Method and system for correcting manual annotation content of three-dimensional point clouds

Abstract

The present invention provides a method and system for correcting manually annotated content of a three-dimensional point cloud, comprising: step S1: obtaining an original three-dimensional point cloud, manually annotating it, and obtaining manually annotated content; step S2: filtering the original three-dimensional point cloud, removing ground points and retaining non-ground points, and obtaining a filtered three-dimensional point cloud; step S3: determining which points in the filtered three-dimensional point cloud are within a manually annotated box based on the manually annotated content, and obtaining a point cloud within the box; step S4: clustering the point cloud within the box and removing noise points to obtain a clustering result; step S5: correcting the manually annotated content based on the clustering result to obtain corrected annotated content. The present invention corrects the manually annotated content by filtering, clustering, projecting, and solving the point cloud within the manually annotated box, thereby obtaining more accurate annotated content.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional point cloud data annotation, and in particular to a method and system for correcting manually annotated content of a three-dimensional point cloud. Background Art

[0002] Autonomous driving technology is currently developing rapidly. From a technical perspective, the core technologies of autonomous driving lie in: perception, which refers to the system's ability to collect and extract relevant information from the environment, such as detecting pedestrians, vehicles, and road signs; planning, which refers to the decision-making process for an autonomous vehicle from its departure point to its destination, requiring comprehensive consideration of safety, obstacle avoidance, and route optimization; and control, which refers to the ability to accurately complete planned actions. Perception is the foundation and prerequisite for subsequent planning and control. Currently, a large number of deep learning algorithms are being applied to autonomous driving technology, such as SSD, R-CNN, and Cascade RCNN for object detection. To apply deep learning algorithms, the collected point cloud data often needs to be annotated in the early stages to facilitate the training of subsequent perception, planning, and control models.

[0003] When building a 3D point cloud annotation platform from scratch, the typical workflow is to first perform manual annotation. Once a preliminary model is established, semi-automatic annotation is performed. Once the model is fully trained, fully automatic annotation can be performed. However, during the manual annotation phase, due to individual differences and understanding levels, efficiency is low and noise points are easily mixed in, resulting in inaccurate 3D box dimensions and multiple objects appearing within a single box. Directly training the model using data with a high mislabeling rate will affect model accuracy.

[0004] Patent document CN114648677A (application number: CN202210313188.X) discloses a laser point cloud annotation device and method, which includes the following steps: 1) a data acquisition module collects spatiotemporally synchronized laser point cloud data and image data; 2) a data annotation module automatically annotates the image data; 3) the automatically annotated image data is reviewed and any image data that fails the review is corrected; 4) a 3D scene reconstruction module establishes a mapping relationship between the image data and the laser point cloud data; 5) based on the mapping relationship, the laser point cloud data is annotated according to the annotations of the image data. The data annotation module in this method already implements automatic annotation and is not suitable for building an annotation platform from scratch. Summary of the Invention

[0005] In view of the defects in the prior art, the purpose of the present invention is to provide a method and system for correcting manual annotation of three-dimensional point clouds.

[0006] The method for correcting manually annotated content of a three-dimensional point cloud provided by the present invention includes:

[0007] Step S1: Obtain an original three-dimensional point cloud, perform manual annotation on it, and obtain manual annotation content, wherein the manual annotation content includes a manual annotation box and a manual label;

[0008] Step S2: filtering the original three-dimensional point cloud to remove ground points and retain non-ground points, thereby obtaining a filtered three-dimensional point cloud;

[0009] Step S3: for the filtered three-dimensional point cloud, determine which points are within the manually marked box according to the information of the manually marked box, and obtain the point cloud within the box;

[0010] Step S4: clustering the point cloud within the frame and removing noise to obtain a clustering result, wherein the clustering result includes several classes of point clouds and the number of classes N;

[0011] Step S5: According to the number of categories N, the manually annotated content is corrected to obtain corrected annotated content, wherein the corrected annotated content includes a corrected annotation box and a corrected label.

