A method and system for segmenting lidar point clouds

Abstract

The application relates to the technical field of laser radar point cloud processing, and discloses a laser radar point cloud segmentation method and system. The system detects a non-rigid instance with a damage mark, matches a skeleton template to generate a defective component topology graph, calculates a component motion parameter set, infers a defective component motion hypothesis based on a kinematic coupling rule, performs layered motion compensation to construct a deformation normalized complete point cloud representation, and finally generates a damaged non-rigid target point cloud segmentation result through a point cloud segmentation network, so that the segmentation and identification problem of a non-rigid moving target with an incomplete structure in a road scene are solved.

Description

Technical Field

[0001] This invention relates to the field of lidar point cloud processing technology, and more specifically, to a lidar point cloud segmentation method and system. Background Technology

[0002] In the fields of autonomous driving and intelligent transportation, LiDAR is widely used for road environment perception, segmenting and identifying dynamic targets such as pedestrians and cyclists by collecting point cloud data. Existing LiDAR point cloud segmentation methods are usually based on training and inference using complete object models, relying on the complete geometric structure and consistent motion characteristics of the target to achieve dynamic target classification.

[0003] However, real-world road scenarios often involve partially damaged or structurally incomplete non-rigid moving targets, such as cyclists missing mudguards, pedestrians using damaged wheelchairs, and pedestrians with limb disabilities. Existing technologies face a dual dilemma for such targets: firstly, skeleton-based non-rigid motion analysis methods rely on establishing motion relationships between components using a complete skeleton topology; missing components cause topological breaks in the skeleton, making it impossible to build an effective motion model. Secondly, point cloud completion methods based on geometric continuity rely on static assumptions and cannot handle dynamic deformations during motion. Therefore, for dynamically incomplete targets, neither motion models can be built nor geometric completion can be performed, creating segmentation blind spots, leading to missed detections or misclassifications, and impacting the safety of autonomous driving systems. Summary of the Invention

[0004] This invention provides a lidar point cloud segmentation method and system, which solves the technical problem in related technologies that cannot accurately handle geometric missing and motion deformation when segmenting point clouds of damaged non-rigid targets.

[0005] This invention discloses a point cloud segmentation method for LiDAR, comprising the following steps: acquiring a temporal point cloud sequence collected by LiDAR; performing non-rigid instance detection on the temporal point cloud sequence; detecting the boundary positions generated by geometric discontinuities within each instance; marking the detected boundary positions as potential damaged boundaries; generating a set of non-rigid instances with damage markings; extracting the geometric features and spatial positions of visible components from the non-rigid instances with damage markings; matching them with a preset skeleton template; identifying the position and type of missing components based on the matching results; generating a topology map of incomplete components; calculating inter-frame motion vectors between adjacent frames for the visible component region; dividing the point cloud within the instance into multiple motion partitions based on the direction similarity and amplitude similarity of the inter-frame motion vectors; and matching the motion partitions with the visible parts in the topology map of incomplete components. Semantic association is performed on component nodes to generate a set of component motion parameters. Based on the skeleton connection relationship in the topology graph of the incomplete component, the motion parameters of adjacent visible components are extracted from the set of component motion parameters. The motion range and trajectory envelope of the missing component are inferred according to the coupling law of human kinematics, and the motion hypothesis of the incomplete component is generated. Based on the motion parameters of the torso component, the visible point cloud is subjected to hierarchical motion compensation processing to restore it to the standard posture. Virtual occupancy point clouds are generated for the missing region according to the motion hypothesis of the incomplete component. The visible point cloud in the standard posture is merged with the virtual occupancy point cloud to construct a deformation-normalized complete point cloud representation. The deformation-normalized complete point cloud representation is input into the point cloud segmentation network to output the point-by-point category prediction result. The point-by-point category prediction result is mapped back to the original motion state to generate the point cloud segmentation result of the damaged non-rigid body target.

[0006] Furthermore, the geometric discontinuity is determined by a geometric continuity index, which is obtained by calculating the weighted sum of the mean angle between the normal vectors of a local region of the point cloud and the standard deviation of the distance between points. When the geometric continuity index is lower than a preset threshold, the corresponding boundary is marked as a potential damaged boundary. The preset threshold is the quantile of the statistical distribution of the geometric continuity index of a normal target point cloud.

[0007] Furthermore, the non-rigid instance detection adopts a motion clustering-based method, which performs point cloud registration on adjacent frames in the temporal point cloud sequence, calculates the displacement vector of each point between adjacent frames, clusters the point cloud according to the direction and magnitude of the displacement vector, and clusters points with similar motion characteristics into one instance. The instance is determined to be a non-rigid target based on whether there are multiple sub-regions with different motion directions inside the instance.

[0008] Furthermore, the skeleton template matching adopts a partial matching strategy: visible parts are used as anchor points, and the geometric similarity between the shape descriptor of the visible part and the shape descriptor of the corresponding part in the skeleton template is calculated by cosine similarity; when the geometric similarity exceeds the matching threshold, a matching relationship is established; based on the position of the part in the skeleton template that is adjacent to the matched part but does not have a corresponding region in the point cloud, the position and type of the missing part are determined.

[0009] Furthermore, the extraction of the visible component is achieved by region segmentation of the non-rigid instance point cloud. Based on the spatial continuity and normal vector consistency of the point cloud, the instance point cloud is divided into multiple connected regions, and each connected region corresponds to a visible component. The shape descriptor is calculated using the point feature histogram method. For each point in the component point cloud, the combined features of the normal vector angle between the point and its neighboring points, the distance between points, and the relative direction are calculated. The combined features are statistically analyzed in the feature space to generate a multidimensional histogram.

[0010] Furthermore, the calculation of the inter-frame motion vector adopts the nearest point matching method. For each point in the current frame, the nearest point in the point cloud of the previous frame is searched as the corresponding point, and the difference between the coordinates of the current frame point and the coordinates of the corresponding point is taken as the inter-frame motion vector of that point. The division of the motion partition adopts the graph cut-based clustering method. The points in the instance are constructed into a graph structure, the weight of the edge is calculated according to the similarity of the motion vectors between two points, and the normalized graph cut algorithm is used to segment the graph structure to obtain the motion partition.

