CN107369158A - The estimation of indoor scene layout and target area extracting method based on RGB D images - Google Patents
The estimation of indoor scene layout and target area extracting method based on RGB D images Download PDFInfo
- Publication number
- CN107369158A CN107369158A CN201710442910.9A CN201710442910A CN107369158A CN 107369158 A CN107369158 A CN 107369158A CN 201710442910 A CN201710442910 A CN 201710442910A CN 107369158 A CN107369158 A CN 107369158A
- Authority
- CN
- China
- Prior art keywords
- plane
- mrow
- segmentation
- rgb
- target area
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/13—Edge detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/50—Depth or shape recovery
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
-
- 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
- G06T2207/10024—Color image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10028—Range image; Depth image; 3D point clouds
-
- 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
- G06T2207/20—Special algorithmic details
- G06T2207/20024—Filtering details
Abstract
The present invention discloses a kind of indoor scene layout estimation based on RGB D images and target area extracting method, including：Scene layout estimates；Over-segmentation is done to pretreated depth map and RGB figures using the partitioning algorithm based on figure and constrained parameters minimal cut algorithm, obtains different size of regional ensemble；Over-segmentation level is grouped, and is carried out region merging technique using four kinds of different measuring similarity modes to complete regional level packet, is obtained the target area of all scale sizes；And object boundary frame matching.The present invention realizes the target area extraction of efficient, the high recall rate of indoor scene.
Description
Technical field
The invention belongs to artificial intelligence computing technique field, particularly a kind of indoor scene layout based on RGB-D images
Estimation and target area extracting method, applied to indoor service robot technology.
Background technology
The research of indoor scene parsing is one of study hotspot of domestic and foreign scholars, for the semantic positioning of Indoor Robot and
Map generation have important application value, simultaneously for solve some high level computer visual problems also have it is very important
Meaning.Target Segmentation and extraction algorithm purpose are target positioning and the example segmentation result for obtaining high quality, are scene parsings
One of committed step.Objective extraction result is usually object candidate area or object boundary frame, by development for many years, target
Extraction algorithm can be divided into two classes at present：The first kind is the algorithm based on sliding window detection thought, and the second class is based on segmentation
Algorithm, include image over-segmentation and segmentation split strategy.First kind algorithm comparison classics are DPM (Deformable Parts
Model) algorithm of target detection, there is very strong robustness using HOG features and SVM classifier, the deformation to target is improved, but
It is that this kind of algorithm calculation cost is larger, and more complicated character representation can not be used.
More classical in second class algorithm is image segmentations of the GBS (graph based segmentation) based on figure
Algorithm, the algorithm is realized simply, speed, can find out visually consistent region, but easily cause over-segmentation；Also
The Target Segmentation algorithm of constrained parameters minimal cut, the algorithm segmentation effect is good, but only includes foreground segmentation region.In recent years by
In the popularization of depth transducer, there are the RGB-D image data sets for largely including depth image, researcher starts with RGB-
D data sets lift effect by increasing geometric properties or depth information, but these algorithms generally all to there is supervision algorithm, it is necessary to
The skeleton pattern being previously obtained is trained, computation complexity is larger, although part improves the degree of accuracy of Objective extraction, target
Classification is less and recall rate is relatively low, and easily ignores plane domain object when detecting.Also there are some unsupervised RGB-D in addition
Objective extraction and partitioning algorithm, calculating speed is very fast, but more sensitive to brightness of image change, noise etc., and robustness is not high.
Although object extraction algorithm is constantly developing, due to the office of the features such as RGB image texture, color, brightness
It is sex-limited, in the indoor scene applied to complexity, problems with still be present：1) occlusion issue, because occlusion detection is less than one
A little big targets；2) the problem of plane domain object and small-size object are easily ignored so that recall rate is relatively low；3) calculate complicated
Spend larger, it is necessary to which pre-training, is unsuitable for real system application；4) it is poor that respond is influenceed on the uncertain factor in image,
Robustness is low.