[0012] Preferably, the method for filtering the original three-dimensional point cloud in step S2 includes a cloth filtering method, a slope filtering method and an adaptive irregular network filtering method.

[0013] Preferably, step S3 includes:

[0014] Step S3.1: According to the angle (θ x ,θ y ,θ z ), get the rotation matrix R and its inverse matrix R -1 :

[0015]

[0016]

[0017] Step S3.2: For all points in the filtered three-dimensional point cloud, use the matrix R -1 Perform inverse rotation transformation on them. If the original coordinates of a point in the reference coordinate system are (x a ,y a ,z a ), then according to

[0018]

[0019] The transformed coordinates of the point are calculated as (x′ a ,y′ a ,z′ a ), if x′ a, y′ a and z′ a satisfy:

[0020]

[0021] Then the point is judged to be within the manual annotation box, otherwise it is judged to be outside the manual annotation box; where (x′0, y′0, z′0) is the coordinate of the center point of the manual annotation box after inverse rotation transformation, l x 、l y 、l z is the length of the manual marking frame in the three axes, d x d y d z is the length margin;

[0022] Step S3.3: After traversing all points in the filtered three-dimensional point cloud, all points located within the manually marked box are selected to obtain a point cloud within the box.

[0023] Preferably, in step S4, the DBSCAN algorithm is used to cluster the point cloud within the frame.

[0024] Preferably, step S5 includes:

[0025] If the number of categories N=0, the manual marking box is deleted;

[0026] If the number of categories N=1, the manual annotation box and manual label of this type of point cloud are reviewed. If they meet the standards, they are not corrected. If they do not meet the standards, they are re-annotated to obtain the corrected annotation box and corrected label of this type of point cloud;

[0027] If the number of categories N>1, each category of point cloud is re-labeled to obtain the corrected annotation box and corrected label of each category of point cloud.

[0028] Preferably, the method for obtaining a corrected annotation box of a certain type of point cloud includes:

[0029] Project the point cloud onto the xoy plane of the coordinate system to obtain a plane projection, then find the minimum area circumscribed rectangle of the plane projection, and finally determine the cube corresponding to the minimum area circumscribed rectangle based on the maximum and minimum values of the point cloud on the z-axis, thus obtaining the corrected annotation box.

[0030] In the calculation of the minimum area circumscribed rectangle, the Graham scanning method is first used to calculate the convex hull of the points included in the plane projection, and then the rotating calculus method is used to calculate the minimum area circumscribed rectangle of the convex hull.

[0031] The system for correcting manually annotated content of a three-dimensional point cloud provided by the present invention includes:

[0032] The original point cloud annotation module is used to obtain the original three-dimensional point cloud, manually annotate it, and obtain the manually annotated content, wherein the manually annotated content includes a manually annotated box and a manually labeled content;

[0033] An original point cloud filtering module is used to filter the original three-dimensional point cloud, remove ground points and retain non-ground points, and obtain a filtered three-dimensional point cloud;

[0034] An in-frame point cloud acquisition module is used to determine which points are within the manually marked frame based on the information of the manually marked frame for the filtered three-dimensional point cloud, thereby obtaining an in-frame point cloud;

[0035] The in-frame point cloud clustering module is used to cluster the in-frame point cloud and remove noise to obtain a clustering result, wherein the clustering result includes several classes of point clouds and the number of classes N;

[0036] The annotation content correction module is used to correct the manually annotated content according to the number of categories N to obtain corrected annotation content, wherein the corrected annotation content includes a corrected annotation box and a corrected label.

[0037] Preferably, the method for filtering the original three-dimensional point cloud by the original point cloud filtering module includes a cloth filtering method, a slope filtering method and an adaptive irregular network filtering method.