[0011] Furthermore, the inference of the motion hypothesis of the missing part adopts the constraint propagation method: the motion parameters of adjacent visible parts are used as boundary conditions; according to the joint degree of freedom constraints defined by the skeleton connection relationship, the range of motion parameter values ​​of the missing part in each degree of freedom direction is calculated; the center value of the range of motion parameter values ​​is used as the estimated motion parameter; the size of the range of motion parameter values ​​is converted into confidence through linear mapping, and the smaller the range of motion parameter values, the higher the confidence.

[0012] Furthermore, the generation of the virtual occupant point cloud adopts a motion constraint sampling method: within the motion range defined by the motion assumption of the incomplete component, the spatial distribution density of virtual points is determined according to the geometric parameters of the corresponding component in the skeleton template. The spatial distribution density is calculated based on the volume information in the geometric parameters and the preset number of points per unit volume. Virtual occupant points are generated by uniformly sampling within the motion range according to the spatial distribution density.

[0013] Furthermore, the method also includes the step of completing the fracture primitive geometry: after generating a set of non-rigid body instances with damage markers, the geometric property continuity of the point cloud regions on both sides of the potential damage boundary is analyzed. The geometric primitive parameters of the point cloud regions on both sides of the potential damage boundary are calculated by fitting using the RANSAC algorithm. When the geometric primitive types fitted on both sides of the potential damage boundary are the same and the difference in geometric primitive parameters is less than a preset threshold, the fracture is determined to be a fracture of the same geometric primitive. For the region determined to be a fracture, virtual extension estimation is performed based on the geometric primitive parameters to generate fracture primitive completion parameters. The fracture primitive completion parameters are used to assist in the generation of virtual occupancy point clouds.

[0014] This invention discloses a lidar point cloud segmentation system, comprising: a damaged instance detection module, used to acquire a temporal point cloud sequence collected by lidar, perform non-rigid instance detection on the temporal point cloud sequence, detect the boundary positions generated by geometric discontinuities within each instance and mark them as potential damaged boundaries, and generate a set of non-rigid instances with damage markings; a fragmented topology generation module, used to extract the geometric features and spatial positions of visible parts from the non-rigid instances with damage markings, match them with a preset skeleton template, identify the position and type of missing parts based on the matching results, and generate a fragmented part topology map; and a motion parameter calculation module, used to calculate the inter-frame motion vector between adjacent frames for the visible part region, divide the point cloud within the instance into multiple motion partitions based on the inter-frame motion vectors, and semantically associate the motion partitions with the visible part nodes in the fragmented part topology map to generate... The system comprises the following modules: a set of component motion parameters; a motion hypothesis inference module, which extracts motion parameters of adjacent visible components from the set of component motion parameters based on the skeleton connection relationship in the topology graph of the incomplete component, infers the motion range and trajectory envelope of the missing component according to the coupling law of human kinematics, and generates motion hypotheses for the incomplete component; a deformation normalization module, which performs layered motion compensation processing on the visible point cloud based on the motion parameters of the torso component to restore it to the standard posture, generates virtual occupancy point clouds for the missing region according to the motion hypotheses of the incomplete component, merges the visible point cloud of the standard posture with the virtual occupancy point cloud, and constructs a deformation-normalized complete point cloud representation; and a segmentation result generation module, which inputs the deformation-normalized complete point cloud representation into the point cloud segmentation network and outputs point-by-point category prediction results, maps the point-by-point category prediction results back to the original motion state, and generates the point cloud segmentation results of the damaged non-rigid body target.

[0015] This invention establishes a skeletal association between the missing region and the visible region through a topological map of the incomplete component. It infers the motion hypothesis of the incomplete component from the motion parameters of the visible component using the coupling law of human kinematics. It normalizes the dynamic deformation point cloud to a standard posture through layered motion compensation and provides occupancy information in the missing region through a virtual occupancy point cloud. This invention solves the technical problem that dynamic incomplete targets cannot establish a complete motion model or perform static geometric completion. It achieves the technical effect of being able to segment and identify non-rigid moving targets with incomplete structures in road scenes, and reduce target omissions and classification errors. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method for segmenting point clouds of damaged non-rigid targets provided in an embodiment of the present invention;

[0017] Figure 2 This is a sub-flowchart of matching skeleton templates and generating a topology map of incomplete components in the point cloud segmentation method for damaged non-rigid targets provided in this embodiment of the invention;

[0018] Figure 3 This is a sub-flowchart of the method for segmenting point clouds of damaged non-rigid targets provided in this embodiment of the invention, which infers the motion assumption of the missing parts based on the kinematic coupling law;

[0019] Figure 4 This is a sub-flowchart of virtual occupant point cloud generation in the damaged non-rigid target point cloud segmentation method provided in the embodiments of the present invention. Detailed Implementation

[0020] In the fields of autonomous driving and intelligent transportation, LiDAR is widely used for road environment perception, segmenting and identifying dynamic targets such as pedestrians and cyclists by collecting point cloud data. Existing LiDAR point cloud segmentation methods are usually based on training and inference using complete object models, relying on the complete geometric structure and consistent motion characteristics of the target to achieve dynamic target classification.

[0021] However, real-world road scenarios often involve partially damaged or structurally incomplete non-rigid moving targets, such as cyclists missing mudguards, pedestrians using damaged wheelchairs, and pedestrians with limb disabilities. Existing technologies face a dual dilemma for such targets: firstly, skeleton-based non-rigid motion analysis methods rely on establishing motion relationships between components using a complete skeleton topology; missing components cause topological breaks in the skeleton, making it impossible to build an effective motion model. Secondly, point cloud completion methods based on geometric continuity rely on static assumptions and cannot handle dynamic deformations during motion. Therefore, for dynamically incomplete targets, neither motion models can be built nor geometric completion can be performed, creating segmentation blind spots, leading to missed detections or misclassifications, and impacting the safety of autonomous driving systems.

[0022] like Figure 1-4The lidar point cloud segmentation method of this embodiment includes the following steps:

[0023] Step 1: Acquire temporal point clouds and detect non-rigid body instances with damage markers;

[0024] Acquire the temporal point cloud sequence collected by LiDAR, perform non-rigid instance detection on the temporal point cloud sequence to identify dynamic target instances such as pedestrians or cyclists, and detect the boundary positions inside each instance caused by geometric discontinuities. Mark the detected boundary positions as potential damaged boundaries and generate a set of non-rigid instances with damage markings.

