CN108876831A - A kind of building three-dimensional point cloud method for registering based on deep learning - Google Patents
A kind of building three-dimensional point cloud method for registering based on deep learning Download PDFInfo
- Publication number
- CN108876831A CN108876831A CN201810584888.6A CN201810584888A CN108876831A CN 108876831 A CN108876831 A CN 108876831A CN 201810584888 A CN201810584888 A CN 201810584888A CN 108876831 A CN108876831 A CN 108876831A
- Authority
- CN
- China
- Prior art keywords
- point
- point cloud
- cloud
- key
- model
- 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.)
- Withdrawn
Links
Classifications
-
- 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
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
- G06T7/33—Determination of transform parameters for the alignment of images, i.e. image registration using feature-based methods
-
- 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
- G06T2207/10004—Still image; Photographic image
- G06T2207/10012—Stereo images
Abstract
The invention discloses a kind of building three-dimensional point cloud method for registering based on deep learning, includes the following steps:The acquisition of S1, training data;The building of S2, deep learning model;S3, model training and tuning;S4, data to be tested pretreatment;S5, test point cloud key area;S6, test point cloud key point;S7, judge crucial point correspondence;S8, transformational relation and registration are calculated.Deep learning is applied to two committed steps of point cloud registering by the present invention:It finds key point and determines matching relationship.The present invention makes full use of deep learning to the characterization ability of point cloud data to find key area, positioning key point step by step, detection speed is greatly accelerated relative to traditional traversal mode, and learns point cloud feature with depth network model and replace manual feature, so that algorithm is more robust and efficient.
Description
Technical field
The present invention relates to city three-dimensional reconstruction and unmanned field more particularly to a kind of buildings based on deep learning
Object three-dimensional point cloud method for registering.
Background technique
In industrial production, life and scientific research, the demand to three-dimensional reconstruction is all widely present, such as AR/VR, 3D
Printing, industrial automation, the protection of historical relic legacy, computer technology are in application of medical domain etc..Nowadays three-dimensional is swept
The precision for retouching instrument is higher and higher, obtains fine city large-scale data and also becomes easy, some developed countries start already
The construction of smart city, and the three-dimensional reconstruction in city is the important component of smart city, for cloud data registration side
The research of method becomes key in three-dimensional reconstruction field.
In traditional method for registering based on point, the rule that key point must manually be set according to certain is found, successively
It traverses all points and calculates response, key point is finally selected according to threshold value.This method can not be based on neighborhood in view of point
All characteristics, cause feature extraction to be not thorough, next takes the mode of traversal computationally intensive, especially for outdoor architecture object
In the processing of larger point cloud.Deep learning is in growing continuously and fast at this stage, so that the research of numerous areas obtains
Great breakthrough and plentiful and substantial scientific achievement, especially obtain considerable progress in the relevant research field of picture, still
Research and application in terms of three-dimensional data is then at the early-stage.It is essentially all to take especially in three-dimensional point cloud application aspect
To low dimension projective, division mode by point cloud pretreatment at certain specification, will cause information in this way and largely lose.
Since laser radar moves towards industry, scientific research field from military project, cost price also gradually declines, so that obtaining a large amount of
Point cloud data be possibly realized, be also more convenient for the training of deep learning model.It is how deep using the training of a large amount of point cloud datas
Learning model is spent, accurately efficiently extracts key point and description point cloud feature as urgent using the sensing capability of model
Business.
Summary of the invention
The purpose of the present invention is to provide a kind of building three-dimensional point cloud method for registering based on deep learning, can have
Effect promotes registration efficiency and precision.
To achieve the above object, the present invention uses following technical scheme:
A kind of building three-dimensional point cloud method for registering based on deep learning comprising following steps:
The acquisition of S1, training sample obtain the sample data for training pattern;
S2, deep learning model is built, so that model is suitable for the processing of point cloud data;
S3, hyper parameter is adjusted according to model training, test feedback information;
S4, pretreatment point cloud, the input by point cloud data piecemeal to be detected, as model;
S5, the input point cloud mass for model, whether detection wherein contains key area, by the point cloud comprising key area
Block retains;
S6, the key point put in cloud mass comprising key area is detected;
S7, the key point for being respectively at difference cloud is inputted into key point Matching Model in couples, judge they whether
Match;
S8, the amount of taking fully correspondence key point calculate transformational relation, it is best from obtained transition matrix to select registration effect
One be used as final result.
