component topological graph, deducing the motion range and the track envelope of the missing component according to the human body motion coupling rule, and generating the motion assumption of the incomplete component; The deformation normalization module is used for performing layered motion compensation processing on the visible point cloud by taking the motion parameters of the trunk part as a reference to restore the visible point cloud to the standard posture, generating a virtual occupied point cloud according to the motion assumption of the incomplete part in the missing area, merging the visible point cloud of the standard posture with the virtual occupied point cloud, and constructing a deformation normalization complete point cloud representation; the segmentation result generation module is used for outputting a point-by-point type prediction result by inputting the deformation normalized whole point cloud representation into the point cloud segmentation network, mapping the point-by-point type prediction result back to the original motion state, and generating a damaged non-rigid body target point cloud segmentation result.

Description

Laser radar point cloud segmentation method and system Technical Field The invention relates to the technical field of laser radar point cloud processing, in particular to a laser radar point cloud segmentation method and a laser radar point cloud segmentation system. Background In the fields of automatic driving and intelligent traffic, the laser radar is widely applied to road environment sensing, and dynamic targets such as pedestrians, riders and the like are segmented and identified by collecting point cloud data. The existing laser radar point cloud segmentation method is usually trained and inferred based on a complete object model, and dynamic target classification is realized by relying on complete geometric structures and consistent motion characteristics of targets. However, there are non-rigid moving objects in real road scenes that are partially damaged or structurally incomplete, such as cyclists lacking a mudguard, pedestrians using damaged wheelchairs, special pedestrians with incomplete limbs, etc. For the targets, the prior art faces double dilemma that on one hand, a non-rigid motion analysis method based on a skeleton model relies on complete skeleton topology to establish motion association among components, the missing components lead to skeleton topology fracture and cannot establish an effective motion model, and on the other hand, a point cloud complement method based on geometric continuity relies on static assumption and cannot process dynamic deformation in a motion process. Therefore, the dynamic incomplete target cannot be built into a motion model or geometrically completed, so that a segmentation blind area is formed, and the target is missed to be detected or misclassified, so that the safety of an automatic driving system is influenced. Disclosure of Invention The invention provides a laser radar point cloud segmentation method and a laser radar point cloud segmentation system, which solve the technical problem that geometric defects and motion deformation cannot be accurately processed when point cloud segmentation is carried out on a damaged non-rigid object in the related technology. The invention discloses a laser radar point cloud segmentation method, which comprises the following steps of acquiring a time sequence point cloud sequence acquired by a laser radar, executing non-rigid body instance detection on the time sequence point cloud sequence, detecting boundary positions inside each instance due to geometric discontinuity, marking the detected boundary positions as potential damage boundaries, and generating a non-rigid body instance set with damage marks; extracting geometric features and spatial positions of visible parts from non-rigid body examples with damage marks, matching with a preset skeleton template, identifying the positions and types of missing parts according to matching results, generating a defective part topological graph, calculating interframe motion vectors between adjacent frames for a visible part area, dividing an example internal point cloud into a plurality of motion partitions based on the directional similarity and the amplitude similarity of the interframe motion vectors, semantically associating the motion partitions with visible part nodes in the defective part topological graph, generating a part motion parameter set, extracting motion parameters of adjacent visible parts from the part motion parameter set based on skeleton connection relations in the defective part topological graph, deducing the motion range and track envelope of the missing parts according to a human body motion coupling rule, generating a defective part motion hypothesis, performing hierarchical motion compensation processing on the visible point cloud by taking the trunk part motion parameters as references to restore the visible point cloud to a standard gesture, generating a virtual occupation point cloud for the missing area according to the defective part motion hypothesis, combining the visible point cloud and the virtual occupation point cloud of the standard gesture, constructing a deformation normalization point cloud representation, outputting a normalization point cloud representation, mapping the normalization point cloud representation into a prediction result, and outputting a prediction point cloud prediction result, namely a prediction result, a prediction point class, and a prediction result, namely a prediction result, and a prediction result, which are obtained by a prediction result, and generating a cloud segmentation result of the damaged non-rigid body target point. Further, the geometric discontinuity is judged through a geometric continuity index, the geometric continuity index is obtained through calculating a weighted sum of a normal vector included angle mean value and an inter-point standard deviation of a local area of the point cloud, when the geometric continuity index is lower than a preset threshold, the corresponding boundary i