[0025] It should be noted that the aforementioned non-rigid instance detection refers to the detection of targets exhibiting non-rigid motion characteristics in point clouds. These non-rigid motion characteristics include relative motion between local regions of the point cloud. The aforementioned geometric discontinuity refers to abnormal breaks in the point cloud in space, manifested as abrupt changes in point density or geometric curvature between adjacent regions. The aforementioned detection of potential damage boundaries is achieved by calculating the geometric continuity index of local regions of the point cloud. When the geometric continuity index is lower than a preset threshold, the corresponding boundary is marked as a potential damage boundary.

[0026] Furthermore, the geometric continuity index is obtained by calculating the angle between the normal vectors and the rate of change of the distance between points in a local region of the point cloud. Specifically, for each point in the point cloud, the mean of the angle between the normal vectors between the point and its neighboring points and the standard deviation of the distance between the points are calculated. The weighted sum of the mean angle between the normal vectors and the standard deviation of the distance between the points is used as the geometric continuity index. The preset threshold is determined based on the statistical distribution of the geometric continuity index of the normal target point cloud, and the 95th percentile of the statistical distribution is taken as the preset threshold.

[0027] Furthermore, the non-rigid instance detection adopts a motion clustering-based method, which is as follows: First, point cloud registration is performed on adjacent frames in the temporal point cloud sequence, and the displacement vector of each point between adjacent frames is calculated. Then, the point cloud is clustered according to the direction and magnitude of the displacement vector, and points with similar motion characteristics are clustered into one instance. Finally, the spatial distribution and motion characteristics of the points in the clustering results are used to determine whether the instance is a non-rigid target. The judgment criterion is that there are multiple sub-regions with different motion directions inside the instance.

[0028] An autonomous driving test vehicle was conducting environmental perception tests at an urban road intersection. The vehicle's onboard 64-line LiDAR collected point cloud data of the surrounding environment at a frequency of 10Hz. At 10:23 AM on March 15, 20XX, the LiDAR detected a cyclist crossing a pedestrian crossing. The cyclist's bicycle was missing its right mudguard, causing an abnormal gap in the point cloud between the rear wheel and the frame. The system acquired three consecutive frames of point cloud data, with an interval of 0.1 seconds between frames.

[0029] In frame 127 of the point cloud data, the system first performs non-rigid instance detection on the point cloud. By calculating the displacement vector of the point cloud between adjacent frames, it identifies instances located at coordinates... A moving target instance exists nearby, with a point cloud containing 2847 points. The geometric continuity index of this instance's point cloud shows that the mean angle of the normal vector detected in the region to the right of the bicycle's rear wheel reaches 68.3 degrees, and the standard deviation of the distance between points is 0.42 meters. The geometric continuity index is [insert index here]. The value exceeds the preset threshold of 35.0, therefore the boundary of this area is marked as a potentially damaged boundary.

[0030] Table 1. Basic information on non-rigid body instances detected in frame 127:

[0031]

[0032] Table 2. Detection results of potential damage boundaries:

[0033]

[0034] Step 2: Match the skeleton template and generate a topology diagram of the incomplete parts;

[0035] For non-rigid body instances with damage markers, extract the geometric features and spatial positions of visible components, match the geometric features and spatial positions of visible components with a preset skeleton template, identify the position and type of missing components based on the matching results, and generate a topology map of the incomplete components.

[0036] It should be noted that the aforementioned skeleton templates include human skeleton templates and human-vehicle combined skeleton templates. The human skeleton template defines the topological connections between various parts of the human body, while the human-vehicle combined skeleton template defines the combined topological relationships between the cyclist and the vehicle. The aforementioned geometric features include the component's dimensional parameters, shape descriptors, and relative positional relationships. The aforementioned incomplete component topology graph refers to graph structure data that records visible component nodes, missing component nodes, and the skeleton connections between nodes.

[0037] In this embodiment, the skeleton template matching adopts a partial matching strategy, specifically:

[0038] Step 2.1: Using the visible parts as anchor points, calculate the geometric similarity between the visible parts and the corresponding parts in the skeleton template. The geometric similarity is calculated by cosine similarity to determine the degree of similarity between the shape descriptor of the visible parts and the shape descriptor of the template parts.

[0039] Step 2.2: Establish a matching relationship when the geometric similarity exceeds the matching threshold;

[0040] Step 2.3: Determine the location and type of missing components based on the location of components in the skeleton template that are adjacent to matched components but do not have corresponding areas in the point cloud.

[0041] Furthermore, the matching threshold is set to 0.7. This matching threshold is determined by statistically analyzing the similarity distribution of correct and incorrect matches on the labeled dataset, and selecting the similarity value that results in the highest matching accuracy as the matching threshold.

[0042] Furthermore, the extraction of visible components is achieved by region segmentation of the non-rigid instance point cloud. Specifically, the instance point cloud is divided into multiple connected regions based on the spatial continuity and normal vector consistency of the point cloud. Each connected region corresponds to a visible component. The bounding box size of each connected region is calculated as a size parameter. The three-dimensional shape descriptor of the point cloud is extracted as a shape feature. The relative position between the centroids of the connected regions is calculated as the relative positional relationship.

[0043] Furthermore, the shape descriptor is calculated using the point feature histogram method. Specifically, for each point in the part's point cloud, the combined features of the normal vector angle between the point and its neighboring points, the distance between points, and the relative direction are calculated. These combined features are statistically analyzed in the feature space to generate a multidimensional histogram as the shape descriptor. This multidimensional histogram can describe the local geometric shape features of the part.

[0044] Visible component extraction is performed on instance INS047. Based on spatial continuity, the point cloud is divided into 8 connected regions, corresponding to the cyclist's torso, head, left arm, right arm, left leg, right leg, bicycle frame, and bicycle rear wheel. The system extracts the bounding box size and shape descriptor of each component and calculates the relative positional relationships between the components.

[0045] The extracted visible component features are matched with a preset human-vehicle composite skeleton template. The human-vehicle composite skeleton template contains 13 standard component nodes: human torso, head, left upper arm, left forearm, right upper arm, right forearm, left thigh, left lower leg, right thigh, right lower leg, vehicle frame, front wheel, and rear wheel.

[0046] Taking the torso component PART01 as an example, its 125-dimensional shape descriptor vector is denoted as... The shape descriptor subvector of the template torso node is denoted as Geometric similarity is calculated using cosine similarity:

[0047] ;

[0048] Similarly, the geometric similarity between the bicycle rear wheel PART08 and the template rear wheel node is:

[0049] ;

[0050] Both exceeded the matching threshold of 0.7, and a matching relationship was established.