The step S1 specifically include it is following step by step:
The point of S11, interception in positions such as door frame angle, window frame angle and metope angles is as key point;
S12, take out using key point as the centre of sphere, radius r sphere in include point, then with these point be respectively the centre of sphere, cut
Positive sample of the radius for the point cloud mass of the sphere of r as key area detection model is taken, so that positive sample all includes key
Point;Negative sample of the sphere of same size as key area detection model is intercepted in the region that cloud does not include key point;
S13, bulbous region point cloud mass is intercepted as the positive sample of critical point detection model, not to close by the centre of sphere of key point
Key point is negative sample of the bulbous region point cloud mass of the centre of sphere as critical point detection model, includes in negative sample or does not include crucial
Point;
S14, the positive sample that key point Matching Model is extracted in two point cloud datas P, Q being registrated:It is taken in cloud P
Point, then the nearest point of the Search Length point is as corresponding points in cloud Q, if closest approach distance is greater than the threshold value of setting,
Without corresponding points;Negative sample is extracted in two point cloud datas P, Q being registrated:It takes a little in cloud P, is then taken in cloud Q
It is greater than the point of given threshold as noncorresponding points away from the point;
S15, using corresponding points to the pairs of bulbous region point cloud mass intercepted for the centre of sphere as key point Matching Model just
Sample, using noncorresponding points to the pairs of bulbous region point cloud mass intercepted for the centre of sphere as the negative sample of key point Matching Model.
The step S2 specifically include it is following step by step:
S21, point cloud classifications model is built based on PointNet network structure, so that model is directly acted in the form of convolution
In the coordinate data of three-dimensional point, the feature extracted finally is mapped in the form of fully-connected network;
S22, twin network model is built based on PointNet network structure, so that model branch is respectively acting on difference
Cloud judges whether some cloud masses correspond to by connecting the output after mapping entirely.
In the step S3, according to model training concrete condition, the hyper parameter of input data is adjusted, hyper parameter includes learning
Practise step-length, learning rate, batch size.
The step S4 specifically include it is following step by step:
S41, farthest point sampling is carried out on to be processed cloud;
S42, using sampled point as the centre of sphere, intercept out the point cloud mass in certain radius sphere, sentence as key area
The input of disconnected model.
The step S6 specifically include it is following step by step:
S61, it takes a little in the point cloud mass comprising key area, and is cut respectively using these points as the centre of sphere with small radii
Take sphere;
S62, by sphere obtained in step S61 as the input of key point discrimination model, reservation is judged as crucial
The point of point.
In the step S8, taking RANSAC strategy, the corresponding key point of the amount of taking fully utilizes SVD to calculate transformational relation every time,
Take RANSAC strategy to obtain a large amount of transformational relations, find matching degree best one, registration effect with LCP value be judgement mark
Standard, formula are expressed as follows:
Wherein piRepresent the point in point cloud P, qiThe point in a cloud Q is respectively represented, R and T respectively indicate rotation and translation square
Battle array, ε value are 0.05m;Formula meaning is maximum public point set, i.e., after point cloud P, Q registration, point q is taken in Qi, find by turning
In point cloud P after changing with qiApart it is less than the point of ε, if it exists such point, as common point, then LCP value is increase accordingly, and is taken
LCP value is highest to be used as final result.
After adopting the above technical scheme, compared with the background technology, the present invention, having the following advantages that:
1, the application by the powerful learning ability of deep neural network and PointNet network structure on unordered cloud,
Training pattern can identify the search range for greatly reducing comprising crucial region and finding key point.
2, point cloud local region is perceived using depth network model, extracts provincial characteristics, rather than engineer's rule mentions
Feature is taken, the structural information of this sampling point cloud is able to be perceptible enough and extract, so that feature is more accurate to the description in region.
