CN113989291A - Building roof plane segmentation method based on PointNet and RANSAC algorithm - Google Patents
Building roof plane segmentation method based on PointNet and RANSAC algorithm Download PDFInfo
- Publication number
- CN113989291A CN113989291A CN202111219599.4A CN202111219599A CN113989291A CN 113989291 A CN113989291 A CN 113989291A CN 202111219599 A CN202111219599 A CN 202111219599A CN 113989291 A CN113989291 A CN 113989291A
- Authority
- CN
- China
- Prior art keywords
- point cloud
- segmentation
- roof
- network
- building
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 230000011218 segmentation Effects 0.000 title claims abstract description 103
- 238000000034 method Methods 0.000 title claims abstract description 44
- 238000002372 labelling Methods 0.000 claims abstract description 9
- 238000005192 partition Methods 0.000 claims abstract description 4
- 238000000605 extraction Methods 0.000 claims description 12
- 230000000694 effects Effects 0.000 claims description 8
- 239000011800 void material Substances 0.000 claims description 6
- 238000012549 training Methods 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 3
- 239000000284 extract Substances 0.000 claims description 2
- 238000005065 mining Methods 0.000 claims description 2
- 230000009467 reduction Effects 0.000 claims description 2
- 239000000203 mixture Substances 0.000 claims 1
- 239000002994 raw material Substances 0.000 claims 1
- 230000008569 process Effects 0.000 abstract description 7
- 238000005516 engineering process Methods 0.000 description 7
- 238000012360 testing method Methods 0.000 description 7
- 238000013135 deep learning Methods 0.000 description 5
- 230000006870 function Effects 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 238000012545 processing Methods 0.000 description 4
- 238000005070 sampling Methods 0.000 description 3
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 238000001914 filtration Methods 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000013178 mathematical model Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000002093 peripheral effect Effects 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10028—Range image; Depth image; 3D point clouds
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a building roof plane segmentation method based on PointNet and RANSAC algorithms, which comprises the steps of marking original point clouds, descending and collecting the marked original point clouds to obtain predicted point clouds; inputting the predicted point cloud into a pre-segmentation network, and performing roof pre-segmentation to obtain a pre-segmentation point cloud; aligning the original point cloud with the pre-segmentation point cloud to obtain a restored point cloud; dividing the roof part with semantic information in the reducing point cloud by using an RANSAC algorithm to obtain a roof point cloud; combining the non-roof point cloud and the roof point cloud to obtain a building point cloud, and completing the partition of the roof plane of the building; the invention can process the integral point cloud of the building, and simultaneously simplifies the manual semantic labeling work from each plane of the labeled roof to the roof part of the labeled building, thereby ensuring the segmentation precision and greatly reducing the integral operation time.
Description
Technical Field
The invention relates to the technical field of roof plane segmentation, in particular to a building roof plane segmentation method based on PointNet and RANSAC algorithms.
Background
With the development of building science and technology, a plurality of novel roof structure forms appear, a scientific and modern management means and technology are needed, the management and design are more artistic, the building roof with more aesthetic feeling is effectively managed and designed in the whole life cycle, people begin to introduce the machine vision technology in artificial intelligence, the application research of the technology of network deep learning on the intelligence aspect of small-area license plates is mature, but up to now, the machine vision technology of the method adopts a laser radar shooting device to obtain scene three-dimensional point cloud, the network deep learning is represented by over-fitting and under-fitting, the network regularization and data enhancement capability is not strong, the deep learning generalization capability of the network is seriously insufficient, the robustness of the system in the implementation technology is poor, and the implementation of the three-dimensional reconstruction process of a large-scene large-area building roof is severely restricted.
The processing of the three-dimensional point cloud comprises the steps of three-dimensional point cloud processing technology including sampling, filtering, registering, segmenting, reconstructing, classifying and the like. The point cloud segmentation principle is to segment point cloud data into a plurality of mutually disjoint subsets according to the attribute features of a point cloud area. Point cloud segmentation is an indispensable key step in a building three-dimensional reconstruction process, wherein plane segmentation of a roof is an especially important step, and particularly, a large number of buildings with complex structures and multiple surfaces and layers are built in the development of modern cities, the point cloud distribution is scattered, noise exists, and fine segmentation of the complex building roofs is an important task.