[0038] Preferably, the method for the in-frame point cloud acquisition module to acquire the in-frame point cloud includes:

[0039] First, according to the angle (θ x ,θ y ,θ z ), get the rotation matrix R and its inverse matrix R -1 :

[0040]

[0041]

[0042] Then, for all points in the filtered three-dimensional point cloud, the matrix R -1 Perform inverse rotation transformation on them. If the original coordinates of a point in the reference coordinate system are (x a ,y a ,z a ), then according to

[0043]

[0044] The transformed coordinates of the point are calculated as (x′ a ,y′ a ,z′a ), if x′ a , y′ a and z′ a satisfy:

[0045]

[0046] Then the point is judged to be within the manual annotation box, otherwise it is judged to be outside the manual annotation box; where (x′0, y′0, z′0) is the coordinate of the center point of the manual annotation box after inverse rotation transformation, l x 、l y 、l z is the length of the manual marking frame in the three axes, d x d y d z is the length margin;

[0047] Finally, after traversing all points in the filtered three-dimensional point cloud, all points located within the manually marked box are selected to obtain the point cloud within the box.

[0048] Preferably, the in-frame point cloud clustering module uses a DBSCAN algorithm to cluster the in-frame point cloud.

[0049] Preferably, the method for the annotation content correction module to obtain the corrected annotation content includes:

[0050] If the number of categories N=0, the manual marking box is deleted;

[0051] If the number of categories N=1, the manual annotation box and manual label of this type of point cloud are reviewed. If they meet the standards, they are not corrected. If they do not meet the standards, they are re-annotated to obtain the corrected annotation box and corrected label of this type of point cloud;

[0052] If the number of categories N>1, each category of point cloud is re-labeled to obtain the corrected annotation box and corrected label of each category of point cloud.

[0053] Preferably, the method for obtaining a corrected annotation box of a certain type of point cloud includes:

[0054] Project the point cloud onto the xoy plane of the coordinate system to obtain a plane projection, then find the minimum area circumscribed rectangle of the plane projection, and finally determine the cube corresponding to the minimum area circumscribed rectangle based on the maximum and minimum values of the point cloud on the z-axis, thus obtaining the corrected annotation box.

[0055] In the calculation of the minimum area circumscribed rectangle, the Graham scanning method is first used to calculate the convex hull of the points included in the plane projection, and then the rotating calculus method is used to calculate the minimum area circumscribed rectangle of the convex hull.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] 1. This invention considers that the ground point cloud may affect the clustering effect under the density-based clustering method. It uses filtering to remove the ground point cloud and only considers the non-ground point cloud, thereby reducing the interference of the clustering algorithm;

[0058] 2. The present invention corrects the manually annotated content by filtering, clustering, projecting and solving the point cloud within the manually annotated box, thereby obtaining more accurate annotation content;

[0059] 3. In the method and system provided by the present invention, when the number of clusters of point clouds in a manually labeled box is greater than 1, it is considered that there are multiple objects in the labeled box. The content of each category is manually re-labeled, and the labeled box of each category is corrected to make the final labeling result more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0061] Figure 1 This is a flow chart of the method for correcting manual annotation content of a three-dimensional point cloud according to the present invention. DETAILED DESCRIPTION

[0062] The present invention will be described in detail below with reference to specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several changes and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0063] Example:

[0064] like Figure 1 The method for correcting manually annotated content of a three-dimensional point cloud provided by the present invention comprises:

[0065] Step S1: Obtain an original three-dimensional point cloud, perform manual annotation on it, and obtain manual annotation content, wherein the manual annotation content includes a manual annotation box and a manual label;

[0066] Step S2: filtering the original three-dimensional point cloud to remove ground points and retain non-ground points, thereby obtaining a filtered three-dimensional point cloud;

[0067] Step S3: for the filtered three-dimensional point cloud, determine which points are within the manually marked box according to the information of the manually marked box, and obtain the point cloud within the box;

[0068] Step S4: clustering the point cloud within the frame and removing noise to obtain a clustering result, wherein the clustering result includes several classes of point clouds and the number of classes N;

[0069] Step S5: According to the number of categories N, the manually annotated content is corrected to obtain corrected annotated content, wherein the corrected annotated content includes a corrected annotation box and a corrected label.