3, so that it is incuded matching relationship with the twin network model of mass data training, model is allowed to learn corresponding rule from sample
Then, to eliminate the judgment criteria that feature space distance is manually set, so that matching is more objective credible.
Detailed description of the invention
Fig. 1 is workflow schematic diagram of the invention.
Fig. 2 is key area detection model positive sample choosing method.
Fig. 3 is to constitute this method to be related to the main structure of model.
Fig. 4 is the twin network model for judging crucial point correspondence.
Fig. 5 is the distribution example of farthest point sampled result.
Fig. 6 is the displaying of critical point detection result.
Fig. 7 is two o'clock cloud registration result example.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.
Embodiment
Referring to Fig. 1, the invention discloses a kind of building three-dimensional point cloud method for registering based on deep learning comprising
Following steps:
The acquisition of S1, training sample obtain the sample data for training pattern;
The point cloud data that this method is used only is needed comprising three-dimensional coordinate without other information, since model is to be used for
Regional area geometry is experienced, so the training data for same model needs to normalize to the sphere of same size
It is interior, and the centre of sphere is needed to be aligned, the gaussian noise that mean value is zero is added when training, is defined as follows:
And the data enhancement methods such as Random-Rotation for taking vertical direction, it is defined as follows:
The point of S11, interception in positions such as door frame angle, window frame angle and metope angles is as key point, due to laser thunder
Point up to scanning is not necessarily at angular vertex, so (the d in a certain range around the corner structure<Point 0.01m) is to close
Key point;
S12, take out using key point as the centre of sphere, radius r (r=0.5m) sphere (dark circle circuit covering as shown in Figure 2
Region) in include point, then with these point be respectively the centre of sphere, interception radius be r sphere point cloud mass as key area
The positive sample (region of light color circle covering as shown in Figure 2) of domain detection model, such positive sample all includes key point;In a cloud
Region not comprising key point intercepts negative sample of the sphere of same size as key area detection model, in such negative sample
Not comprising key point;
S13, bulbous region point cloud mass is intercepted as the positive sample of critical point detection model, not to close by the centre of sphere of key point
Key point is negative sample of the bulbous region point cloud mass of the centre of sphere as critical point detection model, includes in negative sample or does not include crucial
Point.Since the situation about being located near key area when carrying out the judgement of key point, encountered is more, so needing closing
Key range nearby more does negative sample;
S14, the sample (point maked sample that key point Matching Model is extracted in two point cloud datas P, Q being registrated
Key point is known in cloud).Positive sample:Point (key point and non-key point) is taken in cloud P, then in cloud Q
The nearest point of the Search Length point (key point and non-key point) is as corresponding points (if closest approach distance is greater than the threshold of setting
Value, then without corresponding points);Negative sample:Point (key point and non-key point) is taken in cloud P, is then taken in cloud Q away from this
Point is greater than the point of given threshold as noncorresponding points.In a practical situation, judge the corresponding relationship in cloud P, Q between key point
The case where it is more, so, in positive sample,Should be in the majority, in negative sample,It should be in the majority.
S15, with corresponding point to bulbous region point cloud mass is intercepted in couples for the centre of sphere, just as key point Matching Model
Sample, in positive sample, key point is relatively high to accounting for, and non-key point is lower to accounting;With noncorresponding points to intercepting for the centre of sphere
Bulbous region point cloud mass is to the negative sample as key point Matching Model.
S2, deep learning model is built, so that model is suitable for the processing of point cloud data;
The main part of model is shown in that Fig. 3, model act on three-dimensional coordinate a little in the form of convolution, but due to cloud
Randomness is put and is not in contact between the neighbouring relations and point neighbouring relations notationally of three-dimensional space, so each is rolled up
Product core acts only on a single point.Feature extraction is finally the solution point cloud representation sequence tape in the form of maxpooling in model
The influence come.Formula is schematically as follows:
f({x1,…,xn})≈g(h(x1),…,h(xn))
Wherein Xk(k=1 ..., n) is three-dimensional point, and function g is symmetric function, and function h is multi-layer perception (MLP), in this method
In be convolution form, symmetric function by max () realization.
