CN109063753A - A kind of three-dimensional point cloud model classification method based on convolutional neural networks - Google Patents
A kind of three-dimensional point cloud model classification method based on convolutional neural networks Download PDFInfo
- Publication number
- CN109063753A CN109063753A CN201810790133.1A CN201810790133A CN109063753A CN 109063753 A CN109063753 A CN 109063753A CN 201810790133 A CN201810790133 A CN 201810790133A CN 109063753 A CN109063753 A CN 109063753A
- Authority
- CN
- China
- Prior art keywords
- point cloud
- model
- data
- dimensional
- ordering
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computational Linguistics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The three-dimensional point cloud model classification method based on convolutional neural networks that the invention discloses a kind of, comprising steps of S1, selection Princeton ModelNet, it is directed to ModelNet10 and ModelNet40 respectively, required amount of model is chosen as training data and test data from official website, generates training set and data set；S2, signature analysis and building taxonomy model are carried out to point cloud model；S3, ordering is carried out to cloud；S4, by orderly point cloud data two dimensional image；S5, CNN network of the building towards two-dimentional point cloud chart picture.The present invention for the first time directly applies to the CNN of image domains in the classification of three-dimensional point cloud model, achieve 93.97% and 89.75% classification accuracy respectively on ModelNet10 and ModelNet40, it is suitable with current the best way, experimental result sufficiently shows, the CNN of image domains, which is applied to three-dimensional point cloud model classification, has feasibility, proposed PCI2CNN can effectively capture the three-dimensional feature information of point cloud model, be suitable for three-dimensional point cloud model and classify.
Description
Technical field
The present invention relates to the technical fields of computer graphics, computer vision and intelligent recognition, refer in particular to a kind of base
In the three-dimensional point cloud model classification method of convolutional neural networks.
Background technique
With the high speed development of modern computer vision research, pilotless automobile, autonomous robot, real-time SLAM technology
There is breakthrough progress with fields such as virtual three-dimensional models, has promoted the available sexual development of three dimensional point cloud, while
Types of applications research for three dimensional point cloud is expedited the emergence of wherein, the classification of point cloud data is the basis of types of applications research
With crucial
Currently, depth learning technology achieves breakthrough in image and field of speech recognition, this is also threedimensional model
Classification provide beneficial research direction however, the input that deep learning model is capable of handling is regular, orderly, point
Cloud data but have the characteristics that it is irregular, unordered, this enable based on deep learning point cloud classifications work become relatively difficult, study
Work also less
Currently, the work in terms of the deep learning towards point cloud data focus primarily on how to construct be suitable for it is irregular and
In terms of the network point cloud of unordered three dimensional point cloud first network PointNet (Charles R Q, Su H,
Kaichun M,et al.PointNet:Deep Learning on Point Sets for 3D Classification
and Segmentation[C]//Proceedings of Computer Vision and Pattern
Recognition.Los Alamitos:IEEE Computer Society Press, 2017:77-85) it is exactly to utilize T-Net
It realizes alignment of data and feature alignment, is converted using the multiple features that the convolution of 1x1 completes point cloud data itself, utilize Max
The extraction of Pooling symmetric function implementation model global statistics information completes point cloud model classification network certainly based on this
Begin to any adjacent point set operation is not carried out eventually, only obtains the various transformation of point itself and the system comprising all the points in a network
Information is counted, so that it is guaranteed that the overall situation that network is only extracted point cloud model due to PointNet [1] to the adaptability of point cloud data is special
Sign, it is difficult to capture the local feature of point cloud model, therefore team where Charles has also been proposed PointNet++ (Qi C R, Yi
L,Su H,et al.Pointnet++:Deep hierarchical feature learning on point sets in a
metric space[C]//Proceedings of Advances in Neural Information Processing
Systems.Cambridge:MIT Press, 2017:5105-5114): sampling and region division are introduced in a network, by small
To it is big, inside from part to whole region using PointNet network carry out feature extraction, and point is extracted by connection entirely
The feature of cloud model, and then the Li Yangyan et al. for Shandong University of classifying proposes the convolutional network PointCNN (Li for cloud
Y,Bu R,Sun M,et al.PointCNN[OL].[2018-06-30].https://arxiv.org/abs/
1801.07791) the work attempts the convolution operation in analog image field, passes through e-learning and constructs a kind of X- transformation solution
The problem of point cloud data randomness, completes point cloud model by the convolution in the layer-by-layer polymerization analog image field of regional area
Feature extraction and classification.
It is found by analysis, the above work all focuses on the scrambling and randomness for how solving point cloud data, and tries
Figure introduces T-Net, X- transformation, symmetric function solves these problems, and achieves certain effect really however these work all need
Special network is designed for point cloud model, the convolutional neural networks of howling success can not will be obtained in field of image recognition
(Convolution Neural Network, CNN) is applied in point cloud model classification.
Summary of the invention
It is an object of the invention to overcome the deficiencies in the prior art, propose a kind of three-dimensional point based on convolutional neural networks
Cloud model classification method the characteristics of for point cloud model, establishes a kind of general three-dimensional point cloud classifications frame based on CNN, lays equal stress on
How point cloud data is converted to acceptable regular, the orderly two-dimentional point cloud chart of CNN as data by point research, how to be constructed and is suitable for
The CNN network of two-dimentional point cloud chart picture is provided with the trial of benefit for the classification work of point cloud model, is CNN in three-dimensional point cloud model
Directly application in classification provides some supports.
To achieve the above object, a kind of technical solution provided by the present invention are as follows: three-dimensional point based on convolutional neural networks
Cloud model classification method, comprising the following steps:
S1, Princeton ModelNet is selected, is directed to ModelNet10 and ModelNet40 respectively, choose institute from official website
It needs the model of quantity as training data and test data, generates training set and data set；
S2, signature analysis and building taxonomy model are carried out to point cloud model；
S3, ordering is carried out to cloud；
S4, by orderly point cloud data two dimensional image；
S5, CNN network of the building towards two-dimentional point cloud chart picture, comprising: the point cloud model based on medium-sized CNN is classified, is based on
The point cloud model of Small-sized C NN is classified and CNN construction and classification towards two-dimentional point cloud chart picture.
In step sl, select Princeton ModelNet, using official website data, for ModelNet10 and
ModelNet40 chooses 3991,9842 models as training data respectively, and 908,2468 models are as test data.
In step s 2, for a cloud randomness, scrambling, finiteness, sparsity feature, three-dimensional point cloud model is designed
Universal classification frame, including following three modules:
The ordering module of point cloud data, for realizing the ordering of unordered point cloud data；
The two dimensional image module of orderly point cloud data, for realizing the regularization of point cloud data；
CNN module towards two-dimentional point cloud chart picture, the module consist of two parts: deconvolution submodule passes through deconvolution
The problem of operation captures the related information between more point cloud datas, makes up point cloud data sparsity to a certain extent；It is middle-size and small-size
Convolution classification submodule to adapt to the finiteness feature of point cloud data prevent the over-fitting of network.