Since 2015, ResNet reduced the ImageNet image classification error rate to 4%, which is less than 5% of human recognition, and application of deep learning in the machine vision field became increasingly common. In recent years, a large number of point cloud segmentation networks emerge, and PointNet is different from other algorithms for projecting point cloud data to a two-dimensional plane or dividing the point cloud data into voxels with spatial dependency, the PointNet network does not need to process the data into a regular 3D voxel form for processing, the input point cloud sequence has no influence on the output result of the network, and meanwhile, the point cloud data after rotational translation can also be processed. However, the method for segmenting each roof by using the deep learning network only needs a large amount of time and hardware cost, needs to provide a data set with a large data amount for learning, has high algorithm complexity, and has the problems that the segmentation of the roof plane by using the PointNet semantic has large data set preparation workload and can not be performed with instance segmentation.
The random sample consensus estimation algorithm (RANSAC) can iteratively estimate the parameters of a mathematical model from a set of observed data sets containing "outliers". The method is an uncertain algorithm, and a reasonable result can be obtained with a certain probability; the number of iterations must be increased in order to increase the probability and the parameters of the algorithm must be carefully selected. The method has the characteristics of high segmentation precision and high speed, and the efficiency is higher even when a large-scale point cloud scene is processed. However, the plane segmentation only by using the RANSAC algorithm also has practical problems, that is, the segmented plane lacks semantic information, and particularly, when the existing building roof structure is complex and various roofs need to be finely segmented, the work of adding semantic labels at a later stage is increased by using the RANSAC algorithm only.
Disclosure of Invention
This section is for the purpose of summarizing some aspects of embodiments of the invention and to briefly introduce some preferred embodiments. In this section, as well as in the abstract and the title of the invention of this application, simplifications or omissions may be made to avoid obscuring the purpose of the section, the abstract and the title, and such simplifications or omissions are not intended to limit the scope of the invention.
The present invention has been made in view of the above-mentioned conventional problems.
Therefore, the building roof plane segmentation method based on the PointNet and RANSAC algorithms provided by the invention can solve the necessary problems of inaccurate segmentation, incapability of automatic work and the like caused by the fact that manual parameter adjustment is needed in a roof plane segmentation task and plane categories cannot be distinguished in the traditional method.
In order to solve the technical problems, the invention provides the following technical scheme: marking original point clouds, and descending and collecting the marked original point clouds to obtain predicted point clouds; inputting the predicted point cloud into a pre-segmentation network, and performing roof pre-segmentation to obtain a pre-segmentation point cloud; aligning the original point cloud with the pre-segmentation point cloud, and adding semantic information to obtain a restored point cloud; dividing the roof part with semantic information in the reducing point cloud by using an RANSAC algorithm to obtain a roof point cloud; and combining the non-roof point cloud and the roof point cloud to obtain a building point cloud, and finishing the partition of the roof plane of the building.
As a preferable scheme of the building roof plane segmentation method based on PointNet and RANSAC algorithms in the present invention, wherein: marking the original point cloud comprises respectively marking the original point cloud as void and roof by using a point cloud marking tool, and representing the void category by white and representing the roof category by purple in a marking effect.
As a preferable scheme of the building roof plane segmentation method based on PointNet and RANSAC algorithms in the present invention, wherein: the reduction comprises setting initial voxels; iteratively increasing the voxels and randomly deleting the points of the labeled original point cloud, and stopping the iteration when the number of the point clouds is less than 1.2 times of the number of the target point clouds; a predicted point cloud of 6000 x 3 points is obtained.
As a preferable scheme of the building roof plane segmentation method based on PointNet and RANSAC algorithms in the present invention, wherein: the pre-segmentation network comprises a feature extraction sub-network and a segmentation sub-network; the feature extraction sub-network is composed of a plurality of MLP layers and extracts high-dimensional features from 3-dimensional original point clouds, wherein 64-dimensional, two 128-dimensional, 512-dimensional and 2048-dimensional features are extracted respectively; extracting global features through a MaxPool layer, and reducing dimensions of the global features through two MLP layers; fusing a plurality of local features extracted by the feature extraction sub-network and the dimension-reduced global feature by using the segmentation sub-network to obtain an n x 3008 feature network; reducing the dimension of the high-dimensional features of the feature network to 64 dimensions through a plurality of MLP layers, performing two-class prediction on each point of the point cloud through the MLP layer with the activation function being Sigmoid, and outputting a n multiplied by 1 segmentation result.