S21, point cloud classifications model is built based on PointNet network structure, so that model is directly acted in the form of convolution
In the coordinate data of three-dimensional point, the feature that finally mapping is extracted in the form of fully-connected network finally obtains classification results.The model
Applied to detection key area and detection key point;
S22, twin network model is built based on PointNet network structure, as shown in figure 4, making model branch respectively
For difference cloud mass, by connecting the mapping output judgement point matched probability of cloud mass entirely.The model is applied to judge key point
Between matching relationship.
S3, hyper parameter is adjusted according to model training, test feedback information;
According to model training concrete condition, the hyper parameters such as the Learning Step, learning rate, batch size of input data are adjusted.
S4, pretreatment point cloud, the input by point cloud data piecemeal to be detected, as model;
S41, farthest point sampling (Farthest Point Sampling) is carried out on to be processed cloud, sees Fig. 5, in figure
Biggish black color dots and lesser light color point are sampled point.
S42, using sampled point as the centre of sphere, the point cloud mass in the sphere that radius is r=0.5m is intercepted out, as key
The input of region detection model.
Whether S5, the input point cloud mass for model, detection wherein contain key area (region comprising key point), will
Point cloud mass comprising key area retains.Sphere neighborhood representated by larger black color dots is that key area detection model is sentenced in Fig. 5
Break as the region comprising key point;
S6, the key point put in cloud mass comprising key area is detected;
In the region comprising key point detected in S5, traverses each point or suitable point is taken to judge whether at random
Key point.
S61, it is taken a little in the point cloud mass comprising key area, and respectively using these points as the centre of sphere, with small radii (r
=0.1m) interception sphere;
S62, by sphere obtained in S61 as the input of critical point detection model, reservation is judged as key point
Point.Lesser Grey Point is the key point that critical point detection model is found out in Fig. 6, and right side is four effects of partial enlargement in figure
Fruit is schemed, and smaller Grey Point is key point in effect picture.
S7, the crucial vertex neighborhood for being respectively at difference cloud is inputted into key point Matching Model in couples, judges that they are
No matching.Key point of the fetch bit in two point cloud datas partners respectively, and intercepts sphere area by the centre of sphere of key point
Domain point cloud mass inputs key point Matching Model, records crucial point correspondence.
S8, the correspondence key point for taking each amount of taking fully of RANSAC strategy calculate transformational relation using SVD, obtain a large amount of
After transformational relation, take matching degree best one.Point cloud P and Q registration formula is expressed as follows:
qi=Rpi+T
Wherein piAnd qiThe point in cloud P, Q is respectively represented, R and T respectively indicate rotation and translation matrix.
As shown in fig. 7, two point cloud data (black and grey) registration effect a front surface and a side surfaces are shown.Registration effect with
LCP (Largest Common Points) value is judgment criteria, and formula is expressed as follows:
Wherein R, T, pi、qiMeaning is same as described above, and ε value is 0.05m.Formula meaning is maximum public point set, i.e. point cloud
P, after Q registration, point q is taken in Qi, find by conversion after point cloud P in qiApart it is less than the point of ε, if it exists such point,
As common point, then LCP value is increase accordingly.
LCP value is respectively from left to right in Fig. 7:80.3%, 75.5%, 72.1%, take the highest left figure of LCP value as most
Terminate fruit.
After above-mentioned steps, the registration to building three-dimensional point cloud based on deep learning model can be realized.
The foregoing is only a preferred embodiment of the present invention, but scope of protection of the present invention is not limited thereto,
In the technical scope disclosed by the present invention, any changes or substitutions that can be easily thought of by anyone skilled in the art,
It should be covered by the protection scope of the present invention.Therefore, protection scope of the present invention should be with scope of protection of the claims
Subject to.