[0070] The methods for filtering the original three-dimensional point cloud in step S2 include cloth filtering method, slope filtering method and adaptive irregular network filtering method.

[0071] In one embodiment, a cloth filter is used to remove ground points, specifically:

[0072] 1. Flip the terrain:

[0073] 2. Initialize the cloth grid and set the initial vertical position to the highest point;

[0074] 3. Find the nearest point cloud data for each cloth mesh particle, and the height value of the point cloud data is recorded as IHV;

[0075] 4. Calculate the displacement of each particle under the influence of gravity. If the height of the particle is less than IHV, it is set as immovable.

[0076] 5. Calculate the displacement of particles under interaction forces;

[0077] 6. Repeat 4 to 5. Stop when the height difference of a single iteration is less than the threshold or reaches the maximum number of iterations.

[0078] 7. Compare the distance between the point cloud data and the adjacent cloth particles. If the distance is less than a given threshold, it is recorded as a ground point, otherwise it is an object point.

[0079] When executing the above filtering algorithm, since there are no ground truth labels, the maximum values of the x, y, and z axes of the point cloud are used, combined with the image observation of the separation of pedestrians and vehicles on the road, as the parameter adjustment principle, and finally the three-dimensional point cloud is obtained after removing the ground point cloud.

[0080] The step S3 comprises:

[0081] Step S3.1: According to the angle (θ x ,θ y ,θ z ), get the rotation matrix R and its inverse matrix R -1 :

[0082]

[0083]

[0084] Step S3.2: For all points in the filtered three-dimensional point cloud, use the matrix R -1 Perform inverse rotation transformation on them. If the original coordinates of a point in the reference coordinate system are (x a ,y a ,z a ), then according to

[0085]

[0086] The transformed coordinates of the point are calculated as (x′ a ,y′ a ,z′ a ), if x′ a , y′ a and z′ a satisfy:

[0087]

[0088] Then the point is judged to be within the manual annotation box, otherwise it is judged to be outside the manual annotation box; where (x′0, y′0, z′0) is the coordinate of the center point of the manual annotation box after inverse rotation transformation, l x 、l y 、l z is the length of the manual marking frame in the three axes, d x d y d z is the length margin;

[0089] Step S3.3: After traversing all points in the filtered three-dimensional point cloud, all points located within the manually marked box are selected to obtain a point cloud within the box.

[0090] In step S4, the DBSCAN algorithm is used to cluster the point cloud within the frame.

[0091] In this embodiment, DBSCAN is a density-based clustering method that can find all dense areas of sample points and treat the set of the largest densely connected points derived by density as a cluster. The specific implementation steps are as follows:

[0092] Assume that the sample set D=(p1,p2,…,p m ), ε is used to describe a neighborhood distance, and minpt represents the threshold of the number of samples within a certain distance ε.

[0093] 1. Select an unlabeled point p from the dataset i , get the neighboring points in the ε neighborhood of the point, if the number of neighboring points is greater than minpt, record p iFor a core point, add it to the core point queue and create a cluster queue, otherwise re-execute the process;

[0094] 2. Traverse the core point queue: the points in the ε neighborhood of each core point are added to the corresponding cluster and marked, and each point in the neighborhood is searched again for points in the ε neighborhood. If the point is greater than minpt, the current point is considered to be a core point and added to the core point queue;

[0095] 3. After the traversal is completed, if there are still unmarked points, continue to execute 1, otherwise end.

[0096] Use the grid search method to traverse the values of clustering parameters (ε, minpt) and select the group with the largest silhouette coefficient. The specific steps are:

[0097] 1. Assume that k clusters are formed under a certain set of parameters: calculate the point p i The average distance a to other points in the same cluster i , the smaller the distance, the more it should be clustered into the cluster. i Point p i The intra-cluster dissimilarity.

[0098] 2. Calculate p i To another cluster C j The average distance b of all points ij , called point p i with C j The dissimilarity of . Define b i =min{b i1 ,b i2 ,…,b ik} is p i The inter-cluster dissimilarity, b i The larger the value, the more likely the point p is. i The less it belongs to other clusters.