[0051] In the skeleton template, the rear wheel node and the fender node are adjacent through a fixed connection. However, no connected region corresponding to the fender was detected in the current instance point cloud, therefore the fender was identified as a missing component. According to the template definition, the fender should be located above the rear wheel, with a spatial offset of [missing information]. The dimensions of the component are 0.65 meters long, 0.12 meters wide, and 0.08 meters high. The system generates a topology diagram of the incomplete component, which contains 8 visible component nodes and 1 missing component node (mudguard), and records the skeleton connection relationships between the nodes.

[0052] Table 3 shows the component extraction results:

[0053]

[0054] Table 4. Skeleton template matching results:

[0055]

[0056] Step 3: Calculate inter-frame motion vectors and generate a set of component motion parameters;

[0057] For the visible component regions in a non-rigid instance with damage markers, calculate the inter-frame motion vectors of the point clouds between adjacent frames. Based on the direction and magnitude similarity of the inter-frame motion vectors, divide the point cloud within the instance into multiple motion partitions. Semantically associate these motion partitions with the visible component nodes in the damaged component topology graph, calculate the motion parameters of each visible component in the current frame, and generate a set of component motion parameters. ,in This indicates the motion parameters of the first visible component. This indicates the motion parameters of the second visible component. Indicates the first Motion parameters of visible components This indicates the total number of visible components.

[0058] It should be noted that the inter-frame motion vectors mentioned above refer to the position change vectors of the same point in the point cloud between two adjacent frames. The motion partitions mentioned above refer to subsets of point clouds within an instance that have similar motion features. The semantic association mentioned above refers to labeling motion partitions as corresponding component categories based on the correspondence between the spatial location of the motion partitions and the component positions in the skeleton template. The motion parameters mentioned above include the translation vector and rotation parameters of the component in the current frame relative to the previous frame.

[0059] Furthermore, the current frame refers to the frame being processed in the temporal point cloud sequence, the previous frame refers to the frame corresponding to the previous time step of the current frame in the temporal point cloud sequence, and the component motion parameter set records the motion changes of each visible component from the previous frame to the current frame.

[0060] In this embodiment of the application, in order to improve the accuracy of motion partitioning, prior knowledge of the human skeleton is combined when calculating motion partitioning. Specifically, the partitioning is first performed based on motion vector similarity. The motion vector similarity is measured by cosine similarity to ensure the consistency of motion vector direction. Then, the partitioning boundary is adjusted according to the motion coupling constraints of adjacent parts in the skeleton template so that the partitioning boundary is aligned with the position of the skeleton joint.

[0061] Furthermore, the rotation parameters of the component are represented by a rotation matrix. The principal direction of the component in the current frame is obtained by performing principal component analysis on the point cloud within the motion partition. The principal direction of the current frame is compared with the principal direction of the corresponding component in the previous frame, and the rotation matrix between the two is calculated as the rotation parameter. The translation vector is obtained by calculating the displacement of the centroid of the motion partition between adjacent frames.

[0062] Furthermore, the calculation of inter-frame motion vectors adopts the nearest point matching method. Specifically, for each point in the current frame, the nearest point in the point cloud of the previous frame is searched as the corresponding point. The difference between the coordinates of the current frame point and the coordinates of the corresponding point is used as the inter-frame motion vector of that point. To improve the matching accuracy, the search radius is limited when searching for the corresponding point. The search radius is determined based on the measurement accuracy of the lidar and the maximum speed of the target, and is set to 0.5 meters.

[0063] Furthermore, the motion partitioning adopts a graph cut-based clustering method, specifically: the points within the instance are constructed into a graph structure, each point is a graph node, and edges are established between adjacent points. The weight of the edge is calculated based on the similarity of the motion vectors between the two points. The higher the similarity of the motion vectors, the greater the weight. Then, the normalized graph cut algorithm is used to segment the graph structure, so that the sum of edge weights within each subgraph is maximized while the sum of edge weights between subgraphs is minimized. Each subgraph corresponds to a motion partition.

[0064] The system calculates the inter-frame motion vector between frames 126 and 127 for instance INS047. For each point in the current frame, the nearest point within a 0.5-meter radius in the point cloud of the previous frame is searched as the corresponding point, and the position change vector is calculated. The main directions of the inter-frame motion vector of the torso part point cloud are as follows: The amplitude is 0.32 meters per frame. The right leg component, being in a stamping motion, has a motion vector direction of... The length of the torso differs significantly from that of the body.

[0065] Based on the directional similarity of motion vectors between frames, the system uses a graph cut algorithm to divide the instance point cloud into 8 motion partitions. When constructing the graph structure, for two adjacent points within the torso component, their motion vectors are respectively... and The similarity of motion vector directions is calculated using cosine similarity:

[0066] ;

[0067] High similarity indicates that two points belong to the same motion partition, and the edge weight is set to 0.998. For point pairs that cross component boundaries, the motion vector similarity decreases significantly, and the edge weight decreases accordingly, thus achieving accurate partitioning of motion partitions.

[0068] For each motion partition, principal component analysis is used to calculate the principal orientation of the component, and the rotation parameters are obtained by comparing it with the principal orientation of the corresponding component in the previous frame. The rotation angle of the torso component is 3.2 degrees, rotating around the vertical axis; the rotation angle of the right leg component is 15.8 degrees, rotating around the hip joint. The translation vector is obtained by calculating the displacement of the center of mass of each component, generating a set containing the motion parameters of 8 visible components. .

[0069] Table 5. Statistics of inter-frame motion vectors for components:

[0070]

[0071] Table 6: Set of component motion parameters

[0072]

[0073] Step 4: Infer motion assumptions for the missing parts based on kinematic coupling laws;

[0074] Based on the skeleton connection relationship in the topology graph of the missing component, the visible component nodes adjacent to the missing component are obtained. The motion parameters of the adjacent visible components in the current frame are extracted from the component motion parameter set. The motion range and trajectory envelope of the missing component in the current frame are inferred according to the coupling law of human kinematics, and the motion hypothesis of the missing component is generated.