Claims (7)
1. a kind of building three-dimensional point cloud method for registering based on deep learning, which is characterized in that include the following steps:
The acquisition of S1, training sample obtain the sample data for training pattern;
S2, deep learning model is built, so that model is suitable for the processing of point cloud data;
S3, hyper parameter is adjusted according to model training, test feedback information;
S4, pretreatment point cloud, the input by point cloud data piecemeal to be detected, as model;
Whether S5, the input point cloud mass for model, detection wherein contain key area, and the point cloud mass comprising key area is protected
It stays;
S6, the key point put in cloud mass comprising key area is detected;
S7, the key point for being respectively at difference cloud is inputted into key point Matching Model in couples, judges whether they match;
S8, the amount of taking fully correspondence key point calculate transformational relation, selected from obtained transition matrix registration effect it is best one
It is a to be used as final result.
2. a kind of building three-dimensional point cloud method for registering based on deep learning as described in claim 1, which is characterized in that institute
State step S1 specifically include it is following step by step:
The point of S11, interception in positions such as door frame angle, window frame angle and metope angles is as key point;
S12, take out using key point as the centre of sphere, radius r sphere in include point, then with these points be respectively the centre of sphere, interception half
Diameter is positive sample of the point cloud mass of the sphere of r as key area detection model, so that positive sample all includes key point;?
The region that point cloud does not include key point intercepts negative sample of the sphere of same size as key area detection model;
S13, bulbous region point cloud mass is intercepted as the positive sample of critical point detection model, not with key point by the centre of sphere of key point
Include for negative sample of the bulbous region point cloud mass as critical point detection model of the centre of sphere, in negative sample or not include key point equal
It can;
S14, the positive sample that key point Matching Model is extracted in two point cloud datas P, Q being registrated:It is taken a little in cloud P,
Then the nearest point of the Search Length point is as corresponding points in cloud Q, if closest approach distance is greater than the threshold value of setting, without right
Ying Dian;Negative sample is extracted in two point cloud datas P, Q being registrated:It takes a little in cloud P, then takes in cloud Q away from this
Point is greater than the point of given threshold as noncorresponding points;
S15, using corresponding points to the pairs of bulbous region point cloud mass intercepted for the centre of sphere as the positive sample of key point Matching Model,
Using noncorresponding points to the pairs of bulbous region point cloud mass intercepted for the centre of sphere as the negative sample of key point Matching Model.
3. a kind of building three-dimensional point cloud method for registering based on deep learning as described in claim 1, which is characterized in that institute
State step S2 specifically include it is following step by step:
S21, point cloud classifications model is built based on PointNet network structure, so that model directly acts on three in the form of convolution
Tie up the coordinate data of point, the feature that finally mapping is extracted in the form of fully-connected network;
S22, twin network model is built based on PointNet network structure, so that model branch is respectively acting on difference cloud,
Judge whether some cloud masses correspond to by connecting the output after mapping entirely.
4. a kind of building three-dimensional point cloud method for registering based on deep learning as described in claim 1, which is characterized in that
In the step S3, according to model training concrete condition, the hyper parameter of input data is adjusted, hyper parameter includes Learning Step, learns
Habit rate, batch size.
5. a kind of building three-dimensional point cloud method for registering based on deep learning as claimed in claim 2, which is characterized in that institute
State step S4 specifically include it is following step by step:
S41, farthest point sampling is carried out on to be processed cloud;
S42, using sampled point as the centre of sphere, intercept out the point cloud mass in certain radius sphere, judge mould as key area
The input of type.
6. a kind of building three-dimensional point cloud method for registering based on deep learning as claimed in claim 2, which is characterized in that institute
State step S6 specifically include it is following step by step:
S61, it is taken a little in the point cloud mass comprising key area, and respectively using these points as the centre of sphere, ball is intercepted with small radii
Body region;
S62, by sphere obtained in step S61 as the input of key point discrimination model, reservation is judged as key point
Point.
7. a kind of building three-dimensional point cloud method for registering based on deep learning as described in claim 1, it is characterised in that:?