[0099] 3. Click p i Silhouette coefficient So s i ∈[-1,1], the larger the value, the more reasonable the clustering; otherwise, it means that p i It should be classified into other clusters. The silhouette coefficient of all points Select (ε,minpt) that maximizes s.

[0100] After clustering the point cloud within the annotation box and removing noise, we can obtain the point set corresponding to each category, which is the clustering result and the number of categories N.

[0101] The step S5 comprises:

[0102] If the number of categories N=0, the manual marking box is deleted;

[0103] If the number of categories N=1, the manual annotation box and manual label of this type of point cloud are reviewed. If they meet the standards, they are not corrected. If they do not meet the standards, they are re-annotated to obtain the corrected annotation box and corrected label of this type of point cloud;

[0104] If the number of categories N>1, each category of point cloud is re-labeled to obtain the corrected annotation box and corrected label of each category of point cloud.

[0105] Methods for obtaining the corrected annotation box of a certain type of point cloud include:

[0106] Project the point cloud onto the xoy plane of the coordinate system to obtain a plane projection, then find the minimum area circumscribed rectangle of the plane projection, and finally determine the cube corresponding to the minimum area circumscribed rectangle based on the maximum and minimum values of the point cloud on the z-axis, thus obtaining the corrected annotation box.

[0107] In the calculation of the minimum area circumscribed rectangle, the Graham scanning method is first used to calculate the convex hull of the points included in the plane projection, and then the rotating calculus method is used to calculate the minimum area circumscribed rectangle of the convex hull.

[0108] More specifically, the method for calculating the convex hull of a two-dimensional plane point set Q using the Graham scanning method is:

[0109] 1. Let p0 be the point with the smallest y coordinate in Q (if there is more than one point with the smallest y value, p0 is the point with the smallest x value);

[0110] 2. Other points (p1, p2, ..., p n ) are sorted from small to large with p0 as the pole (if the polar angles are the same, sort them from small to large with p0 as the pole 02 The distance is sorted from small to large);

[0111] 3. Push p0, p1, and p2 into stack S;

[0112] 4. Traverse p in counterclockwise order i ∈(p3,p4,…,p n ), assuming that S=(p0,…,p j-1 ,p j ), j <i;

[0113] If the current point p i does not constitute the current convex hull vertex, that is Delete p from S i Otherwise, p i Join to S;

[0114] At this time, the convex hull S corresponding to the two-dimensional point set is obtained.

[0115] Then, the rotating calcaneal method is used to calculate the minimum area circumscribed rectangle that constitutes the convex hull. Consider a convex polygon M with four endpoints p xmin 、p xmax 、p ymin 、p ymax They are the points corresponding to the minimum value of the x-axis, the maximum value of the x-axis, the minimum value of the y-axis, and the maximum value of the y-axis. Algorithm flow:

[0116] 1. Two caliper sets are determined by the tangent lines formed by the above four endpoints;

[0117] 2. When a side coincides with a tangent, the current minimum area is recorded as the area of the rectangle formed by the four lines; otherwise, the minimum area is recorded as positive infinity;

[0118] 3. Rotate clockwise until one of the lines coincides with an edge of the polygon;

[0119] 4. Calculate the area of the new rectangle. If it is less than the minimum area, update the minimum area to the current rectangle area.

[0120] 5. Repeat steps 3-4 until the rotation angle is greater than 90 degrees;

[0121] 6. Output the minimum area.

[0122] For each class obtained by point cloud clustering, find its minimum area circumscribed rectangle in the xoy plane, and then determine the eight point coordinates of the circumscribed box corresponding to the class according to the maximum and minimum values of the z-axis, which is the corrected annotation box.