In step s3, three dimensional point cloud M={ (x is inputtedi,yi,zi), i=1 ..., n }, after ordering, put cloud
Sequence be determined, export as ordered sequence S=((xi,yi,zi), i=1 ..., n), x, y, z are respectively three-dimensional point cloud model
Latitude coordinates point；Here, the basic principle of point cloud data ordering is: reaching the point of setting value in three-dimensional space distance, orderly
Distance is also relatively close after change, and this makes it possible to the features for being maximally maintained original point cloud not to be destroyed, and meets image
Positional relationship between the consecutive points of field is based on this basic principle, designs following three kinds of different ordering methods:
Mass center ranking method: the distance at strong point to object mass center, by closely realizing the ordering of point cloud data to remote sequence, this
The advantages of method being ordered into result with point cloud input sequence and model translation, scaling, rotation it is unrelated, but on the other hand
Have the following problems: opposite mass center is symmetrically put originally spatially each other without neighbouring relations, they but may that after ordering
This is adjacent；
One-dimensional ranking method: model is ajusted in advance to be scanned again, and then along some reference axis, is sorted according to coordinate value size real
The ordering of existing point cloud data, this method can either guarantee ordering result and point cloud input sequence and model translation, scaling,
It rotates unrelated, also avoids problem adjacent after the ordering of symmetric position point in space, and object usually to be scanned is in sky
Between in also can satisfy the positive precedence constraint put, only the point cloud data in this way after sequence only embodies some reference axis
Spatial information can not embody the spatial information of other dimensions；
Two-dimentional ranking method: based on one-dimensional ranking method, point cloud data is obtained to the model scanning ajusted in advance, is obtained a little
Cloud data model, it is equally spaced to be divided into m slice along some reference axis for the point cloud data model, it is fitted when m chooses
When value, it can be seen that the value in the reference axis inside same slice is suitable, i.e., these points are generally aligned in the same plane inside, this
When, to each slice, sorts again according to some reference axis in plane coordinates, complete the ordering of point cloud data, this side
Method not only can guarantee the advantages of one-dimensional ranking method, but also can preferably embody the spatial information of different dimensions.
In step s 4, ordering sequence the S=((x of unordered point cloud data is inputtedi,yi,zi), i=1 ..., n), this step
Suddenly it is intended to reasonably for ordering sequence to be placed on two dimensional image A=(ajk) in p × q, wherein p × q=n, A are two-dimensional matrix,
The corresponding image generated；J and k is respectively the row and column of pixel；ajkThe pixel value arranged for jth row kth in two-dimensional matrix；P, q points
Not Wei matrix line number and columns；N is indicated to put number included in point cloud model, be corresponded between adjacent pixel inside image with meeting
Point cloud data it is near one another on spatial position, for this requirement, design following three kinds of different two dimensional image methods:
Row scanning method: imitate fluorescent screen electron beam movement, orderly point cloud data is successively taken out from front to back, from a left side to
The right side is filled into two dimensional image line by line from top to bottom, and this mode can guarantee laterally adjacent pixel in original point cloud data
Near one another, it is near one another in original point cloud data to be but unable to ensure longitudinally adjacent pixel, that is, does not have isotropism；
Chessboard method: in view of CNN using local receptor field thought extract characteristics of image, if can by the part of cloud with
The part of image is mapped, then can preferably extract the local feature of point cloud data, it is therefore proposed the image of chessboard method
Change method: orderly point cloud data is successively taken out from front to back, is sequentially filled each grid, grid from left to right, from top to bottom
Each pixel is filled still according to mode from left to right, from top to bottom in inside, and when grid value is 8 × 8, each grid is just
A corresponding point cloud local region comprising 64 points, this method do not have isotropism equally；
Spiral method: orderly point cloud data is successively taken out from front to back, since picture centre pixel, by helical trajectory according to
Secondary to be filled, this mode can be good at keeping isotropism, and can be good at keeping in the position by pericenter
The distance relation of luv space point, but there is also the defects of itself: closer to edge, pixel will more disperse, Yi Xie
Space distance after close point filling may become larger.
In step s 5, construction is suitable for the convolutional neural networks of two-dimentional point cloud chart picture, since point cloud data is with limited
Property and sparsity, the large-scale CNN more than the number of plies may cause over-fitting, therefore, will choose first one medium-sized CNN and one it is small-sized
CNN carries out preliminary experiment；
Point cloud model classification based on medium-sized CNN: point cloud data has the characteristics that Limited information and sparsity, and medium-sized
Network towards data size it is big, when the input of AlxeNet is having a size of 224 × 224, point cloud that input size is 1024
The corresponding two dimensional image size of model is only 32 × 32, thus, by point cloud chart picture input before, first to image data into
Row deconvolution operation meets medium-sized CNN input size requirements, avoids over-fitting, while realizing the high-resolution weight of point cloud chart picture
It builds, obtains more space correlation information；
Point cloud model classification based on Small-sized C NN: LeNet is one and connects entirely comprising two layers of convolution, two layers of pond and three layers
The Small-sized C NN network connect, is mainly used for handwriting recongnition, and the input of the network is the image of 32 × 32 sizes, just with comprising
The two-dimentional point cloud chart of 1024 points is matched as size, and therefore, the experiment for directly continuing to use the point cloud model classification based on medium-sized CNN is set
Meter, removes deconvolution submodule, and 32 × 32 two-dimentional point cloud chart picture input LeNet is completed feature extraction and category of model；
CNN construction and classification towards two-dimentional point cloud chart picture: analysis point cloud model own characteristic is tested in conjunction with both the above
As a result, one convolutional neural networks PCI2CNN towards the classification of two-dimentional point cloud chart picture of design, the mentality of designing of the network are as follows:
Comprising one group of totally 2 deconvolution, this group of deconvolution port number is 64, and core size is 2 × 2, and step-length is respectively 2 × 2
With 1 × 1, to operate the super-resolution reconstruction for realizing two-dimentional point cloud chart picture by deconvolution, the related letter between more points cloud is obtained
Breath；
Comprising 3 convolutional layers, and port number is respectively 64,128,256；It include less network layer compared to AlxeNet
Secondary and parameter improves the stability of network training to avoid the complexity of network；It include more parameters compared to LeNet, with
Improve the ability of network fitting training data；
Pondization operation is added after first and third convolutional layer, and pond layer port number is consistent with upper one layer of port number,
Core size is 3 × 3, and step-length is 2 × 2, to obtain information more abundant by overlap sampling.
Compared with prior art, the present invention have the following advantages that with the utility model has the advantages that
1, the invention proposes a kind of general three-dimensional point cloud model taxonomy model based on CNN, overcomes three dimensional point cloud
Disordering, sparsity, finiteness, support the three-dimensional point cloud model based on each class two-dimensional CNN effectively to classify, and classification accuracy
It is higher.
2, the present invention proposes three kinds of ordering methods for the existing feature of point cloud model, realizes unordered point cloud data
Ordering, it is ensured that the relative distance relationship between point cloud data utmostly keeps the integrality of original point cloud feature, meets image
Adjacent pixel each other space correlation the characteristics of, and meet basic demand of all kinds of deep learning algorithms to data order.
3, the present invention proposes three kinds of possible ways for the two dimensional image of orderly point cloud data, realizes orderly point cloud
The two dimensional image of data guarantees that adjacent pixel corresponds to point cloud data spy substantially adjacent to each other on spatial position inside image
Point, and then support three dimensional point cloud feature extraction and classification based on two-dimentional CNN.
4, PCI2CNN proposed by the invention can effectively capture the three-dimensional feature information of point cloud model, achieve higher
Classification accuracy, be suitable for three-dimensional point cloud model classify.
Detailed description of the invention
Fig. 1 is the three-dimensional point cloud model taxonomy model figure based on CNN.
Fig. 2 is two-dimentional ranking method schematic diagram
Fig. 3 is the two dimensional image figure using row scanning method.
Fig. 4 is the two dimensional image figure using chessboard method.
Fig. 5 is the two dimensional image figure using spiral method.
Fig. 6 is the network architecture diagram of AlexNet.
Fig. 7 is deconvolution submodule figure.
Fig. 8 is the point cloud model classification performance figure (ModelNet10) based on AlexNet.
Fig. 9 is LeNet network architecture diagram.
Figure 10 is the point cloud model classification performance figure (ModelNet10) based on LeNet.
Figure 11 is PCI2CNN network architecture diagram.
Figure 12 is the classification performance comparison diagram (ModelNet10) of three networks.
Figure 13 is the classification performance comparison diagram (ModelNet10) of different ordering methods.
Figure 14 is different slice numbers classification performance comparison diagrams (ModelNet10).
Figure 15 is the classification performance comparison diagram (ModelNet10) of different two dimensional image methods.
Figure 16 is the classification performance comparison diagram (ModelNet10) of three cloud networks.
Specific embodiment
The present invention is further explained in the light of specific embodiments.