As a preferable scheme of the building roof plane segmentation method based on PointNet and RANSAC algorithms in the present invention, wherein: the dimension of the last MLP layer of the feature extraction sub-network is n multiplied by 1, a group of 0 and 1 labels are output, 0 represents whether the top is the roof or not, namely blank, and 1 represents the top; wherein n is the number of point clouds input into the pre-segmentation network.
As a preferable scheme of the building roof plane segmentation method based on PointNet and RANSAC algorithms in the present invention, wherein: further comprising, training the loss function loss' of the pre-segmentation network as:
wherein σlossIs the standard deviation of the loss(s),
is a weighted average of the losses.
As a preferable scheme of the building roof plane segmentation method based on PointNet and RANSAC algorithms in the present invention, wherein: the method is characterized in that: further comprising recording the original point cloud as pc0The number of the original point clouds is recorded as n, the roof part with semantic information in the reducing point cloud is segmented through an RANSAC algorithm to obtain a roof point cloud pc ', the number of the roof point clouds is n', and the original point cloud is updated to be pc ═ pc at the moment0-pc ', updating the number of point clouds to n-n'; and if the number of the roof point clouds pc' is more than 5% of the number of the original point clouds pc, continuing to perform segmentation through the RANSAC algorithm, and otherwise, finishing the segmentation.
As a preferable scheme of the building roof plane segmentation method based on PointNet and RANSAC algorithms in the present invention, wherein: adding semantic information comprises establishing a sphere by taking each point of the pre-segmentation point cloud as a sphere center and taking 2 times of the distance between two points in the pre-segmentation point cloud as a radius; and adding semantic information to all points in the sphere.
The invention has the beneficial effects that: the invention can process the integral point cloud of the building, and simultaneously simplifies the manual semantic labeling work from each plane of the labeled roof to the roof part of the labeled building, thereby ensuring the segmentation precision and greatly reducing the integral operation time.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without inventive exercise. Wherein:
fig. 1 is a schematic diagram of a pre-segmentation network structure of a building roof plane segmentation method based on PointNet and RANSAC algorithms according to a first embodiment of the present invention;
fig. 2 is a schematic diagram of a roof plane segmentation process of a building roof plane segmentation method based on PointNet and RANSAC algorithms according to a first embodiment of the present invention;
FIG. 3 is a schematic diagram of a point cloud illustration of a building roof plane segmentation method based on PointNet and RANSAC algorithms according to a second embodiment of the present invention;
fig. 4 is a schematic diagram of an actual effect of a segmentation plane of a building roof plane segmentation method based on PointNet and RANSAC algorithms according to a second embodiment of the present invention.
Detailed Description
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, specific embodiments accompanied with figures are described in detail below, and it is apparent that the described embodiments are a part of the embodiments of the present invention, not all of the embodiments. All other embodiments, which can be obtained by a person skilled in the art without making creative efforts based on the embodiments of the present invention, shall fall within the protection scope of the present invention.
In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention, but the present invention may be practiced in other ways than those specifically described and will be readily apparent to those of ordinary skill in the art without departing from the spirit of the present invention, and therefore the present invention is not limited to the specific embodiments disclosed below.
Furthermore, reference herein to "one embodiment" or "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one implementation of the invention. The appearances of the phrase "in one embodiment" in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments.
The present invention will be described in detail with reference to the drawings, wherein the cross-sectional views illustrating the structure of the device are not enlarged partially in general scale for convenience of illustration, and the drawings are only exemplary and should not be construed as limiting the scope of the present invention. In addition, the three-dimensional dimensions of length, width and depth should be included in the actual fabrication.
Meanwhile, in the description of the present invention, it should be noted that the terms "upper, lower, inner and outer" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the drawings, and are only for convenience of describing the present invention and simplifying the description, but do not indicate or imply that the referred device or element must have a specific orientation, be constructed in a specific orientation and operate, and thus, cannot be construed as limiting the present invention. Furthermore, the terms first, second, or third are used for descriptive purposes only and are not to be construed as indicating or implying relative importance.
The terms "mounted, connected and connected" in the present invention are to be understood broadly, unless otherwise explicitly specified or limited, for example: can be fixedly connected, detachably connected or integrally connected; they may be mechanically, electrically, or directly connected, or indirectly connected through intervening media, or may be interconnected between two elements. The specific meanings of the above terms in the present invention can be understood in specific cases to those skilled in the art.
