CN107369158B - Indoor scene layout estimation and target area extraction method based on RGB-D image - Google Patents

Indoor scene layout estimation and target area extraction method based on RGB-D image Download PDF

Info

Publication number
CN107369158B
CN107369158B CN201710442910.9A CN201710442910A CN107369158B CN 107369158 B CN107369158 B CN 107369158B CN 201710442910 A CN201710442910 A CN 201710442910A CN 107369158 B CN107369158 B CN 107369158B
Authority
CN
China
Prior art keywords
plane
segmentation
target
area
rgb
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.)
Active
Application number
CN201710442910.9A
Other languages
Chinese (zh)
Other versions
CN107369158A (en
Inventor
吴晓秋
霍智勇
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Posts and Telecommunications
Original Assignee
Nanjing University of Posts and Telecommunications
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Nanjing University of Posts and Telecommunications filed Critical Nanjing University of Posts and Telecommunications
Priority to CN201710442910.9A priority Critical patent/CN107369158B/en
Publication of CN107369158A publication Critical patent/CN107369158A/en
Application granted granted Critical
Publication of CN107369158B publication Critical patent/CN107369158B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • 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 invention discloses an indoor scene layout estimation and target area extraction method based on RGB-D images, which comprises the following steps: estimating scene layout; carrying out over-segmentation on the preprocessed depth map and the preprocessed RGB map by using a map-based segmentation algorithm and a constraint parameter minimum segmentation algorithm to obtain region sets with different sizes; performing over-segmentation hierarchical grouping, and performing region merging by using four different similarity measurement modes to complete region hierarchical grouping to obtain target regions of all scales; and target bounding box matching. The method and the device realize the high-efficiency and high-recall-rate target area extraction of the indoor scene.