Three-dimensional point cloud model classification method based on convolutional neural networks provided by the present embodiment, three kinds of major design
The ordering method of three dimensional point cloud, the two dimensional image method of three kinds of orderly point cloud datas are suitable for two-dimentional point cloud chart picture
The convolutional neural networks PCI2CNN of classification, specifically includes the following steps:
S1, Princeton ModelNet is selected, is directed to ModelNet10 and ModelNet40 respectively, choose one from official website
The model of fixed number amount generates training set and data set as training data and test data；Specifically, Princeton is selected
ModelNet chooses 3991,9842 models as instruction for ModelNet10 and ModelNet40 using official website data respectively
Practice data, 908,2468 models are as test data.
S2, signature analysis and building taxonomy model are carried out to point cloud model
The features such as cloud randomness, scrambling, finiteness, sparsity, the present invention devises three-dimensional point cloud model
Universal classification frame, including following three modules, as shown in Figure 1.
The ordering module of point cloud data, realizes the ordering of unordered point cloud data；
The two dimensional image module of orderly point cloud data, realizes the regularization of point cloud data；
CNN module towards two-dimentional point cloud chart picture, the module mainly consist of two parts: deconvolution submodule, by anti-
The problem of convolution operation captures the related information between more point cloud datas, makes up point cloud data sparsity to a certain degree；It is middle-size and small-size
Convolution classification submodule to adapt to the finiteness feature of point cloud data prevent the over-fitting of network.
S3, ordering is carried out to cloud
It inputs three dimensional point cloud M={ (xi, yi, zi), i=1 ..., n }, after ordering, the sequence for putting cloud is true
It is fixed, it exports as ordered sequence S=((xi, yi, zi), i=1 ..., n), x, y, z is respectively the latitude coordinates of three-dimensional point cloud model
Here, the basic principle of point cloud data ordering is point: in the point that three-dimensional space is closer, distance is also opposite after ordering
The feature that nearlyr can thus be maximally maintained original point cloud is not destroyed, and is met between image domains consecutive points
Positional relationship be based on this basic principle, the present invention devises following three kinds different ordering methods:
Mass center ranking method: the distance at strong point to object mass center, by the ordering party for closely realizing point cloud data to remote sequence
The advantages of method being ordered into result with point cloud input sequence and model translation, scaling, the unrelated of rotation but on the other hand also deposit
In following problems: opposite mass center is symmetrically put originally spatially each other without neighbouring relations, they but may be each other after ordering
It is adjacent.
One-dimensional ranking method: model is ajusted in advance to be scanned again, and then along some reference axis, is sorted according to coordinate value size real
Existing this method of the ordering of point cloud data can both guarantee ordering result and point cloud input sequence and model translation, scaling,
Rotate it is unrelated, also avoid problem adjacent after the ordering of symmetric position point in space and, usually object to be scanned is in sky
Between in also to can satisfy the positive precedence constraint put be that the point cloud data after sequence so only embodies some reference axis
Spatial information can not embody the spatial information of other dimensions.
Two-dimentional ranking method: as shown in Fig. 2, based on one-dimensional ranking method, point cloud is obtained to the model scanning ajusted in advance
Data is directed to the point cloud data model, equally spaced to be divided into m slice when m chooses conjunction along some reference axis, such as Z axis
In due course, it is considered that the Z value inside same slice is suitable, i.e., these points are generally aligned in the same plane inside.At this time, to each slice,
Again according to some reference axis sequence in plane coordinates, completing the ordering of point cloud data this method can ensuring method
Two the advantages of, and can preferably embody the spatial information of different dimensions.
S4, by orderly point cloud data two dimensional image
Input ordering sequence the S=((x of unordered point cloud datai,yi,zi), i=1 ..., n) this section is intended to ordering
Sequence is reasonably placed on two dimensional image A=(ajk) in p × q, wherein p × q=n, (A is two-dimensional matrix, the corresponding figure generated
Picture；J and k is respectively the row and column of pixel；ajkThe pixel value arranged for jth row kth in two-dimensional matrix；P, q is respectively the row of matrix
Several and columns；N indicates to put number included in point cloud model), to meet corresponding point cloud data between the adjacent pixel of image inside
Near one another is directed to this requirement on spatial position, we devise following three kinds different two dimensional image method
Row scanning method: as shown in figure 3, imitating the movement of fluorescent screen electron beam, from front to back successively by orderly point cloud data
It takes out, being filled into this mode of in two dimensional image line by line from left to right, from top to bottom can guarantee laterally adjacent pixel original
Near one another in point cloud data, it is near one another in original point cloud data to be but unable to ensure longitudinally adjacent pixel, that is, does not have each
To the same sex.
Chessboard method: in view of CNN extracts characteristics of image using the thought of local receptor field, if we can be by the office of cloud
Portion is mapped with the part of image, then can preferably extract the local feature of point cloud data thus, the invention proposes chesses
The image conversion method of disk method: orderly point cloud data is successively taken out from front to back, according to method shown in Fig. 4, from left to right, from
Top to bottm is sequentially filled each grid, fills each pixel still according to mode from left to right, from top to bottom inside grid and works as
When grid value is 8 × 8, each grid, which just corresponds to the point cloud local region this method of comprising 64 points, not to be had equally
Standby isotropism.
Spiral method: orderly point cloud data is successively taken out from front to back, as shown in figure 5, being pressed since picture centre pixel
Helical trajectory, which is successively filled this mode of, can preferably keep isotropism, and can in the position by pericenter
Keep the distance relation of luv space point well, but there is also the defects of itself: i.e. closer to edge, pixel will be got over
Dispersion, some in space, distance may be larger after closer point filling.
S5, construction are suitable for the convolutional neural networks of two-dimentional point cloud chart picture, since point cloud data has finiteness and sparse
Property, the excessive large-scale CNN of the number of plies may cause over-fitting, therefore, the present invention will choose first one medium-sized CNN and one it is small-sized
CNN carries out preliminary experiment.
Point cloud model classification based on medium-sized CNN: as shown in fig. 6, AlexNet is the convolution mind comprising 8 layers of structure
Through network, medium size network in CNN Models Sets is belonged to for model depth simultaneously, the network is big in image recognition in 2012
Win the championship title in match, recognition capability be worth affirmative thus, select AlexNet as experimental subjects in the present invention first,
It extracts point cloud model feature and classifies.
Point cloud data has the characteristics that Limited information and sparsity, and medium size network towards data size often compared with
Greatly, if the input of AlxeNet is having a size of 224 × 224, and X-Y scheme corresponding to the point cloud model that an input size is 1024
As size be only 32 × 32. thus, by point cloud chart picture input before, the present invention first to image data carry out deconvolution behaviour
Make, meets medium-sized CNN input size requirements, avoid over-fitting；It realizes the super-resolution reconstruction of point cloud chart picture, obtains more empty
Between related information.
As shown in Figure 7, using the three dimensional point cloud comprising 1024 points as standard, i.e., with 32 × 32 two-dimentional point cloud chart picture
For input, we construct the deconv indicated beside the deconvolution submodule deconvolution comprising 3 groups of deconvolution operations
(a, b, c) indicates that the deconvolution operating walk way number is a, and the size of core is b × b, and step-length is c × c.3 group deconvolution operation is logical
Road number is respectively 16,32,64；The size of core is respectively that 2 × 2,4 × 4,6 × 6. every groups of deconvolution operations are internal logical comprising two
Road number is identical with the size of core, and wherein, the step-length of first deconvolution is set as 2 × 2, second to the different convolution operation of step-length
The step-length of a deconvolution is set as 1 × 1, and the size of each deconvolution core can just be divided exactly by step-length, to avoid deconvolution
Occur chessboard effect in journey, improves the generation quality of high-definition picture.
Based on the above deconvolution submodule, 32 × 32 two-dimentional point cloud chart picture can be redeveloped into 279 × 279 high-resolution
Rate point cloud chart picture, then by bilinearity difference by 227 × 227 required by image scaling to AlexNet network, as input
Realize the feature extraction and classification of point cloud model.