[0123] The system for correcting manually annotated content of a three-dimensional point cloud provided by the present invention includes:

[0124] The original point cloud annotation module is used to obtain the original three-dimensional point cloud, manually annotate it, and obtain the manually annotated content, wherein the manually annotated content includes a manually annotated box and a manually labeled content;

[0125] An original point cloud filtering module is used to filter the original three-dimensional point cloud, remove ground points and retain non-ground points, and obtain a filtered three-dimensional point cloud;

[0126] An in-frame point cloud acquisition module is used to determine which points are within the manually marked frame based on the information of the manually marked frame for the filtered three-dimensional point cloud, thereby obtaining an in-frame point cloud;

[0127] The in-frame point cloud clustering module is used to cluster the in-frame point cloud and remove noise to obtain a clustering result, wherein the clustering result includes several classes of point clouds and the number of classes N;

[0128] The annotation content correction module is used to correct the manually annotated content according to the number of categories N to obtain corrected annotation content, wherein the corrected annotation content includes a corrected annotation box and a corrected label.

[0129] The method for filtering the original three-dimensional point cloud by the original point cloud filtering module includes a cloth filtering method, a slope filtering method and an adaptive irregular network filtering method.

[0130] The method for obtaining the point cloud within the frame by the in-frame point cloud obtaining module includes:

[0131] First, according to the angle (θ x ,θ y ,θ z ), get the rotation matrix R and its inverse matrix R -1 :

[0132]

[0133]

[0134] Then, for all points in the filtered three-dimensional point cloud, the matrix R -1 Perform inverse rotation transformation on them. If the original coordinates of a point in the reference coordinate system are (x a ,y a ,z a ), then according to

[0135]

[0136] The transformed coordinates of the point are calculated as (x′ a ,y′ a ,z′ a ), if x′ a , y′ a and z′ a satisfy:

[0137]

[0138] Then the point is judged to be within the manual annotation box, otherwise it is judged to be outside the manual annotation box; where (x′0, y′0, z′0) is the coordinate of the center point of the manual annotation box after inverse rotation transformation, l x 、l y 、l z is the length of the manual marking frame in the three axes, d x d y d z is the length margin;

[0139] Finally, after traversing all points in the filtered three-dimensional point cloud, all points located within the manually marked box constitute an in-box point cloud.

[0140] The in-frame point cloud clustering module uses the DBSCAN algorithm to cluster the in-frame point cloud.

[0141] The method for the annotation content correction module to obtain the corrected annotation content includes:

[0142] If the number of categories N=0, the manual marking box is deleted;

[0143] If the number of categories N=1, the manual annotation box and manual label of this type of point cloud are reviewed. If they meet the standards, they are not corrected. If they do not meet the standards, they are re-annotated to obtain the corrected annotation box and corrected label of this type of point cloud;

[0144] If the number of categories N>1, each category of point cloud is re-labeled to obtain the corrected annotation box and corrected label of each category of point cloud.

[0145] Methods for obtaining the corrected annotation box of a certain type of point cloud include:

[0146] Project the point cloud onto the xoy plane of the coordinate system to obtain a plane projection, then find the minimum area circumscribed rectangle of the plane projection, and finally determine the cube corresponding to the minimum area circumscribed rectangle based on the maximum and minimum values of the point cloud on the z-axis, thus obtaining the corrected annotation box.

[0147] In the calculation of the minimum area circumscribed rectangle, the Graham scanning method is first used to calculate the convex hull of the points included in the plane projection, and then the rotating calculus method is used to calculate the minimum area circumscribed rectangle of the convex hull.

[0148] The method used in the present invention to remove ground points is the CSF filtering method. Other solutions for separating ground points can also be used to achieve the ultimate goal, such as: a slope filtering algorithm developed by comparing the slopes of radar points and their neighboring points; an adaptive triangular irregular network filtering algorithm established; a morphological method for filtering by selecting an appropriate window size;

[0149] The present invention adopts DBSCAN clustering. In addition to this, a series of improved algorithms based on it can also be adopted based on specific data, such as: considering the problem of outliers when processing high-dimensional density-inhomogeneous data, the derived DDBSCAN algorithm; improvements based on the optimization of the algorithm's time complexity, etc.