Example 1
Referring to fig. 1 to 2, a first embodiment of the present invention provides a building roof plane segmentation method based on PointNet and RANSAC algorithms, including:
s1: and marking the original point cloud, and descending and collecting the marked original point cloud to obtain a predicted point cloud.
(1) Labeling an original point cloud
The method comprises the steps that marking work is firstly carried out on original point cloud data obtained by a three-dimensional laser scanner before Segmentation, specifically, void and roof are respectively marked on the original point cloud by utilizing a Semantic-Segmentation-Editor developed in Hitachi, the void category is represented by white in a marking effect, and the roof category is represented by purple.
(2) Descending mining
Because the number of point clouds obtained by laser scanning of a common building after the point cloud labeling is finished is usually large, the point cloud labeling method can usually break through 107Therefore, the PCL and the point cloud marking tools cannot be used, and the subsequent calculation is not stressed; in order to solve the above problems, a great number of original point clouds need to be collected downward, and in this embodiment, a voxel filtering method is used to collect the point clouds downward, and the specific process is as follows:
setting initial voxels;
iteratively adding voxels and randomly deleting the points of the marked original point cloud, and stopping iteration when the number of the point clouds is less than 1.2 times of the number of the target point clouds;
and obtaining the predicted point cloud of 6000 multiplied by 3 points.
S2: and inputting the predicted point cloud into a pre-segmentation network, and performing roof pre-segmentation to obtain a pre-segmentation point cloud.
(1) The pre-segmentation network comprises a feature extraction sub-network and a segmentation sub-network; because too much or too little point cloud quantity can cause the efficiency of the pre-segmentation network to be reduced, the embodiment selects 6000 point standards for training and predicting through experiments; in order to solve the problem that the pre-segmentation network tends to be in one category due to the fact that positive and negative categories in the binary classification are unbalanced, the method carries out positive and negative category balance enhancement processing on a data set; for balancing, the number of positive and negative categories is controlled to be 3000 +/-800, the total number is 6000, and the points of the positive and negative categories are split and respectively processed.
(2) Before segmentation, a pre-segmentation network needs to be trained; the loss function and the regularization in the training are related to the final training effect of the network, point cloud segmentation is different from other binary classification models, the output of the point cloud segmentation is a set of n multiplied by 1 numbers from 0 to 1, the traditional binary cross entropy can only calculate the loss of one result, the range of the loss is from 0 to + ∞, and the result of the binary cross entropy on the model is a loss set of n multiplied by 1; if only the mean value is calculated, the loss of the part with larger category number is dominant, and the final result can cause the output of the model to tend to the category, so that the model is over-fitted to the local optimal solution; the loss function loss' of this embodiment is therefore designed as a weighted loss average of the standard deviation of the loss plus 1:
wherein σlossIs the standard deviation of the loss(s),
is a weighted average of the losses.
Regarding network regularization, in the embodiment, the pre-segmentation network uses a regularization mode of Dropout, and the idea is to remove interdependence in features; when each layer has feature units which are discarded randomly, the network is not biased to a certain feature, and the problem of overfitting can be effectively avoided.
(3) Roof pre-segmentation
Firstly, a feature extraction sub-network is composed of a plurality of MLP layers, and high-dimensional features are extracted from 3-dimensional original point clouds, wherein 64-dimensional, two 128-dimensional, 512-dimensional and 2048-dimensional features are respectively extracted; the dimension of the last MLP layer of the feature extraction sub-network is n multiplied by 1, a group of 0 and 1 labels are output, 0 represents whether the roof is adopted, namely blank, and 1 represents the roof; wherein n is the number of point clouds input into the pre-segmentation network.
And secondly, extracting global features through a MaxPool layer, and reducing dimensions of the global features through two MLP layers.
And thirdly, fusing a plurality of local features extracted by the feature extraction sub-network and the dimension-reduced global feature by using the segmentation sub-network to obtain an n x 3008 feature network.
And fourthly, reducing the dimension of the high-dimensional features of the feature network to 64 dimensions through a plurality of MLP layers, performing two-classification prediction on each point of the point cloud through the MLP layer with the activation function being Sigmoid, and outputting a n multiplied by 1 segmentation result.
The original point cloud can be divided into two types of roofs and blanks through a pre-division network, the blank part without semantic information marking is the wall part of a building, and the whole roof part is divided at the moment, and various roofs are not finely divided, so that further division is needed.