In order to verify the validity of this method, we choose rigid three-dimensional model data collection Princeton ModelNet's
Subset ModelNet10 as test object to the network make a preliminary test due to the threedimensional model in ModelNet10
Forward direction is placed, and in experiment, we include the three of 1024 points using the building of three-dimensional point cloud scanning algorithm provided by PointNet
Point cloud data is tieed up, and its two-dimentional point cloud chart picture is constructed based on Z axis sequence and row scanning mode, is operated by deconvolution shown in Fig. 7
227 × 227 high resolution 2 d point cloud chart picture is obtained with zoom operations, input AlexNet completes feature extraction and classification is real
Testing result as shown in Figure 8 can see, and in 250 step of iteration or so, algorithmic statement, this is also first in 90% or more for classification accuracy
Step demonstrates the validity of inventive algorithm
But AlexNet is a medium-sized CNN, includes 8 layers, there are about 138M for its own parameter amount, add 6 layers of warp
Product so that whole network framework is than cumbersome, and parameter amount is big, time consumption for training, and the relatively low of performance is simultaneously, because point cloud data only wraps
Coordinate information containing cloud itself does not include any topological connection relation, and information content is limited and extremely sparse, excessively complicated net
Network can not extract more useful informations, possibly even cause over-fitting, for this reason, we are by one NN pairs of Small-sized C of reselection
This paper algorithm carries out test
Point cloud model classification based on Small-sized C NN: as shown in figure 9, LeNet, which is one, includes two layers of convolution, two layers of pond
With three layers of Small-sized C NN network connected entirely (Lecun Y, Bottou L, Bengio Y, et al.Gradient-based
learning applied to document recognition[J].Proceedings of the IEEE,1998,86
), (11): 2278-2324 the input for being mainly used for the handwriting recongnition network is the image of 32 × 32 sizes, just with comprising
Therefore as size matching, we directly continue to use the reality of the classification of the point cloud model based on medium-sized CNN to the two-dimentional point cloud chart of 1024 points
Design is tested, deconvolution submodule is removed, 32 × 32 two-dimentional point cloud chart picture input LeNet is completed into feature extraction and model point
Class.
The results are shown in Figure 10 for classification experiments based on LeNet network, in figure 5 it can be seen that 1, in 250 step of iteration or so, algorithm
Convergence, for classification accuracy 88% or so, this demonstrates the validity of this paper algorithm again；2, and based on AlexNet network
Nicety of grading is compared, the nicety of grading under the network reduce 2%. this may be of both caused by reason, first, lack
Deconvolution submodule can not obtain the related information between more point cloud datas；Second, LeNet network are to identify digital handwriting
Body and design, be suitable for capturing linear feature, it is difficult to capture two-dimentional point cloud chart as the feature of this kind of complex image；3,
Compared with the classification performance curve based on AlxeNet network, the classification performance curve under the network is relatively steady, this is because point
The information content that cloud model is capable of providing is limited, and the AlexNet network number of plies is more, and parameter is more, and parameter space is relative complex, therefore parameter
It just will appear during adjustment and change violent problem；And the LeNet network number of plies and the complexity of parameter space and point cloud can
The information content of offer is opposite to be matched, therefore nicety of grading curve is relatively steady.
CNN construction and classification towards two-dimentional point cloud chart picture；Point cloud model own characteristic is analyzed, in conjunction with above two groups of experiments
As a result, we devise one towards two-dimentional point cloud chart picture classification convolutional neural networks PCI2CNN. it is as shown in figure 11, the net
The mentality of designing of network is as follows:
Comprising one group of totally 2 deconvolution, this group of deconvolution port number is 64, and core size is 2 × 2, and step-length is respectively 2 × 2
With 1 × 1, to operate the super-resolution reconstruction for realizing two-dimentional point cloud chart picture by deconvolution, the related letter between more points cloud is obtained
Breath.
Comprising 3 convolutional layers, and it compared to AlxeNet includes less network layer that port number, which is respectively 64,128,256.,
Secondary and parameter improves the stability of network training to avoid the complexity of network；It include more parameters compared to LeNet, with
Improve the ability of network fitting training data.
Pondization operation is added after first and third convolutional layer, and pond layer port number is consistent with upper one layer of port number,
Core size is 3 × 3, and step-length is 2 × 2, to obtain information more abundant by overlap sampling.
The experimental design for continuing to use the point cloud model classification based on medium-sized CNN, the threedimensional model in ModelNet10 is converted
For 32 × 32 two-dimentional point cloud chart picture, input PCI2CNN and complete feature extraction and category of model, and by with AlxeNet and
LeNet compares, and obtain test result as shown in figure 12 as seen from the figure: PCI2CNN proposed by the invention is in ModelNet10
On obtain 92% or so classification accuracy, 88% classification accuracy compared with LeNet is obviously improved and AlexNet phase
Than accuracy rate about improves 2%, and stability, which also has, to be obviously improved；Meanwhile according to a preliminary estimate known to the parameter of the two,
The parameter amount of PCI2CNN more meets two compared to 95.5%. experimental result explanation, the PCI2CNN designed by us is latter reduced
Tie up point cloud chart as the characteristics of, classification performance is good, stability is high and ginseng negligible amounts simultaneously, in summary three groups of experiments can also be said
The validity of the bright taxonomy model towards three-dimensional point cloud model designed by us.
This experimental situation is that the deep learning frame Tensorflow in Ubuntu14.04 operating system, based on open source is real
PCI2CNN is showed, hardware platform is seven rainbow gtx of intel i7 2600K+, 1060 6G+8G RAM.
This experiment is intended to test the above-mentioned three-dimensional point cloud model classification method based on convolutional neural networks of the present embodiment
Classification capacity.
Data set Princeton ModelNet used by this experiment is a rigid three-dimensional model data collection, internal mode
It is herein that type along Z axis forward direction has put, compares for convenience of with other work, selects two subsets of the data set
ModelNet10 and ModelNet40 is that benchmark data set is completed to the test of PCI2CNN wherein: ModelNet10 contains 10 classes
Amount to 4899 rigid models, divided using official, 3991 are used as training sample, and 908 are used as test sample；
ModelNet40 contain 40 classes amount to 12311 rigid models, divided using official, 9842 be used as training sample, 2468
As test sample as benchmark, uniformly adopted on triangular topological relations using Point Cloud Library (PCL) tool
1024 points of sample, and they are normalized in unit ball, the point cloud model of given triangle gridding is obtained with this.
In the training process, each epoch upsets putting in order for training sample at random, and trains as input
PCI2CNN. simultaneously, to expand training data, reduce the over-fitting of network, the robustness of prediction is improved, in training
The data of each batch, random given one belongs to the angle, θ of [0,2 π], and point cloud model is made to rotate θ degree around Z-direction, and
The Gaussian noise for adding (0,0.02) carries out randomized jitter to point cloud data, to expand training data.
Firstly, the present invention has made following test and analysis to the ordering method of point cloud data: using ModelNet10 as base
Quasi- data, based on three kinds of ordering methods proposed by the invention, we construct four kinds of different point cloud datas first
Ordering method: mass center ranking method and 3 kinds of one-dimensional ranking methods, i.e., along X-axis, Y-axis, the sequence of Z axis coordinate；Then row scanning method is used
The two-dimentional point cloud chart picture of point cloud model is established, and inputs PCI2CNN and completes classification, experimental result is as shown in figure 13.
As seen from Figure 13: being 1. above 89% by nicety of grading obtained by different ordering modes；2. pressing the ordering of Z axis coordinate
Classification results afterwards are substantially better than other methods, and nicety of grading is about 92%, improve about 2% compared with other methods；3. being sat by X-axis
Nicety of grading obtained by mark, Y axis coordinate and mass center ranking method is suitable, and in this explanation of 90% or so: 1. this paper algorithm frame has
Validity, classification method entirety classification performance based on this are preferable；2. since the model in ModelNet10 is according to Z
Axis forward direction is put, and the data after this direction sequencing are more in line with the space structure of object and the visual cognition of the mankind, thus its
Accuracy rate highest；3. since object can be rotated at random in Z-direction, i.e., in X-axis, Y-axis both direction the posture of object be with
Meaning, therefore sort along the two reference axis without evident regularity, thus classification accuracy is close, and low compared with sorting along Z axis；4. pressing
The reason of method of centroid distance sequence is because of symmetry is arrived according to point, after relatively far away from sorting there are space length adjacent to each other
Problem, thus result is general.