The content of the invention
The present invention in order to overcome the above-mentioned deficiencies of the prior art, takes into account recall rate and rapidity, solves complex indoor scene
The problem of lower layout estimation and Objective extraction, it is proposed that a kind of estimation of indoor scene layout and target area based on RGB-D images
Domain extracting method, realize the target area extraction of efficient, the high recall rate of indoor scene.
The estimation of indoor scene layout and target area extracting method based on RGB-D images, comprise the following steps,
Step 1, scene layout's estimation：Depth map is converted into intensive 3D point cloud, by three-dimensional European between calculating point cloud
Distance carries out plane segmentation division plane domain and non-planar area, and gained plane domain is classified, and is divided into border and puts down
Face and non-boundary plane；
Step 2, image over-segmentation：Using based on the partitioning algorithm of figure with constrained parameters minimal cut algorithm to pretreated
Depth map does over-segmentation with RGB figures, obtains different size of regional ensemble；
Step 3, the packet of over-segmentation level：Utilize color, texture, size, four kinds of different measuring similarity modes of coincideing
Region merging technique is carried out to complete regional level packet, obtains the target area of all scale sizes；
Step 4, the matching of object boundary frame：To plane domain and the target of non-planar area, it is divided into boundary plane, non-border
In plane, plane domain, four kinds of non-planar area situation take the minimum rectangle bounding box that different strategy matchings includes target,
Obtain target area boundaries frame.
The present invention does plane segmentation and classification using the 3D point cloud of input, make use of the geometric continuity of a cloud to reduce and blocks
Influence to layout estimation, improve the effect of scene layout's estimation；It is minimum using the partitioning algorithm based on figure and constrained parameters
Cut algorithm and over-segmentation is done to pretreated depth map and RGB figures, combine depth information and RGB information, improve segmentation effect
Fruit；Region merging technique is carried out using four kinds of different measuring similarity modes, obtains the target area of all scale sizes, and
A variety of image conditions are considered, add the robustness of algorithm；The target area of plane domain and non-planar area is taken not
Same bounding box matching strategy, had both remained the object in plane domain, and it is undue caused by due to blocking to improve big object again
Cut problem；Rate is overlapped using bounding box and eliminates redundancy bounding box, leaves optimum target zone boundary frame, is producing less candidate side
Object boundary frame recall rate is effectively increased in the case of boundary's frame；Whole process does not need pre-training, and computation complexity is low, is easy to real
Existing, calculating speed is fast.
Brief description of the drawings
Fig. 1 is the flow of the estimation of indoor scene layout and the embodiment of target area extracting method one based on RGB-D images
Figure；
Fig. 2 is depth map and 3D point cloud schematic diagram in Fig. 1 embodiments；
Fig. 3 is plane segmentation and the classifying quality figure of different scenes；
Fig. 4 is homomorphic filtering process chart in Fig. 1 embodiments；
Fig. 5 is the bounding box design sketch of the target area extraction of an implement scene in Fig. 1 embodiments.
Embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, it is right below in conjunction with drawings and Examples
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.
Fig. 1 is estimated by the layout of the indoor scene based on RGB-D images that the embodiment of the present invention proposes and target area carries
Take the overview flow chart of method.The embodiment step is as follows
Step (1) scene layout estimates：Depth map is converted into intensive 3D point cloud first, as shown in Fig. 2 then passing through meter
The three-dimensional Euclidean distance calculated between point cloud carries out plane segmentation division plane domain and non-planar area, and gained plane domain is entered
Row classification, is divided into boundary plane and non-boundary plane.
Step (1.1) plane is split：Carry out uniformity on depth map to sample to obtain triangle point set, for therein each
Triangulation point group uses one candidate plane of RANSAC algorithmic match；Then search point in plane in 3d space, it is each in point can be by
A pixel and its corresponding 3D available points in depth map represent, when the three-dimensional Euclidean distance of a point to the plane is less than
Interior point is apart from tolerance DtolWhen, the interior point that the point is plane is defined, interior point is apart from tolerance DtolCalculate as shown in formula (1)；Finally move
Except the few tiny plane of interior quantity, and split spatial closeness or close to coplanar plane.