S3: and aligning the original point cloud with the pre-segmentation point cloud, and adding semantic information to obtain a restored point cloud.
In order to obtain the maximum efficiency of the RANSAC algorithm, the pre-segmentation point cloud needs to be restored to the original point cloud, and the original point cloud is aligned with the prediction result as the scale of the original point cloud is consistent with that of the prediction result;
secondly, regarding peripheral points of each point of the pre-segmentation point cloud as semantic information with the point, specifically, regarding each point of the pre-segmentation point cloud as a sphere center, and regarding 2 times of the distance between two points in the pre-segmentation point cloud as a radius (the sphere radius is 2 times of the minimum point close distance, the semantic information can be labeled to the maximum extent, and meanwhile, the semantic information range of a prediction result cannot be exceeded), and establishing a sphere; and adding semantic information to all points in the sphere.
S4: and (4) segmenting the roof part with semantic information in the reducing point cloud through an RANSAC algorithm to obtain the roof point cloud.
(1) Recording the original point cloud as pc0The number of the original point clouds is recorded as n, the roof part with semantic information in the reducing point cloud is segmented through an RANSAC algorithm to obtain a roof point cloud pc ', the number of the roof point clouds is n', and the original point cloud is updated to be pc ═ pc at the moment0-pc', updating the number of point clouds ton＝n-n′；
(2) And if the number of the roof point clouds pc' is more than 5% of the number of the original point clouds pc, continuing to perform segmentation through the RANSAC algorithm, and otherwise, finishing the segmentation.
The RANSAC algorithm flow is as follows:
(1) randomly sampling K points;
(2) fitting a model to the K points;
(3) calculating the distance from other points to the model, regarding the distance as an in-office point when the distance is smaller than a certain threshold value, and counting the number of the in-office points;
(4) and repeating the execution for M times, and selecting the model with the most local points as the final model output.
Where M is obtainable by the following formula
z＝1-(1-pk)M
Wherein p represents the probability that a point is an in-office point; k is the number of points which need the least number to solve the model; z is the sampling success rate;
assuming that p is 0.5, since the fitting is a plane, k is 3, and when z is 0.99, M is 34.48, and the best solution is obtained after 35 times of execution.
The input of the algorithm is a restored point cloud with the structure of n multiplied by 3; and (3) quickly fitting a plane model by using a RANSAC algorithm, and obtaining four parameters of a, b, c and d in a plane ax + by + cz ═ d to obtain a plane formula.
Deleting points on the plane from the input point cloud, and continuously fitting the single plane; and repeating the operation until the number of the residual point clouds is less than 5% of the number of the input point clouds.
Preferably, the present embodiment uses the roof part in the point cloud of the whole building as the input of the RANSAC algorithm, and uses the rapidity and accuracy of the RANSAC segmentation plane to finely segment the roof with semantic information.
S5: and combining the non-roof point cloud and the roof point cloud to obtain a building point cloud, and finishing the partition of the roof plane of the building.
And combining the obtained result with the non-roof part point cloud to obtain the final building point cloud and obtain a roof multi-plane segmentation result.
Example 2
The technical effects adopted in the method are verified and explained, and the three test samples are respectively tested by the embodiment, and the real effects of the method are verified by means of the test results of scientific demonstration.
The test samples are respectively a plane pitched roof building, a building with an undulating pitched roof and a large building point cloud, point cloud pictures of three test samples are shown in figure 3, statistics of experimental results of the three test samples are shown in table 1, and actual segmentation effect is shown in figure 4.
Table 1: and testing the experimental result of the sample.
The test result shows that the overall segmentation accuracy of the method reaches 88.2 percent, and the average segmentation accuracy of the actual building roof point cloud can reach 90 percent, and the machine vision learning efficiency is improved by 50 percent compared with a PointNet model; the reality degree of the visual field image of the existing machine vision is improved to a great extent.
It should be noted that the above-mentioned embodiments are only for illustrating the technical solutions of the present invention and not for limiting, and although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that modifications or equivalent substitutions may be made on the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, which should be covered by the claims of the present invention.