[0075] It should be noted that the aforementioned human kinematic coupling law refers to the constraint relationship between the movements of adjacent human body parts, including joint angle range constraints and coordination constraints between adjacent part movements. The aforementioned motion range refers to the spatial area that the missing part may occupy in the current frame's motion state. The aforementioned trajectory envelope refers to the boundary of the spatial range that the missing part may traverse during its motion from the previous frame to the current frame. The aforementioned motion assumptions of the missing part refer to the estimation results of the missing part's motion state in the current frame, including estimated motion parameters and corresponding confidence levels.

[0076] Furthermore, the temporal dimension of the trajectory envelope is represented by the time interval from the previous frame to the current frame, which is determined by the frame rate of the LiDAR. The trajectory envelope covers all possible motion trajectories of the missing component within this time interval.

[0077] In this embodiment of the application, the inference of the motion assumption of the missing component adopts the constraint propagation method, specifically:

[0078] Step 4.1: Use the motion parameters of adjacent visible components in the current frame as boundary conditions;

[0079] Step 4.2: Based on the joint degree-of-freedom constraints defined by the skeleton connection relationship, calculate the range of motion parameters of the missing part in each degree-of-freedom direction. Specifically, for the joints connected to the missing part, determine the allowable rotational degrees of freedom and rotational angle range according to the joint type. Use the rotational parameters of the adjacent visible parts as a reference. Calculate the feasible region of rotational parameters of the missing part within the joint angle range constraints. Combine the feasible region of rotational parameters with the translation vectors of the adjacent visible parts to obtain the range of motion parameters of the missing part.

[0080] Step 4.3: Use the center value of the range of motion parameters as the estimated motion parameters, and convert the range of motion parameters into confidence level through linear mapping. The smaller the range of motion parameters, the higher the confidence level.

[0081] Furthermore, the joint angle range constraint is preset according to the joint type. For ball joints, the rotation angle range is [value missing] in all directions. For hinge joints, rotation is allowed only in one degree of freedom, with a rotation angle range of [range missing]. Confidence level is determined by the formula The calculation yielded, where Indicates the confidence level. This indicates the range of values ​​for the motion parameters. Indicates the preset maximum value range, when Exceed The confidence level is set to 0.

[0082] Furthermore, the joint type is determined based on the joint definitions in the skeletal template. In the human skeletal template, the shoulder and hip joints are defined as ball joints, and the elbow and knee joints are defined as hinge joints. The range of motion parameter values ​​is also considered. The weighted sum of the Euclidean distances between the volume and translation vector range of the feasible domain of the rotation parameters is obtained by calculating the rotation parameter. The weights are set according to the relative importance of the influence of rotation and translation on the spatial position of the component, with the rotation weight being 0.6 and the translation weight being 0.4.

[0083] Furthermore, the trajectory envelope is calculated by performing a spatial union operation on all possible positions of the missing part within the range of motion parameter values. Specifically, multiple sets of motion parameters are uniformly sampled within the range of motion parameter values. For each set of motion parameters, the spatial position of the missing part in the current frame is calculated. The bounding boxes of all spatial positions are merged to form the trajectory envelope. The sampling density is adaptively adjusted according to the size of the range of motion parameter values; the larger the range of motion parameter values, the denser the sampling.

[0084] Based on the incomplete component topology diagram, the missing fender component and the rear wheel component are adjacent through a fixed connection. The motion parameters of the rear wheel component PART08 are extracted from the component motion parameter set: the translation vector is... Meters, rotation angle of 1.8 degrees, rotation axis direction is .

[0085] Since the mudguard is fixedly connected to the rear wheel, with zero degrees of freedom at the joint, the mudguard's motion parameters should be consistent with those of the rear wheel. However, considering that damage to the mudguard structure might cause misalignment during installation, the system sets the tolerance for mudguard misalignment relative to the rear wheel to be [value missing]. Meters. Based on the standard offset in the skeleton template. Meter and position offset tolerance, calculate the range of values ​​for the mudguard motion parameters: translation vector range is meters, rotation angle range is Spend.

[0086] Size of the range of motion parameters The calculation is as follows:

[0087] ;

[0088] Preset maximum value range The confidence level is set to 0.5, therefore:

[0089] ;

[0090] The center value of the range of motion parameters is used as the estimated motion parameters: the translation vector is... The distance is meters, and the rotation angle is 1.8 degrees. Twenty sets of parameters are uniformly sampled within the range of motion parameters. The spatial position of the mudguard under each set of parameters is calculated, and the results are combined to form a trajectory envelope. The spatial range of the envelope is centered on... Meters, dimensions rice.

[0091] Table 7: Results of inferences based on motion assumptions for the missing components:

[0092]

[0093] Table 8. Missing component trajectory envelope parameters:

[0094]

[0095] Step 5: Perform layered motion compensation and construct a deformation-normalized complete point cloud representation;

[0096] The set of component motion parameters is combined with the motion assumption of the missing component. Based on the motion parameters of the torso component, the visible point cloud of the current frame is subjected to hierarchical motion compensation processing to restore the visible point cloud to the standard pose. For the missing area, a virtual occupant point cloud with consistent motion is generated according to the motion assumption of the missing component. The visible point cloud with the standard pose and the virtual occupant point cloud are merged to construct a deformation-normalized complete point cloud representation.

[0097] It should be noted that the aforementioned layered motion compensation refers to a processing method that performs inverse motion transformation sequentially from the torso to the end components according to the skeletal hierarchy. Motion compensation is first performed on the torso components, and then relative motion compensation is performed sequentially on the components connected to the torso, until all visible components are processed. The aforementioned standard pose refers to a preset reference pose used to eliminate the influence of different motion states on the geometric distribution of the point cloud. The aforementioned virtual occupancy point cloud refers to point cloud data generated in the missing region for placement purposes; the spatial distribution of the virtual occupancy point cloud is determined based on the motion range assumed in the motion hypothesis of the missing component.

[0098] In this embodiment of the application, the virtual occupancy point cloud is generated using a motion constraint sampling method, specifically:

[0099] Step 5.1: Within the motion range defined by the motion assumption of the incomplete part, determine the spatial distribution density of virtual points according to the geometric parameters of the corresponding part in the skeleton template. The spatial distribution density is calculated based on the volume information in the geometric parameters and the preset number of points per unit volume.

[0100] Step 5.2: Generate virtual occupant sites by uniformly sampling within the movement range according to the spatial distribution density.