[0150] The present invention seeks the minimum area circumscribed rectangle of a point cloud on the xoy plane, and can also seek the minimum perimeter circumscribed rectangle.

[0151] After performing CSF filtering on the 3D point cloud data to remove ground points, the present invention selects points within the manually labeled box and clusters them. After calculating the minimum area circumscribed rectangle of each cluster projected onto the xoy plane, the circumscribed cube of the corresponding rectangle is calculated based on the maximum value of the z-axis, thereby achieving the effect of correcting the manually labeled box.

[0152] Those skilled in the art will appreciate that, in addition to implementing the system, device, and various modules provided by the present invention in purely computer-readable program code, it is entirely possible to implement the same program in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, embedded microcontrollers, and the like by logically programming the method steps. Therefore, the system, device, and various modules provided by the present invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; the modules for implementing various functions can also be considered both software programs for implementing the method and structures within the hardware component.

[0153] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. The embodiments of this application and the features in the embodiments may be combined with each other in any manner unless there is a conflict.

Claims

1. A method for correcting manually annotated content of a three-dimensional point cloud, characterized in that: include: Step S1: Obtain an original three-dimensional point cloud, perform manual annotation on it, and obtain manual annotation content, wherein the manual annotation content includes a manual annotation box and a manual label; Step S2: filtering the original three-dimensional point cloud to remove ground points and retain non-ground points, thereby obtaining a filtered three-dimensional point cloud; Step S3: for the filtered three-dimensional point cloud, determine which points are within the manually marked box according to the information of the manually marked box, and obtain the point cloud within the box; Step S4: clustering the point cloud data in each frame one by one according to the manually marked frame and removing noise points to obtain a clustering result, which includes several classes of point clouds and the number of classes N; Step S5: Correcting the manually annotated content according to the number of categories N to obtain corrected annotated content, wherein the corrected annotated content includes a corrected annotation frame and a corrected label; The step S5 comprises: If the number of categories N=0, the manual marking box is deleted; If the number of categories N=1, the manual annotation box and manual label of this type of point cloud are reviewed. If they meet the standards, they are not corrected. If they do not meet the standards, they are re-annotated to obtain the corrected annotation box and corrected label of this type of point cloud; If the number of categories N>1, each category of point cloud is re-labeled to obtain the corrected annotation box and corrected label of each category of point cloud; Methods for obtaining the corrected annotation box of a certain type of point cloud include: Project the point cloud onto the xoy plane of the coordinate system to obtain a plane projection, then find the minimum area circumscribed rectangle of the plane projection, and finally determine the cube corresponding to the minimum area circumscribed rectangle based on the maximum and minimum values of the point cloud on the z-axis, thus obtaining the corrected annotation box. In the calculation of the minimum area circumscribed rectangle, the Graham scanning method is first used to calculate the convex hull of the points included in the plane projection, and then the rotating calculus method is used to calculate the minimum area circumscribed rectangle of the convex hull.

2. The method for correcting manual annotation content of a three-dimensional point cloud according to claim 1, characterized in that: The methods for filtering the original three-dimensional point cloud in step S2 include cloth filtering method, slope filtering method and adaptive irregular network filtering method.

3. The method for correcting manual annotation content of a three-dimensional point cloud according to claim 1, characterized in that: The step S3 comprises: Step S3.1: According to the angle (θ x ,θ y ,θ z ), get the rotation matrix R and its inverse matrix R -1 : Step S3.2: For all points in the filtered three-dimensional point cloud, use the matrix R -1 Perform inverse rotation transformation on them. If the original coordinates of a point in the reference coordinate system are (x a ,y a ,z a ), then according to The transformed coordinates of the point are calculated as (x ′ a ,y a ′ ,z ′ a ), if x ′ a 、y a ′ and z ′ a satisfy: Then the point is judged to be within the manual marking frame, otherwise it is judged to be outside the manual marking frame; where (x ′ 0,y ′ 0,z ′ 0) is the coordinate of the center point of the manual annotation frame after inverse rotation transformation, l x 、l y 、l z is the length of the manual marking frame in the three axes, d x d y d z is the length margin; Step S3.3: After traversing all points in the filtered three-dimensional point cloud, all points located within the manually marked box are selected to obtain a point cloud within the box.