Based on known to the above experiment, by the ordering method that Z axis sorts achieve best classification performance thus, to two
When dimension ranking method is tested and analyzed, we sort first, in accordance with Z axis, are then equidistantly sliced, and slice is internal then to press Y
Coordinate rearrangement forms in final ordering result experiment us and has chosen 16,32,64,128,1024 respectively (directly
Sort, be not sliced according to Z axis) slices of five kinds of different numbers, experimental result is as shown in figure 14.
As seen from Figure 14: classifying quality is most preferably sliced for 64, has been up to 93.97% test accuracy rate, phase
Than only improving about 2% by Z coordinate sequence；Classifying quality it is worst be 16 slice, or even not as good as only by Z coordinate sort；
32 slices close with the classifying quality that 128 are sliced when using 64 slices, is sliced this is because relative to 1024 points
Inside averagely possesses 16 points, and Z axis coordinate is approximate, and contains certain plane information, at this time according to Y-axis minor sort again,
Can obtain than only sorting the spatial informations of more various dimensions by Z axis, thus the best of classification performance but work as slice numbers and constantly increase
Added-time, the number for being sliced internal point are constantly reduced, can not one plane of effective expression again geological information, thus at this time again along Y
Axis sequence effect constantly reduces instead；Opposite, when slice numbers are constantly reduced, the internal corresponding point cloud data mistake of slice
It is more, at this time they Z value difference it is larger, then by Y-axis sequence may break originally by Z axis sort advantage, classifying quality into
And it is deteriorated.
According to the equidistant cutting 64 from small to large of Z axis coordinate in subsequent experimental of the present invention, slice is internal to press Y coordinate again
The mode of sequence completes the ordering of point cloud data.
Secondly, the present invention is tested and is analyzed as follows to the two dimensional image method of orderly point cloud: using ModelNet10 as base
Quasi- data, after ordering, we construct orderly point cloud number according to three kinds of row scanning, chessboard method and spiral method different modes respectively
According to two dimensional image, and input PCI2CNN complete classification experimental result it is as shown in figure 15, it can be seen that: in the way of row surface sweeping
Highest classification accuracy is achieved, is higher by about 1% compared with other two methods；Chessboard method and the suitable of spiral method effect carefully divide
Analysis can be found: although not having isotropism, only for single direction, pixel by the two dimensional image method of row scanning
Between relationship and point cloud between relationship there is consistent corresponding relationship each other；And in spiral method and chessboard method, adjacent column
Or the relationship of adjacent pixel in the ranks is not with uniformity, margin location of the closer cloud of script in image such as in spiral method
Setting may meet farther out, and the pixel of adjacent grid intersection is in orderly point cloud but at a distance of an almost grid in chessboard method
Distance, this affects classifying quality to a certain extent.
Finally, the present invention has done to when analysis classification experiments result: having chosen several typical threedimensional models herein
Classification method is compared with PCI2CNN, and including the PointCNN that nearest Shandong University Li Yang man of virtue and ability et al. proposes, experimental data is such as
Shown in table 1, ModelNet10 and ModelNet40 will be directed to respectively below and provide experimental analysis.
Classification experiments result on 1 ModelNet data set of table, "-" indicate corresponding entry information in correlative theses not
It provides
1) experimental result and analysis on ModelNet10
On ModelNet10, this paper algorithm achieves 93.97% classification accuracy, ranking in all methods
Two, are that the network of input can obtain with it since point cloud data includes that information content is few, and has randomness and scrambling simultaneously
The information taken is also few compared with other kinds of network, and therefore, such classification accuracy sufficiently demonstrates the validity of context of methods
It detailed comparison and is analyzed as follows:
Compared with the method based on three-dimensional voxel, context of methods better than in addition to VRN Ensemble all method but
That can see from parameter amount, the parameter of network constructed by this paper be only the 1/45. of VRN Ensemble and other parts and this
The literary comparable voxel method of accuracy rate, if VRN parameter amount is also 9 times of this paper, ORION method then both needs to input voxel model
The directional information for also needing to be provided previously.
Compared with the method based on multiple view, information that the classification accuracy highest of context of methods includes due to multiple view
Amount is big, and often complex as the network of input, also larger is specific for time consumption for training, compares DeepPano, Pairwise
5.31% and 1.17% has been respectively increased in this paper network class accuracy rate, and parameter amount reduces about 2 quantity compared with Pairwise
Grade.
Compared with other networks based on point cloud data, the classification accuracy ratio PointNet and PointNet of context of methods
(vanilla) 0.89% and 2.01%. Figure 16 has been respectively increased and has illustrated the measuring accuracy song of three networks in the training process
Line, abscissa indicates epoch number of training in figure, ordinate expression measuring accuracy as we can see from the figure, from primary iteration
To final convergence, the classification performance of this paper network is superior to PointNet and PointNet (vanilla), and classifying quality is stablized；
And in epoch value 25 or so, the measuring accuracy of context of methods alreadys exceed 90%, and it is good that this also sufficiently demonstrates the network
Generalization Capability.
Further analysis it can be found that, in three dimensions, put the three-dimensional point of cloud all and be it is isolated existing, between points
There is no direct connection relationship, but there is positional relationship to each other in PointNet network, all process of convolution are all
For single 3 D point, extraction is characterized in the abstract of the three-dimensional coordinate of a single point, and is finally obtained using maximum pond layer
Be comprising the global information including all three-dimensional point clouds, thus will to put cloud herein orderly by the feature extraction for lacking to reference point
Switch to two-dimensional matrix after change, remain the relativeness between three-dimensional space point to a certain extent, thus can cross capture compared with
The more characteristic informations of PointNet, therefore classification accuracy is higher.
2) experimental result and analysis on ModelNet40
On ModelNet40, the classification accuracy of this paper algorithm is 89.75%, whole ranking placed in the middle and voxel method
And the comparison of multiple view method is similar with ModelNet10, repeats no more here, selective analysis once various depths based on cloud
Spend learning method.
Compared with network PointNet and PointNet (vanilla) towards point cloud classifications, the classification of this paper algorithm is quasi-
True rate has 0.55% and 2.55% promotion respectively, and parameter amount reduces the analysis of 39%. concrete reasons together compared to PointNet
ModelNet10。
Compared with PointNet++ and PointCNN, the classification accuracy of context of methods reduces 0.95% He respectively
1.95%, this is because it can be more complete, effective in PointNet++ and PointCNN there is also a certain distance
The characteristic information of three-dimensional point cloud network different scale is captured, herein, although capturing local feature information using CNN, by
In that could not comprehensively consider the relationship in ordering, two dimensional image and CNN between deconvolution and convolution kernel and step-length, there may be offices
The problem of portion's Feature capturing is imperfect or is mutually mixed is worth further research and thinking.
Therefore, in conclusion the experiment in ModelNet10 and ModelNet40 has absolutely proved having for context of methods
Effect property, is worthy to be popularized.
Embodiment described above is only the preferred embodiments of the invention, and but not intended to limit the scope of the present invention, therefore
All shapes according to the present invention change made by principle, should all be included within the scope of protection of the present invention.
Claims (6)
1. a kind of three-dimensional point cloud model classification method based on convolutional neural networks, which comprises the following steps:
S1, Princeton ModelNet is selected, is directed to ModelNet10 and ModelNet40 respectively, number needed for being chosen from official website
The model of amount generates training set and data set as training data and test data；
S2, signature analysis and building taxonomy model are carried out to point cloud model；
S3, ordering is carried out to cloud；
S4, by orderly point cloud data two dimensional image；
S5, CNN network of the building towards two-dimentional point cloud chart picture, comprising: the point cloud model based on medium-sized CNN is classified, based on small-sized
The point cloud model of CNN is classified and CNN construction and classification towards two-dimentional point cloud chart picture.