In the step S8, taking RANSAC strategy, the corresponding key point of the amount of taking fully takes RANSAC using SVD calculating transformational relation every time
Strategy obtains a large amount of transformational relations, find matching degree best one, registration effect using LCP value as judgment criteria, formula indicate
It is as follows:
Wherein piRepresent the point in point cloud P, qiThe point in a cloud Q is respectively represented, R and T respectively indicate rotation and translation matrix, and ε takes
Value is 0.05m;Formula meaning is maximum public point set, i.e., after point cloud P, Q registration, point q is taken in Qi, find after conversion
Point cloud P in qiApart it is less than the point of ε, if it exists such point, as common point, then LCP value is increase accordingly, and takes LCP value most
High is used as final result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810584888.6A CN108876831A (en) | 2018-06-08 | 2018-06-08 | A kind of building three-dimensional point cloud method for registering based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810584888.6A CN108876831A (en) | 2018-06-08 | 2018-06-08 | A kind of building three-dimensional point cloud method for registering based on deep learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108876831A true CN108876831A (en) | 2018-11-23 |
Family
ID=64337567
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810584888.6A Withdrawn CN108876831A (en) | 2018-06-08 | 2018-06-08 | A kind of building three-dimensional point cloud method for registering based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108876831A (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110210431A (en) * | 2019-06-06 | 2019-09-06 | 上海黑塞智能科技有限公司 | A kind of point cloud classifications method based on cloud semantic tagger and optimization |
CN111027140A (en) * | 2019-12-11 | 2020-04-17 | 南京航空航天大学 | Airplane standard part model rapid reconstruction method based on multi-view point cloud data |
CN111429563A (en) * | 2020-03-10 | 2020-07-17 | 山东大学 | Pipeline three-dimensional reconstruction method, system, medium and equipment based on deep learning |
CN111612891A (en) * | 2020-05-22 | 2020-09-01 | 北京京东乾石科技有限公司 | Model generation method, point cloud data processing device, point cloud data processing equipment and medium |
CN111882593A (en) * | 2020-07-23 | 2020-11-03 | 首都师范大学 | Point cloud registration model and method combining attention mechanism and three-dimensional graph convolution network |
CN112150523A (en) * | 2020-09-24 | 2020-12-29 | 中北大学 | Three-dimensional point cloud registration method with low overlapping rate |
JP2021515724A (en) * | 2019-01-30 | 2021-06-24 | バイドゥ ドットコム タイムス テクノロジー (ベイジン) カンパニー リミテッド | LIDAR positioning to infer solutions using 3DCNN network in self-driving cars |
CN113139996A (en) * | 2021-05-06 | 2021-07-20 | 南京大学 | Point cloud registration method and system based on three-dimensional point cloud geometric feature learning |
CN113807366A (en) * | 2021-09-16 | 2021-12-17 | 电子科技大学 | Point cloud key point extraction method based on deep learning |
-
2018
- 2018-06-08 CN CN201810584888.6A patent/CN108876831A/en not_active Withdrawn
Cited By (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP7060625B2 (en) | 2019-01-30 | 2022-04-26 | バイドゥドットコム タイムズ テクノロジー (ベイジン) カンパニー リミテッド | LIDAR positioning to infer solutions using 3DCNN network in self-driving cars |
JP2021515724A (en) * | 2019-01-30 | 2021-06-24 | バイドゥ ドットコム タイムス テクノロジー (ベイジン) カンパニー リミテッド | LIDAR positioning to infer solutions using 3DCNN network in self-driving cars |
US11531110B2 (en) | 2019-01-30 | 2022-12-20 | Baidu Usa Llc | LiDAR localization using 3D CNN network for solution inference in autonomous driving vehicles |
CN110210431A (en) * | 2019-06-06 | 2019-09-06 | 上海黑塞智能科技有限公司 | A kind of point cloud classifications method based on cloud semantic tagger and optimization |
CN111027140A (en) * | 2019-12-11 | 2020-04-17 | 南京航空航天大学 | Airplane standard part model rapid reconstruction method based on multi-view point cloud data |
CN111429563A (en) * | 2020-03-10 | 2020-07-17 | 山东大学 | Pipeline three-dimensional reconstruction method, system, medium and equipment based on deep learning |
CN111429563B (en) * | 2020-03-10 | 2021-08-13 | 山东大学 | Pipeline three-dimensional reconstruction method, system, medium and equipment based on deep learning |