4. The method for correcting manual annotation content of a three-dimensional point cloud according to claim 1, characterized in that: In step S4, the DBSCAN algorithm is used to cluster the point cloud within the frame.

5. A system for correcting manually annotated content of a 3D point cloud, characterized in that: include: The original point cloud annotation module is used to obtain the original three-dimensional point cloud, manually annotate it, and obtain the manually annotated content, wherein the manually annotated content includes a manually annotated box and a manually labeled content; An original point cloud filtering module is used to filter the original three-dimensional point cloud, remove ground points and retain non-ground points, and obtain a filtered three-dimensional point cloud; An in-frame point cloud acquisition module is used to determine which points are within the manually marked frame based on the information of the manually marked frame for the filtered three-dimensional point cloud, thereby obtaining an in-frame point cloud; The point cloud clustering module in the frame is used to cluster the point cloud data in each frame one by one according to the manually marked frame and remove noise points to obtain a clustering result. The clustering result includes several classes of point clouds and the number of classes N; An annotation content correction module, configured to correct the manually annotated content according to the number of categories N to obtain corrected annotation content, wherein the corrected annotation content includes a corrected annotation frame and a corrected label; The method for the annotation content correction module to obtain the corrected annotation content includes: If the number of categories N=0, the manual marking box is deleted; If the number of categories N=1, the manual annotation box and manual label of this type of point cloud are reviewed. If they meet the standards, they are not corrected. If they do not meet the standards, they are re-annotated to obtain the corrected annotation box and corrected label of this type of point cloud; If the number of categories N>1, each category of point cloud is re-labeled to obtain the corrected annotation box and corrected label of each category of point cloud; Methods for obtaining the corrected annotation box of a certain type of point cloud include: Project the point cloud onto the xoy plane of the coordinate system to obtain a plane projection, then find the minimum area circumscribed rectangle of the plane projection, and finally determine the cube corresponding to the minimum area circumscribed rectangle based on the maximum and minimum values of the point cloud on the z-axis, thus obtaining the corrected annotation box. In the calculation of the minimum area circumscribed rectangle, the Graham scanning method is first used to calculate the convex hull of the points included in the plane projection, and then the rotating calculus method is used to calculate the minimum area circumscribed rectangle of the convex hull.

6. The system for correcting manual annotation content of a 3D point cloud according to claim 5, characterized in that: The method for filtering the original three-dimensional point cloud by the original point cloud filtering module includes a cloth filtering method, a slope filtering method and an adaptive irregular network filtering method.

7. The system for correcting manual annotation content of a 3D point cloud according to claim 5, characterized in that: The method for obtaining the point cloud within the frame by the in-frame point cloud obtaining module includes: First, according to the angle (θ x ,θ y ,θ z ), get the rotation matrix R and its inverse matrix R -1 : Then, for all points in the filtered three-dimensional point cloud, the matrix R -1 Perform inverse rotation transformation on them. If the original coordinates of a point in the reference coordinate system are (x a ,y a ,z a ), then according to The transformed coordinates of the point are calculated as (x ′ a ,y a ′ ,z ′ a ), if x ′ a 、y a ′ and z ′ a satisfy: Then the point is judged to be within the manual marking frame, otherwise it is judged to be outside the manual marking frame; where (x ′ 0,y ′ 0,z ′ 0) is the coordinate of the center point of the manual annotation frame after inverse rotation transformation, l x 、l y 、l z is the length of the manual marking frame in the three axes, d x d y d z is the length margin; Finally, after traversing all points in the filtered three-dimensional point cloud, all points located within the manually marked box are selected to obtain the point cloud within the box.

8. The system for correcting manual annotation content of a 3D point cloud according to claim 5, characterized in that: The in-frame point cloud clustering module uses the DBSCAN algorithm to cluster the in-frame point cloud.

Images