2. a kind of three-dimensional point cloud model classification method based on convolutional neural networks according to claim 1, feature exist
In: in step sl, Princeton ModelNet is selected, using official website data, for ModelNet10 and ModelNet40
3991,9842 models are chosen respectively as training data, and 908,2468 models are as test data.
3. a kind of three-dimensional point cloud model classification method based on convolutional neural networks according to claim 1, feature exist
In: in step s 2, for a cloud randomness, scrambling, finiteness, sparsity feature, design the logical of three-dimensional point cloud model
With taxonomy model, including following three modules:
The ordering module of point cloud data, for realizing the ordering of unordered point cloud data；
The two dimensional image module of orderly point cloud data, for realizing the regularization of point cloud data；
CNN module towards two-dimentional point cloud chart picture, the module consist of two parts: deconvolution submodule is operated by deconvolution
The problem of capturing the related information between more point cloud datas, making up point cloud data sparsity to a certain extent；Middle-size and small-size volume
Integrating class submodule prevents the over-fitting of network to adapt to the finiteness feature of point cloud data.
4. a kind of three-dimensional point cloud model classification method based on convolutional neural networks according to claim 1, feature exist
In: in step s3, input three dimensional point cloud M={ (xi,yi,zi), i=1 ..., n }, after ordering, put the suitable of cloud
Sequence is determined, and is exported as ordered sequence S=((xi,yi,zi), i=1 ..., n), x, y, z are respectively the dimension of three-dimensional point cloud model
Coordinate points；Here, the basic principle of point cloud data ordering is: reaching the point of setting value in three-dimensional space distance, after ordering
Distance is also relatively close, and this makes it possible to the features for being maximally maintained original point cloud not to be destroyed, and meet image domains
Positional relationship between consecutive points is based on this basic principle, designs following three kinds of different ordering methods:
Mass center ranking method: the distance at strong point to object mass center, by closely realizing the ordering of point cloud data, this method to remote sequence
The advantages of being ordered into result with point cloud input sequence and model translation, scaling, rotation it is unrelated, but on the other hand there is also
Following problems: opposite mass center is symmetrically put originally spatially each other without neighbouring relations, they but may phase each other after ordering
It is adjacent；
One-dimensional ranking method: model is ajusted in advance to be scanned again, and then along some reference axis, is sorted according to coordinate value size and realized point
The ordering of cloud data, this method can either guarantee ordering result and point cloud input sequence and model translation, scaling, rotation
It is unrelated, problem adjacent after the ordering of symmetric position point in space is also avoided, and object usually to be scanned is in space
Also it can satisfy the positive precedence constraint put, only the point cloud data in this way after sequence only embodies the space of some reference axis
Information can not embody the spatial information of other dimensions；
Two-dimentional ranking method: based on one-dimensional ranking method, point cloud data is obtained to the model scanning ajusted in advance, obtains a cloud number
It is equally spaced to be divided into m slice along some reference axis for the point cloud data model according to model, when m chooses appropriate value
When, it can be seen that the value in the reference axis inside same slice is suitable, i.e., these points are generally aligned in the same plane inside, at this point, right
Each slice sorts again according to some reference axis in plane coordinates, completes the ordering of point cloud data, this method can
The advantages of guaranteeing one-dimensional ranking method, and can preferably embody the spatial information of different dimensions.
5. a kind of three-dimensional point cloud model classification method based on convolutional neural networks according to claim 1, feature exist
In: in step s 4, input ordering sequence the S=((x of unordered point cloud datai,yi,zi), i=1 ..., n), this step is intended to
Ordering sequence is reasonably placed on two dimensional image A=(ajk) in p × q, wherein p × q=n, A are two-dimensional matrix, corresponding life
At image；J and k is respectively the row and column of pixel；ajkThe pixel value arranged for jth row kth in two-dimensional matrix；P, q is respectively square
The line number and columns of battle array；N indicates to put number included in point cloud model, to meet corresponding cloud between the adjacent pixel of image inside
Data are near one another on spatial position, for this requirement, design following three kinds of different two dimensional image methods:
Row scanning method: imitating the movement of fluorescent screen electron beam, orderly point cloud data successively taken out from front to back, from left to right,
It is filled into two dimensional image line by line from top to bottom, this mode can guarantee laterally adjacent pixel in original point cloud data each other
Close, it is near one another in original point cloud data to be but unable to ensure longitudinally adjacent pixel, that is, does not have isotropism；
Chessboard method: in view of CNN extracts characteristics of image using the thought of local receptor field, if can be by the part of cloud and image
Part be mapped, then the local feature of point cloud data can be preferably extracted, it is therefore proposed the image conversion side of chessboard method
Method: orderly point cloud data is successively taken out from front to back, is sequentially filled each grid from left to right, from top to bottom, inside grid
Each pixel is filled still according to mode from left to right, from top to bottom, when grid value is 8 × 8, each grid is just corresponded to
One point cloud local region comprising 64 points, this method do not have isotropism equally；
Spiral method: orderly point cloud data is successively taken out from front to back, since picture centre pixel, by helical trajectory successively into
Row filling, this mode can be good at keeping isotropism, and can be good at keeping original in the position by pericenter
The distance relation of spatial point, but there is also the defects of itself: closer to edge, pixel will more disperse, some in space
Distance may become larger after close point filling.
6. a kind of three-dimensional point cloud model classification method based on convolutional neural networks according to claim 1, feature exist
In: in step s 5, construction is suitable for the convolutional neural networks of two-dimentional point cloud chart picture, since point cloud data has finiteness and dilute
Dredge property, the large-scale CNN more than the number of plies may cause over-fitting, therefore, will choose first CNN and Small-sized C NN one medium-sized into
Row preliminary experiment；
Point cloud model classification based on medium-sized CNN: point cloud data has the characteristics that Limited information and sparsity, and medium size network
Towards data size it is big, when the input of AlxeNet is having a size of 224 × 224, point cloud model that input size is 1024
Corresponding two dimensional image size is only 32 × 32, thus, before by the input of point cloud chart picture, image data is carried out first anti-
Convolution operation meets medium-sized CNN input size requirements, avoids over-fitting, while realizing the super-resolution reconstruction of point cloud chart picture, obtain
Take more space correlation information；
Point cloud model classification based on Small-sized C NN: LeNet is one and connects entirely comprising two layers of convolution, two layers of pond and three layers
Small-sized C NN network, is mainly used for handwriting recongnition, and the input of the network is the image of 32 × 32 sizes, just with including 1024
The two-dimentional point cloud chart of a point is matched as size, therefore, directly continues to use the experimental design of the point cloud model classification based on medium-sized CNN,