In formula, f is focal length, and b is the baseline length of sensor, and m is linear normalization parameter, Z representative depth values.
Step (1.2) plane is classified：According to obtained dominant plane region, it is assumed that the normal line vector of plane is towards observation
Person, calculate in the ratio between point cloud total quantity for putting cloud number and whole scene of plane another side, will be less than the plane of certain threshold value point
Class is boundary plane, and the plane more than certain threshold value is categorized as non-boundary plane.Ideally the threshold value is 0, it is contemplated that
Influence of noise, it is arranged to 0.01.Plane is classified as shown in Figure 3 with the final effect of segmentation.
Step (2) image over-segmentation：Using the partitioning algorithm based on figure and constrained parameters minimal cut algorithm to pretreatment after
Depth map and RGB figures do over-segmentation, obtain different size of regional ensemble R={ r1..., rn}。
With reference to RGB information and depth information, the image of multiple pixel scales from unlike signal passage is carried out first
Different degrees of over-segmentation, the other image of region class is obtained, bottom-up group technology is then utilized to area according to provincial characteristics
Domain carries out level packet until whole image turns into a region, to obtain the group rank figure for including all size target areas
Picture.
The segmentation of step (2.1) based on RGI color spaces：RGB triple channel images are converted into normalization RG passages to add
Brightness I passages, i.e. RGI color spaces, it is then different degrees of undue using doing three kinds to RGI images based on the dividing method of figure
Cut.
The segmentation of step (2.2) based on the gray-scale map after homomorphic filtering：First RGB image is handled plus homomorphic filtering, place
Then reason flow using the dividing method based on figure to the gray-scale map exported after processing as shown in figure 4, carry out three kinds in various degree
Over-segmentation.
The segmentation of step (2.3) based on the depth map after hole-filling：Depth map is entered using global optimization color method
Row hole-filling, three kinds of different degrees of over-segmentations then are carried out to the depth map after filling up using based on the dividing method of figure.
The segmentation of step (2.4) based on RGB-D hybrid channels：Using the foreground segmentation method of constrained parameters minimal cut, knot
RGB image information and depth information are closed, over-segmentation is carried out to the image of RGB-D hybrid channels, this method is based on such as following formula (2) institute
The energy theorem shown：
μ in formula, v ∈ N, λ ∈ R, v are the set of all pixels point, edge aggregations of the ε between adjacent pixel.CλFor cost
Function, a cost can be produced when assigning label to each pixel.Binary potential function VμvHerein as penalty, when to
When similar adjacent pixel assigns different labels, a penalty value will be produced.
Step (2.4.1) calculation cost function Cλ：
V in formulafRepresent foreground seeds, vbBackground seed is represented, λ is offset, and formula f is defined as follows formula (4)：
f(xμ)=lnpf(xμ)-lnpb(xμ) (4)
P in formulafRepresent that pixel μ belongs to the probability distribution of foreground area, add p after depth informationfIt is defined as follows formula
(5) shown in：
D is depth map in formula, and I is RGB image.The representative pixel points of j expression seed regions, these pixels are equal by K-
Value-based algorithm (k=5) is selected as regional center, and α is scale factor, and γ is scale factor.
Step (2.4.2) calculates penalty Vμv：
The similitude g (μ, υ) of two neighborhood pixels is calculated according to pixel μ and υ gPb values in formula：
σ in formula2For edge sharpening parameter, binary item V is controlledμvSlickness.RGB image and depth map are all carried out
GPb is calculated, and its linear combination is risen and is used as the final gPb values of each pixel：
GPb=α gPbr+(1-α)·gPbd (8)
GPb in formularRepresent the gPb values of pixel in RGB figures, gPbdThe gPb values of pixel in depth map are represented, α is set to
0.3。
Step (3) over-segmentation level is grouped：Region merging technique is carried out to complete using four kinds of different measuring similarity modes
Regional level is grouped, and obtains the target area of all scale sizes.
Step (3.1) calculates the similarity s (r of all adjacent areas two-by-two firsti, rj) and be added to similarity collection
Close in S, find out two region r that similarity is maximum in set SiAnd rj, it is merged into a region rt, and it is added to region
In set R；
Step (3.2) removes r from similarity setiAnd rjThe similarity in region adjacent thereto, i.e. S=S s (ri, r*),
Calculate new region rtThe similarity in region adjacent thereto, its result is added in similarity set S；
Step (3.3) repeats (3.1) and arrives (3.2) step, until whole image turns into a big region, completes the layer in region
Secondary packet, obtain the target area of all scale sizes.
Step (3.1.1) calculates color similarity：The face for the 25bins for obtaining tri- passages of RGB is normalized using L1 norms
Color Histogram so that each region obtains the vector of one 75 dimensionThen according to the vector as shown in formula (9)
Color similarity between zoning：
Step (3.1.2) calculates texture similarity：The height that variance is 1 is calculated to 8 different directions of each Color Channel
This differential, each direction of each Color Channel obtain 10bins histogram so that each region obtains one 240 dimension
VectorThen the texture similarity between the zoning as shown in formula (10)：
Step (3.1.3) calculates size similarity：Accounting as shown in formula (11) according to each area size in the picture
Size similarity between rate zoning：
I refers to whole image in formula.
Step (3.1.4) calculates similarity of coincideing：Calculate the minimum rectangle bounding box B for including two combined regionijWith two
The difference of the size in region, then according to identical similar between phase difference accounting rate zoning in the picture as shown in formula (12)
Degree：
Step (3.1.5) combines four kinds of similarities：Last similarity s (ri, rj) calculation by more than four kinds of similarities
Shown in linear combination such as following formula (13)：
s(ri, rj)=a1sc(ri, rj)+a2st(ri, rj)+a3ss(ri, rj)+a4sf(ri, rj) (13)
A in formulai∈ { 0,1 }, represents whether the similarity is used.
Step (4)：Object boundary frame matches：To plane domain and the target of non-planar area, it is divided into four kinds of situations and takes
Different strategy matchings includes the minimum rectangle bounding box of target, obtains target area boundaries frame.
Step (4.1) is all directly used for plane domain, boundary plane region；To each non-boundary plane, look for
The minimum euclidean distance with other non-boundary planes is calculated to its boundary point, the plane split that distance is less than to certain threshold value rises
Come, using the non-boundary plane region after split；And positioned at the target of plane domain, only retain the target as caused by RGB image
Region.
Step (4.2) is for non-planar area, in addition to the target area too small with non-planar area overlapping area, its
He is directly used target area.
Step (4.3) matches object boundary frame：The target area of all uses is converted into shade, to these shades one by one
Matching includes the minimum rectangle bounding box including shade, obtains shown in bounding box set B such as formulas (14)：
B=BBP+BMPR+BNPR+BPR (14)
BP represents boundary plane region in formula, and MPR represents split plane domain, and NPR represents the mesh positioned at non-planar area
Mark, PR represent the target positioned at plane domain.
Then the tiny bounding box in set B is removed, and bounding box is sorted by size, is from top to bottom iterated to calculate
Overlapping rate between bounding box, overlap rate Q (bi, bj) computational methods such as following formula (15), then filter off overlapping rate and be more than certain threshold value
Bounding box is repeated, obtains optimum target zone boundary frame set to the end, effect is as shown in Figure 5.
B in formulai, bj∈ B, a (bi) represent bounding box biArea.
Technological means disclosed in the present invention program is not limited only to the technological means disclosed in above-mentioned embodiment, in addition to
Formed technical scheme is combined by above technical characteristic.
Claims (5)
1. the estimation of indoor scene layout and target area extracting method based on RGB-D images, it is characterised in that：Including following step
Suddenly,
Step 1, scene layout's estimation：Depth map is converted into intensive 3D point cloud, by calculating the three-dimensional Euclidean distance between point cloud
Carry out plane segmentation division plane domain and non-planar area, and gained plane domain classified, be divided into boundary plane with
Non- boundary plane；
Step 2, image over-segmentation：Using the partitioning algorithm based on figure and constrained parameters minimal cut algorithm to pretreated depth
Figure does over-segmentation with RGB figures, obtains different size of regional ensemble；
Step 3, the packet of over-segmentation level：Carried out using color, texture, size, four kinds of different measuring similarity modes of coincideing
Region merging technique obtains the target area of all scale sizes to complete regional level packet；
Step 4, the matching of object boundary frame：To plane domain and the target of non-planar area, it is divided into boundary plane, non-border is put down
In face, plane domain, four kinds of non-planar area situation take different strategy matchings to include the minimum rectangle bounding box of target, obtain
To target area boundaries frame.
2. the estimation of indoor scene layout and target area extracting method according to claim 1 based on RGB-D images, its
It is characterised by：Step 1 detailed process is,
Step 1.1, plane segmentation：Carry out uniformity on depth map to sample to obtain triangle point set, for each triangle therein
Point group uses one candidate plane of RANSAC algorithmic match；Then 3D point cloud space search plane in point, when a point arrive this
The three-dimensional Euclidean distance of plane is less than interior point apart from tolerance DtolWhen, the interior point that the point is plane is defined, interior point is apart from tolerance Dtol
Calculate such as formula (1)；The tiny plane of point in finally removing, and split spatial closeness or close to coplanar plane；
<mrow>
<msub>
<mi>D</mi>
<mrow>
<mi>t</mi>
<mi>o</mi>
<mi>l</mi>
</mrow>
</msub>
<mo>=</mo>
<mrow>
<mo>(</mo>
<mfrac>
<mi>m</mi>
<mrow>
<mi>f</mi>
<mi>b</mi>
</mrow>
</mfrac>
<mo>)</mo>
</mrow>
<mo>*</mo>
<msup>
<mi>Z</mi>
<mn>2</mn>
</msup>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</mrow>
In formula, f is focal length, and b is the baseline length of sensor, and m is linear normalization parameter, Z representative depth values；
Step 1.2, plane classification：According to obtained dominant plane region, it is assumed that the normal line vector of plane calculates towards observer
In the ratio between point cloud total quantity for putting cloud number and whole scene of plane another side, the plane less than threshold value is categorized as border and put down
Face, the plane more than threshold value are categorized as non-boundary plane.
3. the estimation of indoor scene layout and target area extracting method according to claim 1 based on RGB-D images, its
It is characterised by：Step 2 detailed process is,
Step 2.1, the segmentation based on RGI color spaces：RGB triple channel images are converted into normalization RG passages and add brightness I
Passage, i.e. RGI color spaces, over-segmentation then is done to RGI images using based on the dividing method of figure；
Step 2.2, the segmentation based on the gray-scale map after homomorphic filtering：RGB image is handled plus homomorphic filtering, to defeated after processing
The gray-scale map gone out carries out over-segmentation using the dividing method based on figure；
Step 2.3, the segmentation based on the depth map after hole-filling：Cavity is carried out to depth map using global optimization color method
Fill up, over-segmentation is carried out to the depth map after filling up using based on the dividing method of figure；
Step 2.4, the segmentation based on RGB-D hybrid channels：Using the foreground segmentation method of constrained parameters minimal cut, with reference to RGB
Image information and depth information, over-segmentation, the energy theorem based on formula (2) are carried out to the image of RGB-D hybrid channels：
<mrow>
<msup>
<mi>E</mi>
<mi>&lambda;</mi>
</msup>
<mrow>
<mo>(</mo>
<mi>X</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<munder>
<mi>&Sigma;</mi>
<mrow>
<mi>&mu;</mi>
<mo>&Element;</mo>
<mi>&gamma;</mi>
</mrow>
</munder>
<msub>
<mi>C</mi>
<mi>&lambda;</mi>
</msub>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>&mu;</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>+</mo>
<munder>
<mi>&Sigma;</mi>
<mrow>
<mo>(</mo>
<mi>&mu;</mi>
<mo>,</mo>
<mi>v</mi>
<mo>)</mo>
<mo>&Element;</mo>
<mi>&epsiv;</mi>
</mrow>
</munder>
<msub>
<mi>V</mi>
<mrow>
<mi>&mu;</mi>
<mi>v</mi>
</mrow>
</msub>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>&mu;</mi>
</msub>
<mo>,</mo>
<msub>
<mi>x</mi>
<mi>v</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>2</mn>
<mo>)</mo>
</mrow>
</mrow>
μ in formula, v ∈ N, λ ∈ R, v are the set of all pixels point, edge aggregations of the ε between adjacent pixel, CλFor cost function,
A cost, binary potential function V can be produced when assigning label to each pixelμvAs penalty, when to similar adjacent
When pixel assigns different labels, a penalty value will be produced.
4. the estimation of indoor scene layout and target area extracting method according to claim 1 based on RGB-D images, its
It is characterised by：Step 3 detailed process is,
Step 3.1, the similarity set for calculating all adjacent areas two-by-two, find out two maximum region r of similarityiAnd rj,
It is merged into a new region rt, it is added in regional ensemble；
Step 3.2, remove r from similarity setiAnd rjThe similarity in region adjacent thereto, calculate new region rtIt is adjacent thereto
The similarity in region, it is added in similarity set；
Step 3.3, repeat step 3.1~3.2, until whole image turns into a big region, the level for completing image is grouped, and is obtained
Take the target area of all scale sizes.
5. the estimation of indoor scene layout and target area extracting method according to claim 1 based on RGB-D images, its
It is characterised by：Step 4 detailed process is,
Step 4.1, all directly it is used for plane domain, boundary plane region；To each non-boundary plane, its is found
Boundary point calculates the minimum euclidean distance with other non-boundary planes, and the plane that distance is less than to threshold value pieces together, using spelling
Non- boundary plane region after conjunction；And positioned at the target of plane domain, only retain the target area as caused by RGB image；
Step 4.2, for non-planar area, in addition to the target area too small with non-planar area overlapping area, other mesh
Mark region is all directly used；
Step 4.3, matching object boundary frame：The target area of all uses is converted into shade, these shades are matched one by one
Comprising the minimum rectangle bounding box including shade, tiny bounding box is then removed, and bounding box is sorted by size, by upper
The overlapping rate between bounding box is computed repeatedly under, the bounding box that overlapping rate is more than certain threshold value is filtered off, obtains target area to the end
Domain bounding box.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710442910.9A CN107369158B (en) | 2017-06-13 | 2017-06-13 | Indoor scene layout estimation and target area extraction method based on RGB-D image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710442910.9A CN107369158B (en) | 2017-06-13 | 2017-06-13 | Indoor scene layout estimation and target area extraction method based on RGB-D image |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107369158A true CN107369158A (en) | 2017-11-21 |
| CN107369158B CN107369158B (en) | 2020-11-13 |
Family
ID=60306413