[0101] Furthermore, the preset number of points per unit volume is set to 1000 points per cubic meter. This number of points per unit volume matches the point cloud density of the target by the lidar at normal detection distance. The motion range is obtained by applying the estimated motion parameters in the motion hypothesis of the incomplete parts to the standard geometry of the corresponding parts in the skeleton template. The standard geometry is represented by a bounding box.

[0102] Furthermore, the specific execution method of layered motion compensation is as follows: First, identify the point cloud region corresponding to the torso component, invert the rotation parameters of the torso component to obtain the inverse rotation matrix, subtract the torso translation vector from the coordinates of each point in the torso point cloud and then multiply by the inverse rotation matrix to complete the motion compensation of the torso component. Then, according to the hierarchical order of the skeleton topology, for each non-torso component, first transform the point cloud of the non-torso component to the coordinate system of its parent component, and then apply the inverse motion parameters of the non-torso component relative to the parent component for compensation. This process is repeated until all visible components are restored to the standard posture.

[0103] Furthermore, the standard posture is defined as the human body standing upright or the cyclist's standard riding posture. In the standard posture, the main direction of the torso is vertically upward, and the joint angle between the limbs and the torso is a preset standard angle. For the human body standing upright, the arms hang naturally and the legs are straight. For the cyclist's standard riding posture, the torso leans forward 30 degrees, the arms extend forward to hold the handlebars, and the legs are kept in the middle position of the pedaling posture.

[0104] The system will collect the motion parameters of the components. Combining the motion assumptions of the incomplete parts, layered motion compensation is performed based on the motion parameters of the torso part PART01. First, motion compensation is performed on the torso part point cloud; the translation vector of the torso is... Meters, rotation angle of 3.2 degrees, rotation axis of Calculate the inverse rotation matrix, subtract the translation vector from the coordinates of the 687 points in the torso point cloud, and then multiply by the inverse rotation matrix on the left to complete the torso motion compensation.

[0105] Following the skeletal hierarchy, relative motion compensation was performed sequentially on the head, left arm, right arm, left leg, and right leg. The rotation angle of the right leg component PART06 was 15.8 degrees, with an additional rotation angle relative to the torso. After transforming the right leg point cloud to the torso coordinate system, a 12.6-degree inverse rotation was applied to complete motion compensation. Motion compensation was then performed on the frame and rear wheel components to restore all visible point clouds to the rider's standard riding posture, in which the torso is leaning forward 30 degrees and the legs are in the middle of the pedaling position.

[0106] A virtual occupancy point cloud is generated for the missing mudguard component. Based on the skeleton template, the standard geometry of the mudguard is a bounding box with a length of 0.65 meters, a width of 0.12 meters, and a height of 0.08 meters, with a volume of... Cubic meters. Based on a unit volume point count of 1000 per cubic meter, the following needs to be generated: Each point is rounded up to obtain 7 virtual occupant points. These 7 virtual occupant points are generated uniformly within the spatial range defined by the trajectory envelope, with their coordinates distributed around the center of the envelope. Nearly 1 meter.

[0107] The visible point cloud (2847 points in total) in standard pose was merged with the virtual occupancy point cloud (7 points) to construct a deformation-normalized complete point cloud representation, totaling 2854 points. The virtual occupancy points were labeled "virtual" for identification in subsequent processing.

[0108] Step 6: Input the point cloud segmentation network and map it to generate the point cloud segmentation result of the damaged non-rigid target;

[0109] The deformation-normalized complete point cloud representation is input into the point cloud segmentation network, and the point-by-point category prediction results are output. The prediction results of the virtual occupancy area are labeled as the corresponding component category and an uncertainty label is added. Based on the component motion parameter set and the motion parameters in the motion assumption of the incomplete component, the point-by-point category prediction results are mapped back to the original motion state according to the deformation parameters, and the damaged non-rigid body target point cloud segmentation results are generated.

[0110] It should be noted that the aforementioned point cloud segmentation network refers to a neural network model used to perform point-by-point semantic segmentation of point clouds. The aforementioned uncertainty label refers to an identifier attached to the prediction results of virtual occupancy regions, used to indicate that the point-by-point category prediction results for these virtual occupancy regions are based on inference rather than actual observation data. The aforementioned mapping back to the original motion state refers to performing a motion transformation on the point-by-point category prediction results under standard pose conditions that is the opposite of hierarchical motion compensation, so that the point-by-point category prediction results correspond to the spatial positions of the original point cloud.

[0111] Furthermore, the point cloud segmentation network adopts a semantic segmentation network based on the PointNet++ architecture. The input of this point cloud segmentation network is the point coordinates and point features in the deformed normalized complete point cloud representation, and the output is the class probability distribution of each input point. The point cloud segmentation network extracts local and global features of the point cloud through multiple ensemble abstraction layers. The last layer of the point cloud segmentation network is a fully connected layer, which maps the extracted features to the probability distribution of each class. The cross-entropy loss function is used to measure the difference between the predicted class and the true class, and the Adam optimization algorithm is used for training.

[0112] Furthermore, when training the point cloud segmentation network, a supervised learning approach is adopted. The training data includes complete point clouds in standard pose and their corresponding point-by-point category labels. For points in the virtual occupancy region, labels are made according to the category of the corresponding component in the skeleton template during training. The loss value of the virtual occupancy region is weighted during training. The weighting weight is determined based on the confidence of the motion hypothesis of the incomplete component. Specifically, the cross-entropy loss value is multiplied by the confidence value to obtain the weighted loss value.

[0113] In this embodiment of the application, in order to improve the reliability of the point-by-point category prediction results, the confidence level of the point-by-point category prediction results of the virtual occupant area is adjusted. Specifically, the category prediction probability of the virtual occupant area is weighted according to the confidence level in the motion hypothesis of the incomplete part. The category prediction probability of the virtual occupant area with a lower confidence level is reduced accordingly. The weighting method is to multiply the category prediction probability by the confidence level.

[0114] In this embodiment, the method further includes a step of completing the fracture primitive geometry. Specifically, after step 1, the geometric attribute continuity of the point cloud regions on both sides of the potential damage boundary is analyzed, the geometric primitive parameters of the point cloud regions on both sides of the potential damage boundary are calculated, it is determined whether the damage fractures on both sides of the potential damage boundary belong to the same geometric primitive, and virtual extension estimation is performed on the regions determined to be damage fractures based on the geometric primitive parameters to generate fracture primitive completion parameters. The aforementioned geometric primitives include basic geometric shapes such as planes, cylindrical surfaces, and spheres. The aforementioned fracture primitive completion parameters are used in step 5 to assist in the generation of virtual occupancy point clouds, so that the virtual occupancy point clouds and the visible point clouds maintain geometric continuity at the fracture boundary.