Claims (8)
1. A building roof plane segmentation method based on PointNet and RANSAC algorithm is characterized in that: comprises the steps of (a) preparing a mixture of a plurality of raw materials,
marking the original point cloud, and descending and collecting the marked original point cloud to obtain a predicted point cloud;
inputting the predicted point cloud into a pre-segmentation network, and performing roof pre-segmentation to obtain a pre-segmentation point cloud;
aligning the original point cloud with the pre-segmentation point cloud, and adding semantic information to obtain a restored point cloud;
dividing the roof part with semantic information in the reducing point cloud by using an RANSAC algorithm to obtain a roof point cloud;
and combining the non-roof point cloud and the roof point cloud to obtain a building point cloud, and finishing the partition of the roof plane of the building.
2. The building roof plane segmentation method based on PointNet and RANSAC algorithms of claim 1, wherein: labeling the point cloud of origin includes,
and respectively labeling the original point cloud with void and roof by using a point cloud labeling tool, and expressing the void category by white and expressing the roof category by purple in a labeling effect.
3. The building roof plane segmentation method based on PointNet and RANSAC algorithms of claim 2, wherein: the downward mining comprises the steps of,
setting an initial voxel;
iteratively increasing the voxels and randomly deleting the points of the labeled original point cloud, and stopping the iteration when the number of the point clouds is less than 1.2 times of the number of the target point clouds;
a predicted point cloud of 6000 x 3 points is obtained.
4. The building roof plane segmentation method based on PointNet and RANSAC algorithms of claim 1 or 2, characterized in that: the pre-segmentation network comprises a feature extraction sub-network and a segmentation sub-network;
the feature extraction sub-network is composed of a plurality of MLP layers and extracts high-dimensional features from 3-dimensional original point clouds, wherein 64-dimensional, two 128-dimensional, 512-dimensional and 2048-dimensional features are extracted respectively;
extracting global features through a MaxPool layer, and reducing dimensions of the global features through two MLP layers;
fusing a plurality of local features extracted by the feature extraction sub-network and the dimension-reduced global feature by using the segmentation sub-network to obtain an n x 3008 feature network;
reducing the dimension of the high-dimensional features of the feature network to 64 dimensions through a plurality of MLP layers, performing two-class prediction on each point of the point cloud through the MLP layer with the activation function being Sigmoid, and outputting a n multiplied by 1 segmentation result.
5. The building roof plane segmentation method based on PointNet and RANSAC algorithms of claim 4, wherein: also comprises the following steps of (1) preparing,
the dimension of the last MLP layer of the feature extraction sub-network is n multiplied by 1, a group of 0 and 1 labels are output, 0 represents whether the roof is adopted, namely blank, and 1 represents the roof;
wherein n is the number of point clouds input into the pre-segmentation network.
6. The building roof plane segmentation method based on PointNet and RANSAC algorithms of claim 5, wherein: also comprises the following steps of (1) preparing,
the loss function loss' for training the pre-segmentation network is:
wherein σlossIs the standard deviation of the loss(s),
is a weighted average of the losses.
7. The building roof plane segmentation method based on PointNet and RANSAC algorithms of claim 5 or 6, wherein: also comprises the following steps of (1) preparing,
recording the original point cloud as pc0The number of the original point clouds is recorded as n, and a roof part with semantic information in the reduction point cloud is segmented through an RANSAC algorithm to obtain a roof point cloud pc'The number is n', and the original point cloud is updated to pc ═ pc at the moment0-pc ', updating the number of point clouds to n-n';
and if the number of the roof point clouds pc' is more than 5% of the number of the original point clouds pc, continuing to perform segmentation through the RANSAC algorithm, and otherwise, finishing the segmentation.