Remove deconvolution submodule, 32 × 32 two-dimentional point cloud chart picture input LeNet is completed into feature extraction and category of model；
CNN construction and classification towards two-dimentional point cloud chart picture: analysis point cloud model own characteristic is tested in conjunction with both the above and is tied
Fruit, one convolutional neural networks PCI2CNN towards the classification of two-dimentional point cloud chart picture of design, the mentality of designing of the network are as follows:
Comprising one group of totally 2 deconvolution, this group of deconvolution port number is 64, and core size is 2 × 2, step-length is respectively 2 × 2 and 1 ×
1, to operate the super-resolution reconstruction for realizing two-dimentional point cloud chart picture by deconvolution, obtain the relevant information between more points cloud；
Comprising 3 convolutional layers, and port number is respectively 64,128,256；Compared to AlxeNet include less network layer and
Parameter improves the stability of network training to avoid the complexity of network；It include more parameters compared to LeNet, to improve
The ability of network fitting training data；
Pondization operation is added after first and third convolutional layer, and pond layer port number is consistent with upper one layer of port number, core is big
Small is 3 × 3, and step-length is 2 × 2, to obtain information more abundant by overlap sampling.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810790133.1A CN109063753B (en) | 2018-07-18 | 2018-07-18 | Three-dimensional point cloud model classification method based on convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810790133.1A CN109063753B (en) | 2018-07-18 | 2018-07-18 | Three-dimensional point cloud model classification method based on convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109063753A true CN109063753A (en) | 2018-12-21 |
| CN109063753B CN109063753B (en) | 2021-09-14 |
Family
ID=64817263
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810790133.1A Active CN109063753B (en) | 2018-07-18 | 2018-07-18 | Three-dimensional point cloud model classification method based on convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109063753B (en) |
Cited By (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109840938A (en) * | 2018-12-30 | 2019-06-04 | 芜湖哈特机器人产业技术研究院有限公司 | One kind being used for Complicated Automobile point cloud model method for reconstructing |
| CN109919046A (en) * | 2019-02-19 | 2019-06-21 | 清华大学 | A kind of three-dimensional point cloud feature learning method and apparatus based on relationship characteristic |
| CN110069993A (en) * | 2019-03-19 | 2019-07-30 | 同济大学 | A kind of target vehicle detection method based on deep learning |
| CN110097047A (en) * | 2019-03-19 | 2019-08-06 | 同济大学 | A kind of vehicle checking method using single line laser radar based on deep learning |
| CN110109995A (en) * | 2019-05-14 | 2019-08-09 | 中国矿业大学 | A kind of multi-source heterogeneous data fusion method of fully-mechanized mining working based on deep learning |
| CN110738194A (en) * | 2019-11-05 | 2020-01-31 | 电子科技大学中山学院 | three-dimensional object identification method based on point cloud ordered coding |
| CN110942110A (en) * | 2019-12-31 | 2020-03-31 | 新奥数能科技有限公司 | Feature extraction method and device of three-dimensional model |
| CN111028151A (en) * | 2019-12-03 | 2020-04-17 | 西安科技大学 | Point cloud data splicing method based on graph residual error neural network fusion |
| CN111274909A (en) * | 2020-01-16 | 2020-06-12 | 重庆邮电大学 | Human body point cloud framework extraction method based on deep learning |
| CN111310811A (en) * | 2020-02-06 | 2020-06-19 | 东华理工大学 | Large-scene three-dimensional point cloud classification method based on multi-dimensional feature optimal combination |
| CN111444927A (en) * | 2020-03-26 | 2020-07-24 | 广州市炜城智能科技有限公司 | Method for identifying animal and plant pests by using laser technology in multidimensional manner |
| CN111460193A (en) * | 2020-02-28 | 2020-07-28 | 天津大学 | Three-dimensional model classification method based on multi-mode information fusion |
| CN111832582A (en) * | 2019-04-15 | 2020-10-27 | 中国矿业大学(北京) | Method for classifying and segmenting sparse point cloud by using point cloud density and rotation information |
| CN112070105A (en) * | 2020-07-14 | 2020-12-11 | 辽宁师范大学 | Non-rigid three-dimensional model classification method based on two-channel convolutional neural network learning |
| CN112115744A (en) * | 2019-06-20 | 2020-12-22 | 北京京东叁佰陆拾度电子商务有限公司 | Point cloud data processing method and device, computer storage medium and electronic equipment |
| CN112435329A (en) * | 2020-12-01 | 2021-03-02 | 山东鲁能软件技术有限公司 | Power transmission equipment programmed modeling method and device based on laser point cloud data |
| WO2021052283A1 (en) * | 2019-09-16 | 2021-03-25 | 华为技术有限公司 | Method for processing three-dimensional point cloud data and computing device |
| CN112689997A (en) * | 2020-06-10 | 2021-04-20 | 深圳市大疆创新科技有限公司 | Point cloud sorting method and device |
| CN112906829A (en) * | 2021-04-13 | 2021-06-04 | 成都四方伟业软件股份有限公司 | Digital recognition model construction method and device based on Mnist data set |
| CN112967296A (en) * | 2021-03-10 | 2021-06-15 | 重庆理工大学 | Point cloud dynamic region graph convolution method, classification method and segmentation method |
| CN113076958A (en) * | 2021-04-25 | 2021-07-06 | 华南理工大学 | Three-dimensional point cloud model classification method and system with rotation invariance |
| CN113361473A (en) * | 2021-06-30 | 2021-09-07 | 北京百度网讯科技有限公司 | Image processing method, model training method, device, apparatus, storage medium, and program |
| KR20210119539A (en) * | 2019-06-21 | 2021-10-05 | 텐센트 테크놀로지(센젠) 컴퍼니 리미티드 | Image identification method and apparatus, identification model training method and apparatus, and storage medium |
| CN115290650A (en) * | 2022-09-29 | 2022-11-04 | 南京航空航天大学 | Method and system for detecting hole characteristics of composite material wallboard based on point cloud |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102592538A (en) * | 2010-09-30 | 2012-07-18 | 卡西欧计算机株式会社 | Display drive device, display device, driving control method and electronic device |
| CN105630905A (en) * | 2015-12-14 | 2016-06-01 | 西安科技大学 | Scattered-point cloud data based hierarchical compression method and apparatus |
| CN107092859A (en) * | 2017-03-14 | 2017-08-25 | 佛山科学技术学院 | A kind of depth characteristic extracting method of threedimensional model |
| CN107578463A (en) * | 2017-09-28 | 2018-01-12 | 奇瑞汽车股份有限公司 | The method and apparatus that rasterizing processing is carried out to radar cloud data |
| CN108230329A (en) * | 2017-12-18 | 2018-06-29 | 孙颖 | Semantic segmentation method based on multiple dimensioned convolutional neural networks |
-
2018
- 2018-07-18 CN CN201810790133.1A patent/CN109063753B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102592538A (en) * | 2010-09-30 | 2012-07-18 | 卡西欧计算机株式会社 | Display drive device, display device, driving control method and electronic device |
| CN105630905A (en) * | 2015-12-14 | 2016-06-01 | 西安科技大学 | Scattered-point cloud data based hierarchical compression method and apparatus |
| CN107092859A (en) * | 2017-03-14 | 2017-08-25 | 佛山科学技术学院 | A kind of depth characteristic extracting method of threedimensional model |
| CN107578463A (en) * | 2017-09-28 | 2018-01-12 | 奇瑞汽车股份有限公司 | The method and apparatus that rasterizing processing is carried out to radar cloud data |
| CN108230329A (en) * | 2017-12-18 | 2018-06-29 | 孙颖 | Semantic segmentation method based on multiple dimensioned convolutional neural networks |
Non-Patent Citations (2)
| Title |
|---|
| ALEX KRIZHEVSKY ET AL: "ImageNet Classification with Deep Convolutional Neural Networks", 《COMMUNICATIONS OF THE ACM》 * |
| 解则晓等: "一种散乱分层点云的有序化精简方法", 《图学学报》 * |
Cited By (37)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109840938B (en) * | 2018-12-30 | 2022-12-23 | 芜湖哈特机器人产业技术研究院有限公司 | Point cloud model reconstruction method for complex automobile |
| CN109840938A (en) * | 2018-12-30 | 2019-06-04 | 芜湖哈特机器人产业技术研究院有限公司 | One kind being used for Complicated Automobile point cloud model method for reconstructing |
| CN109919046A (en) * | 2019-02-19 | 2019-06-21 | 清华大学 | A kind of three-dimensional point cloud feature learning method and apparatus based on relationship characteristic |