[0115] Furthermore, the geometric primitive parameters are obtained by fitting using the RANSAC algorithm. For planar primitives, the geometric primitive parameters include the plane normal vector and the distance to the origin. For cylindrical primitives, the geometric primitive parameters include the axial direction, axial position, and radius. The criteria for determining whether the two sides of the potential damage boundary belong to the same geometric primitive for damage fracture are: the geometric primitives fitted on both sides of the potential damage boundary are of the same type and the difference in geometric primitive parameters is less than a preset threshold. The threshold for the difference in plane normal vector is 15 degrees, the threshold for the difference in axial direction of cylindrical surface is 15 degrees, and the threshold for the difference in radius is 10% of the component size.

[0116] Furthermore, the specific method for virtual extension estimation is as follows: when it is determined that the damage fracture on both sides of the potential damage boundary belongs to the same geometric primitive, the geometric primitive parameters on both sides of the potential damage boundary are averaged to obtain unified geometric primitive parameters. Then, based on the unified geometric primitive parameters, a virtual point cloud conforming to the geometric primitive shape is generated in the fracture area. The generation range of the virtual point cloud is determined by the spatial position of the point cloud area on both sides of the potential damage boundary. The density of the virtual point cloud is consistent with the density of the visible point cloud. The generated virtual point cloud is used as part of the fracture primitive completion parameters and is merged with the virtual occupancy point cloud in step 5.

[0117] The system inputs a deformed, normalized complete point cloud representation (2854 points) into a pre-trained PointNet++ point cloud segmentation network. This point cloud segmentation network consists of 4 ensemble abstraction layers and 3 feature propagation layers, outputting 13 categories including: torso, head, upper limbs, lower limbs, frame, wheels, and fenders. The point cloud segmentation network outputs a 13-dimensional category probability vector for each input point.

[0118] For the 2847 points in the visible point cloud region, the point cloud segmentation network outputs normal category prediction results. For example, the average probability of the torso component point cloud being predicted as the "torso" category is 0.92, and the average probability of the right leg component point cloud being predicted as the "lower limb" category is 0.88. For the 7 points in the virtual occupant region, the point cloud segmentation network outputs a prediction probability of 0.76 for the "mudguard" category. After weighted adjustment based on the confidence level of 0.804 for the motion hypothesis of the missing component, the adjusted category prediction probability is... It is also marked with the uncertainty label "virtual inference".

[0119] Based on the set of component motion parameters and the motion parameters in the motion assumptions of the incomplete components, a motion transformation opposite to the hierarchical motion compensation is performed on the point-by-point category prediction results, mapping the prediction results under the standard posture back to the original motion state. First, a positive motion transformation is performed on non-torso components, applying a 12.6-degree positive rotation to the right leg prediction result. Then, the positive motion parameters of the torso are applied to all components, along with the translation vector. The meter and rotation angle of 3.2 degrees were used to make the prediction result correspond to the spatial position of the original point cloud in frame 127.

[0120] The generated point cloud segmentation result for the damaged non-rigid target contains category labels for 2854 points, of which 2847 are predictions from actual observation points and 7 are inferred results from virtual occupant points. After mapping the spatial location of the virtual occupant points back to the original coordinate system, they are located in the area above the rear wheel where the mudguard should be, with coordinates ranging from [missing information]. rice.

[0121] Table 9. Statistics on category predictions output by the point cloud segmentation network:

[0122]

[0123] This implementation addresses the segmentation problem of damaged non-rigid targets in LiDAR point clouds. It establishes a skeleton-like association between missing and visible regions using a topological map of the missing components, ensuring topological connections between components are maintained even when missing components cause topological breaks in the skeleton. By utilizing the coupling principles of human kinematics to infer motion hypotheses of missing components from their motion parameters, it overcomes the difficulty of directly observing the motion state of missing components. Layered motion compensation normalizes the dynamically deformed point cloud to a standard pose, eliminating the influence of motion state on geometric distribution, enabling the point cloud segmentation network to perform segmentation processing in a unified pose space. The generation of virtual occupant point clouds provides occupant information in missing regions, avoiding segmentation boundary errors caused by region gaps. Therefore, this implementation solves the dual dilemma of neither establishing a complete motion model nor performing static geometric completion for dynamically incomplete targets, enabling the segmentation and identification of structurally incomplete non-rigid moving targets in road scenes, reducing target misses and classification errors.

[0124] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for segmenting point clouds using lidar, characterized in that, Includes the following steps: Acquire the temporal point cloud sequence collected by LiDAR, perform non-rigid instance detection on the temporal point cloud sequence, detect the boundary positions caused by geometric discontinuities inside each instance, mark the detected boundary positions as potential damaged boundaries, and generate a set of non-rigid instances with damage markings; Extract the geometric features and spatial positions of visible components from non-rigid body instances with damage markers, match them with a preset skeleton template, identify the position and type of missing components based on the matching results, and generate a topology map of the incomplete components. For the visible component region, calculate the inter-frame motion vector between adjacent frames. Based on the direction similarity and amplitude similarity of the inter-frame motion vector, divide the point cloud within the instance into multiple motion partitions. Semantically associate the motion partitions with the visible component nodes in the incomplete component topology map to generate a set of component motion parameters. Based on the skeleton connection relationship in the topology diagram of the incomplete component, the motion parameters of adjacent visible components are extracted from the set of component motion parameters. The motion range and trajectory envelope of the missing component are inferred according to the coupling law of human kinematics, and the motion hypothesis of the incomplete component is generated. Based on the motion parameters of the torso components, the visible point cloud is subjected to layered motion compensation processing to restore it to the standard posture. For the missing area, a virtual occupant point cloud is generated according to the motion assumption of the missing component. The visible point cloud of the standard posture and the virtual occupant point cloud are merged to construct a deformation-normalized complete point cloud representation. The deformation-normalized complete point cloud representation is input into the point cloud segmentation network, which outputs point-by-point category prediction results. These point-by-point category prediction results are then mapped back to the original motion state to generate point cloud segmentation results for damaged non-rigid targets.