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710442910.9A Active CN107369158B (en) | 2017-06-13 | 2017-06-13 | Indoor scene layout estimation and target area extraction method based on RGB-D image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107369158B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108742159A (en) * | 2018-04-08 | 2018-11-06 | 浙江安精智能科技有限公司 | Intelligent control device of water dispenser based on RGB-D cameras and its control method |
| CN109636814A (en) * | 2018-12-18 | 2019-04-16 | 联想(北京)有限公司 | A kind of image processing method and electronic equipment |
| CN110263692A (en) * | 2019-06-13 | 2019-09-20 | 北京数智源科技有限公司 | Container switch gate state identification method under large scene |
| CN111611919A (en) * | 2020-05-20 | 2020-09-01 | 西安交通大学苏州研究院 | Road scene layout analysis method based on structured learning |
| CN111815696A (en) * | 2019-04-11 | 2020-10-23 | 曜科智能科技（上海）有限公司 | Depth map optimization method, device, equipment and medium based on semantic instance segmentation |
| WO2020258793A1 (en) * | 2019-06-26 | 2020-12-30 | 北京市商汤科技开发有限公司 | Target detection and training of target detection network |
| CN113256776A (en) * | 2021-06-21 | 2021-08-13 | 炫我信息技术(北京)有限公司 | Image processing method and device, electronic equipment and computer readable storage medium |
| CN115272529A (en) * | 2022-09-28 | 2022-11-01 | 中国海洋大学 | Layout-first multi-scale decoupling ocean remote sensing image coloring method and system |
| CN116740809A (en) * | 2023-06-05 | 2023-09-12 | 嘉兴米兰映像家具有限公司 | Intelligent sofa control method based on user gesture |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090297049A1 (en) * | 2005-07-07 | 2009-12-03 | Rafael Advanced Defense Systems Ltd. | Detection of partially occluded targets in ladar images |
| CN102436654A (en) * | 2011-09-02 | 2012-05-02 | 清华大学 | Adaptive segmentation method of building point cloud |
| CN104809187A (en) * | 2015-04-20 | 2015-07-29 | 南京邮电大学 | Indoor scene semantic annotation method based on RGB-D data |
| CN105488809A (en) * | 2016-01-14 | 2016-04-13 | 电子科技大学 | Indoor scene meaning segmentation method based on RGBD descriptor |
-
2017
- 2017-06-13 CN CN201710442910.9A patent/CN107369158B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090297049A1 (en) * | 2005-07-07 | 2009-12-03 | Rafael Advanced Defense Systems Ltd. | Detection of partially occluded targets in ladar images |
| CN102436654A (en) * | 2011-09-02 | 2012-05-02 | 清华大学 | Adaptive segmentation method of building point cloud |
| CN104809187A (en) * | 2015-04-20 | 2015-07-29 | 南京邮电大学 | Indoor scene semantic annotation method based on RGB-D data |
| CN105488809A (en) * | 2016-01-14 | 2016-04-13 | 电子科技大学 | Indoor scene meaning segmentation method based on RGBD descriptor |
Non-Patent Citations (1)
| Title |
|---|
| SAURABH GUPTA等: "Perceptual Organization and Recognition of Indoor Scenes from RGB-D Images", 《COMPUTER VISION FOUNDATION》 * |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108742159A (en) * | 2018-04-08 | 2018-11-06 | 浙江安精智能科技有限公司 | Intelligent control device of water dispenser based on RGB-D cameras and its control method |
| CN109636814A (en) * | 2018-12-18 | 2019-04-16 | 联想(北京)有限公司 | A kind of image processing method and electronic equipment |
| CN111815696B (en) * | 2019-04-11 | 2023-08-22 | 曜科智能科技（上海）有限公司 | Depth map optimization method, device, equipment and medium based on semantic instance segmentation |
| CN111815696A (en) * | 2019-04-11 | 2020-10-23 | 曜科智能科技（上海）有限公司 | Depth map optimization method, device, equipment and medium based on semantic instance segmentation |
| CN110263692A (en) * | 2019-06-13 | 2019-09-20 | 北京数智源科技有限公司 | Container switch gate state identification method under large scene |
| WO2020258793A1 (en) * | 2019-06-26 | 2020-12-30 | 北京市商汤科技开发有限公司 | Target detection and training of target detection network |
| TWI762860B (en) * | 2019-06-26 | 2022-05-01 | 大陸商北京市商湯科技開發有限公司 | Method, device, and apparatus for target detection and training target detection network, storage medium |
| CN111611919A (en) * | 2020-05-20 | 2020-09-01 | 西安交通大学苏州研究院 | Road scene layout analysis method based on structured learning |