Description

Indoor scene layout estimation and target area extraction method based on RGB-D image
Technical Field
The invention belongs to the technical field of artificial intelligence calculation, in particular to an indoor scene layout estimation and target area extraction method based on RGB-D images, which is applied to an indoor service robot technology.
Background
The research of indoor scene analysis is one of the research hotspots of scholars at home and abroad, has important application value for indoor robot semantic positioning and map generation, and has very important significance for solving some high-level computer vision problems. The target segmentation and extraction algorithm aims to obtain high-quality target positioning and instance segmentation results, and is one of the key steps of scene analysis. The target extraction result is usually a target candidate region or a target bounding box, and through years of development, target extraction algorithms can be currently divided into two types: the first kind is algorithm based on sliding window detection idea, the second kind is algorithm based on segmentation, including image over-segmentation and segmentation amalgamation strategy. The first kind of algorithms is more classical than a DPM (deformable Parts model) target detection algorithm, and has strong robustness to deformation of a target by adopting an improved HOG feature and an SVM classifier, but the algorithms are high in calculation cost and cannot use complex feature representation.
The second type of algorithm is more classical image segmentation algorithm based on GBS (graph based segmentation), which is simple to implement and high in speed, can find out visually consistent regions, but easily causes over-segmentation; and the target segmentation algorithm with the minimum constraint parameter segmentation is good in segmentation effect and only comprises a foreground segmentation area. In recent years, due to the popularization of depth sensors, a large number of RGB-D image data sets including depth images appear, researchers begin to improve effects by increasing geometric features or depth information by using the RGB-D data sets, but these algorithms are usually supervised algorithms, a pre-obtained contour model needs to be trained, the calculation complexity is high, although the accuracy of target extraction is partially improved, the target types are few and the recall rate is low, and planar area objects are easily ignored during detection. In addition, some unsupervised RGB-D target extraction and segmentation algorithms exist, the calculation speed is high, but the algorithms are sensitive to image brightness change, noise and the like, and the robustness is low.
Although the target extraction algorithm is continuously developed, due to the limitation of features such as texture, color, brightness, etc. of RGB images, when applied to a complex indoor scene, the following problems still exist: 1) occlusion problems, where some large objects cannot be detected due to occlusion; 2) the problem that objects in a plane area and small-size objects are easy to ignore is solved, so that the recall rate is low; 3) the calculation complexity is high, pre-training is needed, and the method is not suitable for practical system application; 4) the influence on uncertain factors in the image has poor reaction capability and low robustness.
Disclosure of Invention
In order to overcome the defects of the prior art, take the recall rate and the rapidity into consideration, solve the problems of layout estimation and target extraction in a complex indoor scene, provide an indoor scene layout estimation and target area extraction method based on RGB-D images, and realize the high-efficiency and high-recall-rate target area extraction of the indoor scene.
An indoor scene layout estimation and target area extraction method based on RGB-D images comprises the following steps,
step 1, scene layout estimation: converting the depth map into dense 3D point clouds, performing plane segmentation by calculating three-dimensional Euclidean distances among the point clouds to divide a plane area and a non-plane area, and classifying the obtained plane area into a boundary plane and a non-boundary plane;
step 2, image over-segmentation: carrying out over-segmentation on the preprocessed depth map and the preprocessed RGB map by using a map-based segmentation algorithm and a constraint parameter minimum segmentation algorithm to obtain region sets with different sizes;
step 3, over-segmentation hierarchical grouping: combining the regions by utilizing four different similarity measurement modes of color, texture, size and coincidence to complete region level grouping and obtain target regions of all scales;
step 4, matching a target bounding box: and adopting different strategies to match the minimum rectangular bounding box containing the target under four conditions of dividing the target of the plane area and the non-plane area into a bounding plane, a non-bounding plane, the plane area and the non-plane area to obtain the bounding box of the target area.
According to the method, the input 3D point cloud is used for plane segmentation and classification, the influence of occlusion on layout estimation is reduced by using the geometric continuity of the point cloud, and the effect of scene layout estimation is improved; the preprocessed depth map and the preprocessed RGB map are subjected to over-segmentation by using a map-based segmentation algorithm and a constraint parameter minimum segmentation algorithm, and the depth information and the RGB information are combined, so that the segmentation effect is improved; four different similarity measurement modes are utilized to carry out region combination, target regions with all sizes are obtained, multiple image conditions are considered, and the robustness of the algorithm is improved; different bounding box matching strategies are adopted for target areas of the planar area and the non-planar area, so that objects in the planar area are reserved, and the problem of over-segmentation of large objects due to shielding is solved; redundant bounding boxes are eliminated by utilizing the bounding box overlapping rate, the best target area bounding box is left, and the target bounding box recall rate is effectively improved under the condition that fewer candidate bounding boxes are generated; the whole process does not need pre-training, the calculation complexity is low, the realization is easy, and the calculation speed is high.
Drawings
FIG. 1 is a flow chart of an embodiment of a RGB-D image based indoor scene layout estimation and target region extraction method;
FIG. 2 is a schematic view of a depth map and a 3D point cloud of the embodiment of FIG. 1;
FIG. 3 is a diagram of the effect of plane segmentation and classification in different scenes;
FIG. 4 is a flow chart of homomorphic filtering process in the embodiment of FIG. 1;
FIG. 5 is a diagram illustrating the effect of a bounding box in extracting a target region of an implementation scenario in the embodiment of FIG. 1.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Fig. 1 is a general flowchart of an indoor scene layout estimation and target area extraction method based on RGB-D images according to an embodiment of the present invention. The procedure of this example is as follows
Step (1) scene layout estimation: firstly, a depth map is converted into dense 3D point clouds, as shown in fig. 2, then, plane segmentation is carried out by calculating three-dimensional Euclidean distances among the point clouds to divide a plane area and a non-plane area, and the obtained plane area is classified into a boundary plane and a non-boundary plane.
Step (1.1) plane segmentation: carrying out consistent sampling on the depth map to obtain a triangular point set, and matching each triangular point group in the triangular point set with a candidate plane by adopting a RANSAC algorithm; then searching in-plane points in the 3D space, wherein each in-plane point can be represented by a pixel point in the depth map and a corresponding 3D effective point, and when the three-dimensional Euclidean distance from one point to the plane is smaller than the distance tolerance D of the in-plane pointstolDefining the point as an inner point of the plane, the distance tolerance D of the inner pointtolCalculating as shown in formula (1); and finally removing the tiny planes with few inner points, and splicing the planes which are close to or close to the same plane in space.
Figure GDA0002660270520000031
In the formula, f is a focal length, b is a base length of the sensor, m is a linear normalization parameter, and Z represents a depth value.