2. The lidar point cloud segmentation method according to claim 1, characterized in that, The geometric discontinuity is determined by the geometric continuity index, which is obtained by calculating the weighted sum of the mean angle between the normal vectors of a local region of the point cloud and the standard deviation of the distance between points. When the geometric continuity index is lower than a preset threshold, the corresponding boundary is marked as a potential damaged boundary. The preset threshold is the quantile of the statistical distribution of the geometric continuity index of the normal target point cloud.

3. The lidar point cloud segmentation method according to claim 1, characterized in that, The non-rigid instance detection method adopts motion clustering. It performs point cloud registration on adjacent frames in the temporal point cloud sequence, calculates the displacement vector of each point between adjacent frames, and clusters the point cloud according to the direction and magnitude of the displacement vector. Points with similar motion characteristics are clustered into one instance. Whether the instance is a non-rigid target is determined based on whether there are multiple sub-regions with different motion directions inside the instance.

4. The lidar point cloud segmentation method according to claim 1, characterized in that, The skeleton template matching adopts a partial matching strategy: visible parts are used as anchor points, and the geometric similarity between the shape descriptor of the visible part and the shape descriptor of the corresponding part in the skeleton template is calculated by cosine similarity; when the geometric similarity exceeds the matching threshold, a matching relationship is established; based on the position of the part in the skeleton template that is adjacent to the matched part but does not have a corresponding region in the point cloud, the position and type of the missing part are determined.

5. The lidar point cloud segmentation method according to claim 4, characterized in that, The extraction of the visible component is achieved by segmenting the non-rigid instance point cloud into regions. Based on the spatial continuity and normal vector consistency of the point cloud, the instance point cloud is divided into multiple connected regions, and each connected region corresponds to a visible component. The shape descriptor is calculated using the point feature histogram method. For each point in the component point cloud, the combined features of the normal vector angle between the point and its neighboring points, the distance between points, and the relative direction are calculated. The combined features are statistically analyzed in the feature space to generate a multi-dimensional histogram.

6. The lidar point cloud segmentation method according to claim 1, characterized in that, The calculation of the inter-frame motion vector adopts the nearest point matching method. For each point in the current frame, the nearest point in the point cloud of the previous frame is searched as the corresponding point, and the difference between the coordinates of the current frame point and the coordinates of the corresponding point is used as the inter-frame motion vector of that point. The division of motion partitions adopts a graph-cut-based clustering method. The points in the instance are constructed into a graph structure, and the weight of the edge is calculated according to the similarity of the motion vectors between two points. The normalized graph cut algorithm is used to segment the graph structure to obtain motion partitions.

7. The lidar point cloud segmentation method according to claim 1, characterized in that, The inference of the motion hypothesis of the missing part adopts the constraint propagation method: the motion parameters of adjacent visible parts are used as boundary conditions; according to the joint degree of freedom constraints defined by the skeleton connection relationship, the range of motion parameter values ​​of the missing part in each degree of freedom direction is calculated; the center value of the range of motion parameter values ​​is used as the estimated motion parameter; the size of the range of motion parameter values ​​is converted into confidence through linear mapping, and the smaller the range of motion parameter values, the higher the confidence.

8. The lidar point cloud segmentation method according to claim 1, characterized in that, The virtual occupant point cloud is generated using a motion constraint sampling method: within the motion range defined by the motion assumption of the incomplete component, the spatial distribution density of virtual points is determined according to the geometric parameters of the corresponding component in the skeleton template. The spatial distribution density is calculated based on the volume information in the geometric parameters and the preset number of points per unit volume. Virtual occupant points are generated by uniformly sampling within the motion range according to the spatial distribution density.

9. The lidar point cloud segmentation method according to claim 1, characterized in that, It also includes the step of completing the geometry of the fracture primitive: after generating a set of non-rigid body instances with damage markers, the geometric property continuity of the point cloud regions on both sides of the potential damage boundary is analyzed, and the geometric primitive parameters of the point cloud regions on both sides of the potential damage boundary are calculated by fitting using the RANSAC algorithm. When the geometric primitives fitted on both sides of the potential damage boundary are of the same type and the difference in geometric primitive parameters is less than a preset threshold, they are determined to be damaged fractures of the same geometric primitive. For the regions determined to be damaged fractures, virtual extension estimation is performed based on the geometric primitive parameters to generate fracture primitive completion parameters. The fracture primitive completion parameters are used to assist in the generation of virtual occupancy point clouds.

10. A lidar point cloud segmentation system, used to execute the lidar point cloud segmentation method according to any one of claims 1 to 9, characterized in that, include: The damaged instance detection module is used to acquire the time-series point cloud sequence collected by the lidar, perform non-rigid instance detection on the time-series point cloud sequence, detect the boundary positions caused by geometric discontinuities inside each instance and mark them as potential damaged boundaries, and generate a set of non-rigid instances with damage markings. The incomplete topology generation module is used to extract the geometric features and spatial positions of visible parts from non-rigid body instances with damage markers, match them with preset skeleton templates, identify the position and type of missing parts based on the matching results, and generate a topology map of incomplete parts. The motion parameter calculation module is used to calculate the inter-frame motion vector between adjacent frames for the visible component region. Based on the inter-frame motion vector, the point cloud within the instance is divided into multiple motion partitions. The motion partitions are semantically associated with the visible component nodes in the incomplete component topology map to generate a set of component motion parameters. The motion hypothesis inference module is used to extract motion parameters of adjacent visible parts from the set of motion parameters of the parts based on the skeleton connection relationship in the topology map of the incomplete parts, infer the motion range and trajectory envelope of the missing parts according to the coupling law of human kinematics, and generate motion hypotheses of the incomplete parts. The deformation normalization module is used to perform layered motion compensation processing on the visible point cloud based on the motion parameters of the torso parts to restore it to the standard posture. For the missing area, a virtual occupant point cloud is generated according to the motion assumption of the missing parts. The visible point cloud in the standard posture is merged with the virtual occupant point cloud to construct a deformation normalized complete point cloud representation. The segmentation result generation module is used to input the deformed normalized complete point cloud representation into the point cloud segmentation network and output the point-by-point category prediction results. The point-by-point category prediction results are then mapped back to the original motion state to generate the point cloud segmentation results of the damaged non-rigid target.

Images