| CN110097047A (en) * | 2019-03-19 | 2019-08-06 | 同济大学 | A kind of vehicle checking method using single line laser radar based on deep learning |
| CN110097047B (en) * | 2019-03-19 | 2021-10-08 | 同济大学 | Vehicle detection method based on deep learning and adopting single line laser radar |
| CN110069993A (en) * | 2019-03-19 | 2019-07-30 | 同济大学 | A kind of target vehicle detection method based on deep learning |
| CN111832582A (en) * | 2019-04-15 | 2020-10-27 | 中国矿业大学(北京) | Method for classifying and segmenting sparse point cloud by using point cloud density and rotation information |
| CN111832582B (en) * | 2019-04-15 | 2023-07-21 | 中国矿业大学(北京) | Method for classifying and segmenting sparse point cloud by utilizing point cloud density and rotation information |
| CN110109995A (en) * | 2019-05-14 | 2019-08-09 | 中国矿业大学 | A kind of multi-source heterogeneous data fusion method of fully-mechanized mining working based on deep learning |
| CN112115744A (en) * | 2019-06-20 | 2020-12-22 | 北京京东叁佰陆拾度电子商务有限公司 | Point cloud data processing method and device, computer storage medium and electronic equipment |
| CN112115744B (en) * | 2019-06-20 | 2024-05-21 | 北京京东叁佰陆拾度电子商务有限公司 | Point cloud data processing method and device, computer storage medium and electronic equipment |
| KR20210119539A (en) * | 2019-06-21 | 2021-10-05 | 텐센트 테크놀로지(센젠) 컴퍼니 리미티드 | Image identification method and apparatus, identification model training method and apparatus, and storage medium |
| KR102645533B1 (en) * | 2019-06-21 | 2024-03-07 | 텐센트 테크놀로지(센젠) 컴퍼니 리미티드 | Image identification method and device, identification model training method and device, and storage medium |
| WO2021052283A1 (en) * | 2019-09-16 | 2021-03-25 | 华为技术有限公司 | Method for processing three-dimensional point cloud data and computing device |
| CN110738194A (en) * | 2019-11-05 | 2020-01-31 | 电子科技大学中山学院 | three-dimensional object identification method based on point cloud ordered coding |
| CN110738194B (en) * | 2019-11-05 | 2023-04-07 | 电子科技大学中山学院 | Three-dimensional object identification method based on point cloud ordered coding |
| CN111028151A (en) * | 2019-12-03 | 2020-04-17 | 西安科技大学 | Point cloud data splicing method based on graph residual error neural network fusion |
| CN110942110A (en) * | 2019-12-31 | 2020-03-31 | 新奥数能科技有限公司 | Feature extraction method and device of three-dimensional model |
| CN111274909A (en) * | 2020-01-16 | 2020-06-12 | 重庆邮电大学 | Human body point cloud framework extraction method based on deep learning |
| CN111274909B (en) * | 2020-01-16 | 2022-05-20 | 重庆邮电大学 | Human body point cloud framework extraction method based on deep learning |
| CN111310811A (en) * | 2020-02-06 | 2020-06-19 | 东华理工大学 | Large-scene three-dimensional point cloud classification method based on multi-dimensional feature optimal combination |
| CN111460193A (en) * | 2020-02-28 | 2020-07-28 | 天津大学 | Three-dimensional model classification method based on multi-mode information fusion |
| CN111444927A (en) * | 2020-03-26 | 2020-07-24 | 广州市炜城智能科技有限公司 | Method for identifying animal and plant pests by using laser technology in multidimensional manner |
| CN111444927B (en) * | 2020-03-26 | 2023-05-26 | 广州市炜城智能科技有限公司 | Method for identifying animal and plant pests by using multi-dimensional morphology of laser technology |
| CN112689997A (en) * | 2020-06-10 | 2021-04-20 | 深圳市大疆创新科技有限公司 | Point cloud sorting method and device |
| CN112070105A (en) * | 2020-07-14 | 2020-12-11 | 辽宁师范大学 | Non-rigid three-dimensional model classification method based on two-channel convolutional neural network learning |
| CN112070105B (en) * | 2020-07-14 | 2023-07-11 | 辽宁师范大学 | Non-rigid three-dimensional model classification method based on double-channel convolutional neural network learning |
| CN112435329A (en) * | 2020-12-01 | 2021-03-02 | 山东鲁能软件技术有限公司 | Power transmission equipment programmed modeling method and device based on laser point cloud data |
| CN112967296B (en) * | 2021-03-10 | 2022-11-15 | 重庆理工大学 | Point cloud dynamic region graph convolution method, classification method and segmentation method |
| CN112967296A (en) * | 2021-03-10 | 2021-06-15 | 重庆理工大学 | Point cloud dynamic region graph convolution method, classification method and segmentation method |
| CN112906829A (en) * | 2021-04-13 | 2021-06-04 | 成都四方伟业软件股份有限公司 | Digital recognition model construction method and device based on Mnist data set |
| CN113076958B (en) * | 2021-04-25 | 2023-07-18 | 华南理工大学 | Three-dimensional point cloud model classification method and system with rotation invariance |
| CN113076958A (en) * | 2021-04-25 | 2021-07-06 | 华南理工大学 | Three-dimensional point cloud model classification method and system with rotation invariance |
| CN113361473A (en) * | 2021-06-30 | 2021-09-07 | 北京百度网讯科技有限公司 | Image processing method, model training method, device, apparatus, storage medium, and program |
| CN113361473B (en) * | 2021-06-30 | 2023-12-08 | 北京百度网讯科技有限公司 | Image processing method, model training method, image processing device, model training apparatus, storage medium, and program |
| CN115290650B (en) * | 2022-09-29 | 2023-01-17 | 南京航空航天大学 | Method and system for detecting hole characteristics of composite material wallboard based on point cloud |
| CN115290650A (en) * | 2022-09-29 | 2022-11-04 | 南京航空航天大学 | Method and system for detecting hole characteristics of composite material wallboard based on point cloud |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109063753B (en) | 2021-09-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109063753A (en) | A kind of three-dimensional point cloud model classification method based on convolutional neural networks | |
| CN106778821B (en) | Classification of Polarimetric SAR Image method based on SLIC and improved CNN | |
| CN107563422B (en) | A kind of polarization SAR classification method based on semi-supervised convolutional neural networks | |
| CN105426919B (en) | The image classification method of non-supervisory feature learning is instructed based on conspicuousness | |
| CN106650830A (en) | Deep model and shallow model decision fusion-based pulmonary nodule CT image automatic classification method | |
| CN110458249A (en) | A kind of lesion categorizing system based on deep learning Yu probability image group | |
| CN108596248A (en) | A kind of classification of remote-sensing images model based on improvement depth convolutional neural networks | |
| CN109410321A (en) | Three-dimensional rebuilding method based on convolutional neural networks | |
| CN106203487A (en) | A kind of image classification method based on Multiple Kernel Learning Multiple Classifier Fusion and device | |
| CN110298831A (en) | A kind of magic magiscan and its method based on piecemeal deep learning | |
| CN110276402A (en) | A kind of salt body recognition methods based on the enhancing of deep learning semanteme boundary | |
| CN107292336A (en) | A kind of Classification of Polarimetric SAR Image method based on DCGAN | |
| CN115546032B (en) | Single-frame image super-resolution method based on feature fusion and attention mechanism | |
| CN110222767A (en) | Three-dimensional point cloud classification method based on nested neural and grating map | |
| CN110263858A (en) | A kind of bolt image composition method, device and relevant device | |
| CN106981067A (en) | A kind of Texture Segmentation Methods based on full convolutional network | |
| CN115331104A (en) | Crop planting information extraction method based on convolutional neural network | |
| CN111738052B (en) | Multi-feature fusion hyperspectral remote sensing ground object classification method based on deep learning | |
| Li et al. | Few-shot hyperspectral image classification with self-supervised learning | |
| CN111833432B (en) | Three-dimensional reconstruction method based on core two-dimensional gray scale image | |
| CN114219049A (en) | Fine-grained pencil and stone image classification method and device based on hierarchy constraint | |
| CN110188811A (en) | Underwater target detection method based on normed Gradient Features and convolutional neural networks | |
| Ma et al. | Feature selection module for CNN based object detector | |
| Sun et al. | ESinGAN: Enhanced single-image GAN using pixel attention mechanism for image super-resolution | |
| CN108446723A (en) | A kind of multiple dimensioned empty spectrum synergetic classification method of high spectrum image |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |