阅读说明：本技术 一种运动物体点云模型配准方法 (Moving object point cloud model registration method ) 是由 邱鹏 李赛红 于 2021-07-27 设计创作，主要内容包括：本发明涉及机器视觉技术领域,一种运动物体点云模型配准方法,其特征在于：包括如下步骤,输入同一尺寸物体两次运动过程中扫描重建所得点云模型；分别提取同一尺寸物体两个模型中鲜明、且相同的若干局部点云,基于提取的相同局部点云,对两个物体模型做初次配准；基于配准结果调整物体模型得到初次匹配结果；分别提取两个物体模型主体的相同部分,基于提取物体的相同部分做最终配准。该方法适用于点云密度比较小的较为稀疏的物体点云模型,适用于两个存在较大差异部分的物体点云模型,同时可提高匹配速度。(The invention relates to the technical field of machine vision, in particular to a moving object point cloud model registration method, which is characterized by comprising the following steps: inputting a point cloud model obtained by scanning and reconstructing in the process of two movements of an object with the same size; respectively extracting a plurality of distinct and same local point clouds in two models of objects with the same size, and performing primary registration on the two object models based on the extracted same local point clouds; adjusting the object model based on the registration result to obtain a primary matching result; and respectively extracting the same parts of the two object model main bodies, and performing final registration based on the extracted same parts of the objects. The method is suitable for a relatively sparse object point cloud model with relatively small point cloud density, is suitable for two object point cloud models with relatively large difference, and can improve the matching speed.)

1. A method for registering a point cloud model of a moving object is characterized by comprising the following steps: comprises the following steps of (a) carrying out,

step 1, inputting a point cloud model obtained by scanning and reconstructing in the process of two movements of an object with the same size;

step 2, respectively extracting a plurality of distinct and same local point clouds in two models of the object with the same size, and performing primary registration on the two object models based on the extracted same local point clouds;

step 3, adjusting the object model based on the registration result to obtain a primary matching result;

and 4, respectively extracting the same parts of the two object model bodies, and performing final registration based on the extracted same parts of the objects.

2. The method of registering a point cloud model of a moving object according to claim 1, characterized in that: in step 1, the input point cloud model registration object is a large-size sparse point cloud model with a certain difference.

3. The method of registering a point cloud model of a moving object according to claim 2, characterized in that: in step 2, matching of the input point cloud model, taking the significant same parts of the two object models as a first matching input reference, the parts can be extracted by methods such as the proportion of the whole structure, and the like, taking the first matching result as the reference, taking the main body part with small object feature difference as the input reference of the second matching, the main body part can be obtained based on the voxel difference size after octree segmentation, and combining with the PCL conditional filtering, the same part structure required by the second registration can be accurately obtained.

4. The method of registering a point cloud model of a moving object according to claim 3, characterized in that: in step 2, calculating the coordinates of the positions of the centroids of the structures of the parts with the same significance, adjusting the point cloud by taking the centroids as the origin of the coordinate system, coinciding the coordinates of the positions of the centroids of the two structures with the same significance, and performing ICP registration to realize the primary registration of the two object models.

5. The method of registering a point cloud model of a moving object according to claim 4, characterized in that: in step 2, after the registration of the significantly identical partial structures is completed, a rigid transformation matrix of the two point clouds is obtained, the rigid transformation matrix describes parameters of the other significantly identical partial structure position obtained by performing rotational translation on one of the significantly identical partial structures in the three-dimensional space, and based on the rigid transformation matrix, the point cloud position of one object model is subjected to rigid transformation to the point cloud position of the other object model.

6. The method of registering a point cloud model of a moving object according to claim 5, characterized in that: in step 4, extracting other local same part point clouds except for the obviously same part structure, calculating and solving a point cloud normal vector, judging and removing part of the point clouds of which the normal vector is vertical to a specific plane according to the normal vector, filtering the processed part of the point clouds, performing voxel segmentation by adopting an octree, comparing each voxel of the two point clouds, reserving the same part of the point clouds, obtaining two similar point clouds subjected to primary coarse registration, wherein the point clouds have smaller difference and roughly overlapped spatial positions, and dynamically adjusting the size of an object point cloud model and the density of the sparse point clouds according to the point cloud density under the ICP convergence condition based on an ICP algorithm framework arranged in a PCL library.

7. The method of registering a point cloud model of a moving object according to claim 6, characterized in that: and optimizing the resolving efficiency of the objective function in the ICP algorithm flow based on a Google open-source Ceres convex optimization algorithm library.

Technical Field

The invention relates to the technical field of machine vision, in particular to a moving object point cloud model registration method.

Background

At present, a mainstream method for Point cloud model registration mainly includes a registration method based on feature points and an ICP (Iterative Closest Point algorithm) based method, and global overall registration is performed on a Point cloud model. At present, the mainstream method is generally applied to scenes with dense point clouds, small size and small difference between registered point clouds.

The application scene is the sparse point cloud registration of moving objects with large difference and large size. The point cloud registration refers to solving the parameter of coincidence between two similar point clouds after translation and rotation. In the field of 3D computer vision, point cloud models acquired and processed by different sensors are fused to perform three-dimensional reconstruction.

The prior art model registration method has the following disadvantages:

1. the method has low applicability to the point cloud model of the object with larger size and the precision is difficult to meet. Taking a large truck as an example, the size range of the truck point cloud can reach 2.5m x 3m x 20m, and a general point cloud registration algorithm is difficult to apply.

2. The method has low applicability to object point cloud models with low density and sparse point clouds, and the precision is difficult to meet. Some objects often have sparse point cloud density of the objects obtained by scanning and reconstructing the sensor due to large size, and the resolution of a scanning common large truck reaches a coordinate point interval of 5 cm.

3. The method has low applicability to two object point cloud models with large difference parts, and the precision is difficult to meet. The object may carry other redundant objects during the secondary scanning, so that the two scanning results have a large difference, and the general point cloud registration algorithm is difficult to apply.

4. The efficiency is low and the registration takes a long time. Based on a related point cloud registration algorithm built in a PCL (personal computer) library, the time consumption for carrying out ICP (inductively coupled plasma) registration under strict one-to-one condition is up to two minutes when the object point cloud with the scale of 6 ten thousand points is subjected to the ICP registration.

Disclosure of Invention

In order to solve the problems, the invention provides a moving object point cloud model registration method which is suitable for a sparse object point cloud model with small point cloud density and two object point cloud models with large difference parts and can improve the matching speed.

In order to solve the problems, the invention adopts the technical scheme that:

a method for registering a point cloud model of a moving object comprises the following steps,

step 1, inputting a point cloud model obtained by scanning and reconstructing in the process of two movements of an object with the same size;

step 2, respectively extracting a plurality of distinct and same local point clouds in two models of the object with the same size, and performing primary registration on the two object models based on the extracted same local point clouds;

step 3, adjusting the object model based on the registration result to obtain a primary matching result;

and 4, respectively extracting the same parts of the two object model bodies, and performing final registration based on the extracted same parts of the objects.

Preferably, in step 1, the input point cloud model registration object is a large-size sparse point cloud model with a certain degree of difference.

Preferably, in step 2, the matching of the input point cloud model takes the significantly same parts of the two object models as the first matching input reference, the parts can be extracted by methods such as the proportion of the whole structure, and the like, and then takes the first matching result as the reference, takes the main part with small object feature difference as the input reference of the second matching, the main part can be obtained based on the voxel difference size after octree segmentation, and the same part structure required by the second registration can be accurately obtained by combining with the conditional filtering of PCL.

Preferably, in step 2, the coordinates of the positions of the centroids of the structures of the significant same parts are calculated, the point cloud is adjusted by using the centroid as the origin of the coordinate system, and after the coordinates of the positions of the centroids of the two structures of the significant same parts are overlapped, the ICP registration is performed, so that the primary registration of the two object models is realized.

Preferably, in step 2, after the registration of the significantly identical partial structures is completed, a rigid transformation matrix of the two point clouds is obtained, the rigid transformation matrix describes parameters of one of the significantly identical partial structures, which is subjected to rotational translation in a three-dimensional space to obtain the position of the other significantly identical partial structure, and based on the rigid transformation matrix, the position of one of the object model point clouds is subjected to rigid transformation to the position of the other one of the object model point clouds.

Preferably, in step 4, extracting other local same part point clouds except for obviously same part structures, calculating and solving a point cloud normal vector, judging and removing part of the point clouds of which the normal vector is vertical to a specific plane according to the normal vector, filtering the processed part of the point clouds, performing voxel segmentation by using an octree, comparing each voxel of the two point clouds, retaining the same part of the point clouds, obtaining two similar point clouds subjected to primary coarse registration, wherein the point clouds have smaller difference and roughly overlapped spatial positions, and dynamically adjusting the size of an object point cloud model and the density of the sparse point clouds and the ICP convergence condition according to the point cloud density based on an ICP algorithm framework built in a PCL library.

Preferably, the efficiency of the solution of the objective function in the ICP algorithm flow is optimized based on a Google open-source Ceres convex optimization algorithm library.

The invention has the beneficial effects that:

1. the method supports and is suitable for a sparse object point cloud model with small point cloud density. Referring to truck model registration, the method can be applied to point cloud registration with sparsity degree of which the distance between points in the point cloud is up to 5 cm.

2. The method supports and is suitable for two object point cloud models with large difference parts. With reference to truck model registration, the present invention can accommodate differential model registration that resembles a 30% -40% portion of the car when the truck is unloaded.

3. The method improves the registration efficiency and reduces the time consumption. The invention is based on sectional type, only extracts the same part point cloud for registration, and simultaneously performs efficiency optimization based on Ceres. And testing under the condition that the number of the point clouds is 6 ten thousand points in a standard truck point cloud, and reducing the time spent on registering from PCL built-in ICP for two minutes to half minute.

Drawings

FIG. 1 is a flowchart of a method for registering a point cloud model of a moving object according to the present invention.

FIG. 2 is a no-load truck point cloud model in the moving object point cloud model registration method of the present invention.

FIG. 3 is a loaded truck point cloud model in the moving object point cloud model registration method of the present invention.

Detailed Description

The present invention is described in detail below with reference to the attached drawings.

As shown in fig. 1, a method for registering a point cloud model of a moving object includes the following steps of 1, inputting point cloud models obtained by scanning and reconstructing in two moving processes of the object with the same size; step 2, respectively extracting a plurality of distinct and same local point clouds in two models of the object with the same size, and performing primary registration on the two object models based on the extracted same local point clouds; step 3, adjusting the object model based on the registration result to obtain a primary matching result; and 4, respectively extracting the same parts of the two object model bodies, and performing final registration based on the extracted same parts of the objects.

As shown in fig. 2 and fig. 3, in step 1, the registration object of the input point cloud model is obtained by three-dimensional reconstruction after a large-size moving object is scanned by a laser radar during the moving process.

In step 2, matching of the input point Cloud model, taking the proportion of the significant same partial structures of the two object models in the overall structure as a primary matching reference, taking the difference of the distinguishing characteristics of the significant same partial structures in the overall model as a secondary matching reference, and accurately obtaining the significant same partial structures in the point Cloud model through PCL (Point Cloud library) based condition filtering.

In step 2, calculating the coordinates of the positions of the centroids of the structures of the parts with the same significance, adjusting the point cloud by taking the centroids as the origin of the coordinate system, coinciding the coordinates of the positions of the centroids of the two structures with the same significance, and performing ICP registration to realize the primary registration of the two object models. And obtaining rigid transformation matrixes of the two point clouds after the registration of the parts with the same significance is completed, wherein the matrixes describe parameters of the other part with the same significance in the three-dimensional space obtained by the rotation and translation of one part with the same significance, and one object model point cloud position is subjected to rigid transformation to the other model point cloud position of the object based on the rigid transformation matrixes.

In step 4, extracting other local same part point clouds except for the obviously same part structure, calculating and solving a point cloud normal vector, judging and removing part of point clouds of which the normal vector is vertical to a specific plane according to the normal vector, filtering the processed part of point clouds, performing voxel segmentation by adopting an octree, comparing each voxel of the two point clouds, reserving the same part of point clouds, obtaining two similar point clouds subjected to primary coarse registration, wherein the point clouds have smaller difference and roughly overlapped spatial positions, and dynamically adjusting the size of an object point cloud model and the density of the sparse point clouds according to the density of the point clouds based on an ICP (iterative closed point) algorithm frame arranged in a PCL (PCL) library. And optimizing the resolving efficiency of the objective function in the ICP algorithm flow based on a Google open-source Ceres convex optimization algorithm library.

Example 1

This example details the procedure of the matching method. Taking a truck as an example, the flow of the method is as follows.

Inputting object point cloud model data: the registration object applicable to the invention is obtained by three-dimensional reconstruction after the large-size moving object is scanned by the laser radar in the moving process. In the two scanning processes, an object may carry an additional object, so that the input data is a large-size sparse point cloud with two similar main parts and a larger difference part. Taking a common large truck as an example, the truck carries an excessive cargo part through scanning in a loading state, and the point cloud model of the unloaded truck and the loaded truck has a large difference part.

Extracting the same small part of the point cloud main body: based on the condition that the object still has the same small part with obvious characteristics although the object contains a large difference part, the object can be extracted to be subjected to primary registration so as to solve the problem that direct overall registration is affected by the difference part. Still taking a large truck as an example, it can be found by analysis that the resulting head portion is unchanged, whether the large truck is loaded or unloaded. Therefore, the vehicle head parts of the input no-load and loading models can be respectively extracted, the specific principle is that the position near the joint of the vehicle head and the carriage can be found based on the proportion of the vehicle head in the whole vehicle body, the position of the split vehicle head carriage can be found based on the characteristic that the height change of the joint of the vehicle head and the carriage is obvious, and finally the vehicle head can be successfully extracted from the truck model based on PCL conditional filtering.

Same fraction registration: taking the headstock extracted in the above step as an example, the headstock is used as a part of the object, and the headstock is a rigid part no matter the headstock is loaded or unloaded, and the headstock registration result can be used as a reference for overall registration of the object. The vehicle head registration process comprises the steps of calculating position coordinates of two vehicle head mass centers, adjusting point clouds by taking the mass centers as the origin of a coordinate system, performing ICP registration after the two mass centers are overlapped, and directly calling related implementation of an ICP algorithm in a PCL library when the quantity of partial point clouds at the vehicle head is small. And adjusting the object model based on the small part of the registration result to obtain a primary registration result: after the locomotive registration is completed, a rigid transformation matrix of the two point clouds is obtained, and the matrix describes how one locomotive rotates and translates in a three-dimensional space to obtain the parameters of the position of the other locomotive. Based on the matrix, the entire unloaded point cloud is rigidly transformed to a loaded point cloud location.

Extracting the same part of the object body: the object is large in size, and there are many other local identical parts in the subject, except for a relatively obvious small portion of the point cloud. And extracting all the point cloud data to finish the final point cloud registration. Still taking a truck as an example, when the truck is unloaded and loaded, the difference between the parts of the car is large, and the registration accuracy is affected by the difference between the whole part of the car and the registration. Therefore, the same part of the vehicle body is extracted for final registration, the specific extraction step is to calculate and solve a point cloud normal vector, and the point cloud of the part, the normal vector of which is vertical to the ground, is removed according to the judgment of the normal vector. Most of the carriage bottom and goods parts are removed through processing, voxel segmentation is carried out through an octree after filtering processing, each voxel of the two point clouds is compared, and the same partial point clouds are reserved.

Final registration is done based on extracting the same part of the object: after the processing of the previous step, two similar point clouds subjected to primary coarse registration are obtained, the difference between the point clouds is small, and the spatial positions are roughly overlapped. The final step at this point can be fine point cloud registration based on the framework of ICP registration algorithm. The registration algorithm is realized based on an ICP algorithm framework built in a PCL library, and the ICP convergence condition is dynamically adjusted according to the point cloud density and the size of the object point cloud model and the sparse point cloud density. Meanwhile, aiming at the time-consuming problem, the efficiency of the target function calculation in the ICP algorithm flow is optimized based on a Google open-source Ceres convex optimization algorithm library.

The method firstly extracts a small part with obvious characteristics for primary registration, then extracts the same part of the main body of the object for segmented secondary registration, and other methods generally perform registration based on the whole object point cloud. The invention is based on the registration optimization aiming at large-size and sparse point clouds, and other point cloud registration methods are generally only suitable for the registration of dense point clouds with small size in a small range. The method extracts the point clouds of the same part of the object body to carry out final registration, can avoid the point clouds containing larger difference parts, and is superior to other registration methods which are generally only suitable for scenes with high point cloud similarity.

The foregoing is only a preferred embodiment of the present invention, and many variations in the detailed description and the application range can be made by those skilled in the art without departing from the spirit of the present invention, and all changes that fall within the protective scope of the invention are therefore considered to be within the scope of the invention.