Step (1.2) plane classification: and according to the obtained main plane area, assuming that a normal vector of the plane faces an observer, calculating the ratio of the point cloud number on the other surface of the plane to the total point cloud number of the whole scene, classifying the plane smaller than a certain threshold value as a boundary plane, and classifying the plane larger than the certain threshold value as a non-boundary plane. Ideally, the threshold is 0, and is set to 0.01 in consideration of the influence of noise. The final effect of the plane classification and segmentation is shown in fig. 3.
And (2) image over-segmentation: and (3) carrying out over-segmentation on the preprocessed depth map and the preprocessed RGB map by using a graph-based segmentation algorithm and a constraint parameter minimum segmentation algorithm to obtain a region set R (R) with different sizes1，…，rn}。
Combining RGB information and depth information, firstly, carrying out over-segmentation of a plurality of pixel-level images from different signal channels to different degrees to obtain area-level images, and then carrying out hierarchical grouping on the areas by using a bottom-up grouping method according to area characteristics until the whole image becomes an area so as to obtain a group-level image containing all large and small target areas.
Step (2.1) is based on segmentation of the RGI color space: the RGB three-channel image is converted into a normalized RG channel and a brightness I channel, namely an RGI color space, and then the RGI image is subjected to three over-segmentation in different degrees by adopting a graph-based segmentation method.
And (2.2) segmentation based on the homomorphic filtered gray level image: homomorphic filtering is firstly carried out on the RGB image, the processing flow is shown in figure 4, and then the processed and output gray level image is subjected to over-segmentation of three different degrees by adopting a graph-based segmentation method.
And (2.3) segmenting the depth map after filling the cavity: and filling holes in the depth map by using a global optimization coloring method, and then performing over-segmentation on the filled depth map in three different degrees by using a graph-based segmentation method.
Step (2.4) segmentation based on RGB-D hybrid channels: the method is based on an energy formula shown as the following formula (2):
Figure GDA0002660270520000041
in the formula, lambda belongs to R, and v is a set of all pixel points and is an edge set between adjacent pixels. CλAs a cost function, a cost is incurred when assigning a label to each pixel. Binary potential function VμvHere, as a penalty function, a penalty value is generated when similar neighboring pixels are assigned different labels.
Step (2.4.1) of calculating a cost function Cλ：
Figure GDA0002660270520000042
In the formula vfRepresents the foreground seed, vbRepresents the background seed, λ is the offset, and formula f is defined as formula (4):
f(xμ)＝lnpf(xμ)-lnpb(xμ) (4)
in the formula pfRepresenting the probability distribution of pixel point mu belonging to foreground region, p after adding depth informationfThe definition is shown in the following formula (5):
Figure GDA0002660270520000043
in the formula, D is a depth map, and I is an RGB image. j represents representative pixel points of the seed region, the pixel points are selected by a K-means algorithm (K is 5) to serve as the center of the region, alpha is a scale factor, and gamma is a scale factor.
Step (2.4.2) calculating penalty function Vμv：
Figure GDA0002660270520000044
The similarity g (μ, ν) between two neighboring pixels is calculated according to gPb values of the pixel points μ and v:
Figure GDA0002660270520000045
in the formula sigma2For the edge-sharpening parameter, the binary term V is controlledμvThe smoothness of the surface. gPb calculation is performed on the RGB image and the depth map, and the RGB image and the depth map are linearly combined to be used as a final gPb value of each pixel:
gPb＝α·gPbr+(1-α)·gPbd (8)
formula (III) gPbrgPb values, gPb, representing pixels in an RGB mapdThe gPb value representing a pixel in the depth map is set to 0.3.
Step (3) over-segmentation hierarchical grouping: and performing region combination by using four different similarity measurement modes to complete region level grouping and obtain target regions with all dimensions.
Step (3.1) first of all the similarity s (r) of every two adjacent regions is calculatedi，rj) Adding the similarity into a similarity set S, and finding out two regions r with the maximum similarity in the set SiAnd rjMerge them into one region rtAnd added to the region set R;
step (3.2) removing r from the similarity setiAnd rjSimilarity with its adjacent region, i.e. S ═ S \ S (r)i，r*) Calculating a new region rtAdding the similarity of the adjacent regions to a similarity set S;
and (3.3) repeating the steps (3.1) to (3.2) until the whole image becomes a large area, completing the hierarchical grouping of the areas and acquiring the target areas with all dimensions.
Calculating the color similarity by the step (3.1.1): color histograms of 25bins for the three RGB channels were obtained using L1 norm normalization, so that each region resulted in a 75-dimensional vector
Figure GDA0002660270520000051
Then, the color similarity between the regions is calculated according to the vector as shown in formula (9):
Figure GDA0002660270520000052
calculating the texture similarity by the steps of (3.1.2): calculating Gaussian differential with variance of 1 for 8 different directions of each color channel, and acquiring a 10bins histogram for each direction of each color channel, so that each region obtains a 240-dimensional vector
Figure GDA0002660270520000053
Then, the texture similarity between the regions is calculated as shown in equation (10):
Figure GDA0002660270520000054
calculating the size similarity by the step (3.1.3): the size similarity between the regions is calculated according to the ratio of each region size in the image as shown in equation (11):
Figure GDA0002660270520000055
wherein I refers to the entire image.
Calculating the anastomosis similarity according to the following steps (3.1.4): calculating the minimum rectangular bounding box B containing two merged regionsijAnd the difference between the sizes of the two regions, and then calculating the coincidence similarity between the regions according to the ratio of the difference value in the image as shown in formula (12):
Figure GDA0002660270520000056
step (3.1.5) combines four similarities: final similarity s (r)i，rj) The calculation method of (2) is shown by the following formula (13) by linear combination of the above four similarities:
s(ri，rj)＝a1sc(ri，rj)+a2st(ri，rj)+a3ss(ri，rj)+a4sf(ri，rj) (13)
in the formula aiE {0, 1}, indicates whether the similarity is used.
And (4): matching a target bounding box: and for the targets in the planar area and the non-planar area, adopting different strategies to match the minimum rectangular bounding box containing the targets under four conditions to obtain the bounding box of the target area.
Step (4.1) for the plane areas, the boundary plane areas are directly adopted; for each non-boundary plane, finding boundary points of the non-boundary plane, calculating the minimum Euclidean distance between the non-boundary plane and other non-boundary planes, splicing the planes with the distance less than a certain threshold value, and adopting the spliced non-boundary plane areas; while the object located in the plane area, only the object area generated by the RGB image remains.
And (4.2) for the non-planar area, other target areas except the target area with the small overlapping area with the non-planar area are directly adopted.
Step (4.3) matching the target bounding box: converting all adopted target areas into masks, and matching the masks with the minimum rectangular bounding boxes containing the masks one by one to obtain a bounding box set B shown as a formula (14):
B＝BBP+BMPR+BNPR+BPR (14)
where BP represents a bounding plane region, MPR represents a stitching plane region, NPR represents an object located in a non-planar region, and PR represents an object located in a planar region.
Then removing the tiny boundary frames in the set B, sequencing the boundary frames according to the area size, and iteratively calculating the overlapping rate between the boundary frames from top to bottom, wherein the overlapping rate O (B)i，bj) The calculation method is as the following formula (15), then the repeated bounding boxes with the overlapping rate larger than a certain threshold are filtered out to obtain the final optimal target area bounding box set, and the effect is as shown in fig. 5Shown in the figure.
Figure GDA0002660270520000061
In the formula bi，bj∈B，a(bi) Represents a bounding box biThe area of (a).
The technical means disclosed in the invention scheme are not limited to the technical means disclosed in the above embodiments, but also include the technical scheme formed by any combination of the above technical features.