CN111612891A (en) * | 2020-05-22 | 2020-09-01 | 北京京东乾石科技有限公司 | Model generation method, point cloud data processing device, point cloud data processing equipment and medium |
CN111612891B (en) * | 2020-05-22 | 2023-08-08 | 北京京东乾石科技有限公司 | Model generation method, point cloud data processing method, device, equipment and medium |
CN111882593A (en) * | 2020-07-23 | 2020-11-03 | 首都师范大学 | Point cloud registration model and method combining attention mechanism and three-dimensional graph convolution network |
CN111882593B (en) * | 2020-07-23 | 2022-06-17 | 首都师范大学 | Point cloud registration model and method combining attention mechanism and three-dimensional graph convolution network |
CN112150523B (en) * | 2020-09-24 | 2022-04-12 | 中北大学 | Three-dimensional point cloud registration method with low overlapping rate |
CN112150523A (en) * | 2020-09-24 | 2020-12-29 | 中北大学 | Three-dimensional point cloud registration method with low overlapping rate |
CN113139996A (en) * | 2021-05-06 | 2021-07-20 | 南京大学 | Point cloud registration method and system based on three-dimensional point cloud geometric feature learning |
CN113139996B (en) * | 2021-05-06 | 2024-02-06 | 南京大学 | Point cloud registration method and system based on three-dimensional point cloud geometric feature learning |
CN113807366A (en) * | 2021-09-16 | 2021-12-17 | 电子科技大学 | Point cloud key point extraction method based on deep learning |
CN113807366B (en) * | 2021-09-16 | 2023-08-08 | 电子科技大学 | Point cloud key point extraction method based on deep learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108876831A (en) | A kind of building three-dimensional point cloud method for registering based on deep learning | |
CN108765475A (en) | A kind of building three-dimensional point cloud method for registering based on deep learning | |
CN110378196B (en) | Road visual detection method combining laser point cloud data | |
CN106127204B (en) | A kind of multi-direction meter reading Region detection algorithms of full convolutional neural networks | |
CN103268729B (en) | Based on mobile robot's tandem type map creating method of composite character | |
CN108364272A (en) | A kind of high-performance Infrared-Visible fusion detection method | |
CN101770581B (en) | Semi-automatic detecting method for road centerline in high-resolution city remote sensing image | |
WO2017020528A1 (en) | Lane line recognition modeling method, apparatus, storage medium, and device, recognition method and apparatus, storage medium, and device | |
CN108288088A (en) | A kind of scene text detection method based on end-to-end full convolutional neural networks | |
CN106780592A (en) | Kinect depth reconstruction algorithms based on camera motion and image light and shade | |
CN105005769A (en) | Deep information based sign language recognition method | |
CN105160686B (en) | A kind of low latitude various visual angles Remote Sensing Images Matching Method based on improvement SIFT operators | |
CN103489193B (en) | OO high-resolution remote sensing image change detecting method based on convergence strategy | |
Xue et al. | Dual network structure with interweaved global-local feature hierarchy for transformer-based object detection in remote sensing image | |
CN106447662A (en) | Combined distance based FCM image segmentation algorithm | |
CN106022342A (en) | Image feature extraction method based on KAZE algorithm | |
Cheng et al. | Generating high-resolution climate prediction through generative adversarial network | |
CN109829459A (en) | Based on the vision positioning method for improving RANSAC | |
CN107330436B (en) | Scale criterion-based panoramic image SIFT optimization method | |
CN115082778B (en) | Multi-branch learning-based homestead identification method and system | |
CN106558065A (en) | The real-time vision tracking to target is realized based on color of image and texture analysiss | |
CN116128024A (en) | Multi-view contrast self-supervision attribute network abnormal point detection method | |
CN104732190A (en) | Synthetic aperture sonar target detection method based on orthogonal texture correlation analysis | |
CN109583461A (en) | A kind of template matching method based on edge feature | |
Wu et al. | Image inpainting algorithm based on adaptive template direction |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
WW01 | Invention patent application withdrawn after publication |
Application publication date: 20181123 |
|
WW01 | Invention patent application withdrawn after publication |