| CN113256776A (en) * | 2021-06-21 | 2021-08-13 | 炫我信息技术(北京)有限公司 | Image processing method and device, electronic equipment and computer readable storage medium |
| CN115272529B (en) * | 2022-09-28 | 2022-12-27 | 中国海洋大学 | Layout-first multi-scale decoupling ocean remote sensing image coloring method and system |
| CN115272529A (en) * | 2022-09-28 | 2022-11-01 | 中国海洋大学 | Layout-first multi-scale decoupling ocean remote sensing image coloring method and system |
| CN116740809A (en) * | 2023-06-05 | 2023-09-12 | 嘉兴米兰映像家具有限公司 | Intelligent sofa control method based on user gesture |
| CN116740809B (en) * | 2023-06-05 | 2024-03-29 | 嘉兴米兰映像家具有限公司 | Intelligent sofa control method based on user gesture |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107369158B (en) | 2020-11-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107369158A (en) | The estimation of indoor scene layout and target area extracting method based on RGB D images | |
| CN108665481B (en) | Self-adaptive anti-blocking infrared target tracking method based on multi-layer depth feature fusion | |
| CN105389584B (en) | Streetscape semanteme marking method based on convolutional neural networks with semantic transfer conjunctive model | |
| CN103810503A (en) | Depth study based method for detecting salient regions in natural image | |
| CN106909902B (en) | Remote sensing target detection method based on improved hierarchical significant model | |
| Alidoost et al. | A CNN-based approach for automatic building detection and recognition of roof types using a single aerial image | |
| CN103337072B (en) | A kind of room objects analytic method based on texture and geometric attribute conjunctive model | |
| CN105528575B (en) | Sky detection method based on Context Reasoning | |
| CN105261017A (en) | Method for extracting regions of interest of pedestrian by using image segmentation method on the basis of road restriction | |
| CN104517095B (en) | A kind of number of people dividing method based on depth image | |
| CN107564022A (en) | Saliency detection method based on Bayesian Fusion | |
| Asokan et al. | Machine learning based image processing techniques for satellite image analysis-a survey | |
| Gupta et al. | Automatic trimap generation for image matting | |
| CN113408584B (en) | RGB-D multi-modal feature fusion 3D target detection method | |
| Wang et al. | An overview of 3d object detection | |
| CN104766065A (en) | Robustness prospect detection method based on multi-view learning | |
| CN109086777A (en) | A kind of notable figure fining method based on global pixel characteristic | |
| Li et al. | Unsupervised road extraction via a Gaussian mixture model with object-based features | |
| CN111401380A (en) | RGB-D image semantic segmentation method based on depth feature enhancement and edge optimization | |
| Li et al. | Transmission line detection in aerial images: An instance segmentation approach based on multitask neural networks | |
| Bao et al. | Unpaved road detection based on spatial fuzzy clustering algorithm | |
| CN113592893B (en) | Image foreground segmentation method for determining combination of main body and accurate edge | |
| CN105354547A (en) | Pedestrian detection method in combination of texture and color features | |
| Wang et al. | A region-line primitive association framework for object-based remote sensing image analysis | |
| Ren et al. | Research on infrared small target segmentation algorithm based on improved mask R-CNN |
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 | ||
| CB02 | Change of applicant information | ||
| CB02 | Change of applicant information |
Address after: Room 201, building 2, phase II, No.1 Kechuang Road, Yaohua street, Qixia District, Nanjing City, Jiangsu Province, 210003 Applicant after: NANJING University OF POSTS AND TELECOMMUNICATIONS Address before: 210003 Gulou District, Jiangsu, Nanjing new model road, No. 66 Applicant before: NANJING University OF POSTS AND TELECOMMUNICATIONS |
|
| GR01 | Patent grant | ||
| GR01 | Patent grant |