8. The building roof plane segmentation method based on PointNet and RANSAC algorithms of claim 1, wherein: the adding of the semantic information includes adding the semantic information,
taking each point of the pre-segmentation point cloud as a sphere center, and taking 2 times of the distance between two points in the pre-segmentation point cloud as a radius to establish a sphere;
and adding semantic information to all points in the sphere.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111219599.4A CN113989291A (en) | 2021-10-20 | 2021-10-20 | Building roof plane segmentation method based on PointNet and RANSAC algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111219599.4A CN113989291A (en) | 2021-10-20 | 2021-10-20 | Building roof plane segmentation method based on PointNet and RANSAC algorithm |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113989291A true CN113989291A (en) | 2022-01-28 |
Family
ID=79739550
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111219599.4A Pending CN113989291A (en) | 2021-10-20 | 2021-10-20 | Building roof plane segmentation method based on PointNet and RANSAC algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113989291A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114565772A (en) * | 2022-02-14 | 2022-05-31 | 清华大学 | Set feature extraction method and device, electronic equipment and storage medium |
| CN115239951A (en) * | 2022-06-08 | 2022-10-25 | 广东领慧建筑科技有限公司 | Wall surface segmentation and identification method and system based on point cloud data processing |
| CN115393583A (en) * | 2022-07-21 | 2022-11-25 | 泰瑞数创科技(北京)股份有限公司 | Method for carrying out artificial intelligence semantic segmentation on wall |
-
2021
- 2021-10-20 CN CN202111219599.4A patent/CN113989291A/en active Pending
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114565772A (en) * | 2022-02-14 | 2022-05-31 | 清华大学 | Set feature extraction method and device, electronic equipment and storage medium |
| CN115239951A (en) * | 2022-06-08 | 2022-10-25 | 广东领慧建筑科技有限公司 | Wall surface segmentation and identification method and system based on point cloud data processing |
| CN115239951B (en) * | 2022-06-08 | 2023-09-15 | 广东领慧数字空间科技有限公司 | Wall surface segmentation recognition method and system based on point cloud data processing |
| CN115393583A (en) * | 2022-07-21 | 2022-11-25 | 泰瑞数创科技(北京)股份有限公司 | Method for carrying out artificial intelligence semantic segmentation on wall |
| CN115393583B (en) * | 2022-07-21 | 2023-09-29 | 泰瑞数创科技(北京)股份有限公司 | Method for carrying out artificial intelligence semantic segmentation on wall |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113989291A (en) | Building roof plane segmentation method based on PointNet and RANSAC algorithm | |
| CN104679818B (en) | A kind of video key frame extracting method and system | |
| CN105303195B (en) | A kind of bag of words image classification method | |
| US20230206603A1 (en) | High-precision point cloud completion method based on deep learning and device thereof | |
| CN111008337B (en) | Deep attention rumor identification method and device based on ternary characteristics | |
| CN110322453A (en) | 3D point cloud semantic segmentation method based on position attention and auxiliary network | |
| CN101271469A (en) | Two-dimension image recognition based on three-dimensional model warehouse and object reconstruction method | |
| CN109492596B (en) | Pedestrian detection method and system based on K-means clustering and regional recommendation network | |
| CN110210431B (en) | Point cloud semantic labeling and optimization-based point cloud classification method | |
| CN110472652B (en) | Small sample classification method based on semantic guidance | |
| CN114092697B (en) | Building facade semantic segmentation method with attention fused with global and local depth features | |
| CN109993100A (en) | The implementation method of facial expression recognition based on further feature cluster | |
| CN108595558B (en) | Image annotation method based on data equalization strategy and multi-feature fusion | |
| CN113032613B (en) | Three-dimensional model retrieval method based on interactive attention convolution neural network | |
| CN110377659A (en) | A kind of intelligence chart recommender system and method | |
| CN114330516A (en) | Small sample logo image classification based on multi-graph guided neural network model | |
| CN114926469A (en) | Semantic segmentation model training method, semantic segmentation method, storage medium and terminal | |
| CN107832753B (en) | Face feature extraction method based on four-value weight and multiple classification | |
| CN110363299A (en) | Space reasoning by cases method towards delamination-terrane of appearing | |
| CN114373099A (en) | Three-dimensional point cloud classification method based on sparse graph convolution | |
| CN115713605A (en) | Commercial building group automatic modeling method based on image learning | |
| CN114358279A (en) | Image recognition network model pruning method, device, equipment and storage medium | |
| CN111652309A (en) | Visual word and phrase co-driven bag-of-words model picture classification method | |
| CN117788810A (en) | Learning system for unsupervised semantic segmentation | |
| CN116935138A (en) | Picture subject content diversity calculation and automatic selection method and system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| EE01 | Entry into force of recordation of patent licensing contract |
Application publication date: 20220128 Assignee: SHANGHAI SILIANG ELECTRONIC TECHNOLOGY CO.,LTD. Assignor: Shanghai University of Electric Power Contract record no.: X2024310000032 Denomination of invention: A Building Roof Plane Segmentation Method Based on PointNet and RANSAC Algorithms License type: Common License Record date: 20240327 |
|
| EE01 | Entry into force of recordation of patent licensing contract |