Claims (5)

1. An indoor scene layout estimation and target area extraction method based on RGB-D images is characterized in that: comprises the following steps of (a) carrying out,
step 1, scene layout estimation: converting the depth map into dense 3D point clouds, performing plane segmentation by calculating three-dimensional Euclidean distances among the point clouds to divide a plane area and a non-plane area, and classifying the obtained plane area into a boundary plane and a non-boundary plane;
step 2, image over-segmentation: carrying out over-segmentation on the preprocessed depth map and the preprocessed RGB map by using a map-based segmentation algorithm and a constraint parameter minimum segmentation algorithm to obtain region sets with different sizes;
step 3, over-segmentation hierarchical grouping: combining the regions by utilizing four different similarity measurement modes of color, texture, size and coincidence to complete region level grouping and obtain target regions of all scales;
step 4, matching a target bounding box: for the target of the plane area, the boundary plane area is directly adopted, for each non-boundary plane, the boundary point of the non-boundary plane is found, the minimum Euclidean distance between the non-boundary plane and other non-boundary planes is calculated, the planes with the distance smaller than a certain threshold value are spliced, the spliced non-boundary plane area is adopted, the target in the plane area is adopted, and only the target area generated by the RGB image is reserved; for the target of the non-planar area, except the target area with the small overlapping area with the non-planar area, other target areas are directly adopted, all the adopted target areas are converted into masks, and the masks are matched with the minimum rectangular bounding box containing the masks one by one to obtain the bounding box.
2. The RGB-D image based indoor scene layout estimation and target area extraction method as claimed in claim 1, wherein: the specific process of the step 1 is that,
step 1.1, plane segmentation: carrying out consistent sampling on the depth map to obtain a triangular point set, and matching each triangular point group in the triangular point set with a candidate plane by adopting a RANSAC algorithm; then searching the plane inner point in the 3D point cloud space, and when the three-dimensional Euclidean distance from one point to the plane is smaller than the distance tolerance D of the inner pointtolDefining the point as an inner point of the plane, the distance tolerance D of the inner pointtolCalculating the formula (1); finally, removing the inner point fine plane, and splicing planes which are close to or close to the same plane in space;
Figure FDA0002660270510000011
in the formula, f is a focal length, b is a base length of the sensor, m is a linear normalization parameter, and Z represents a depth value;
step 1.2, plane classification: and according to the obtained main plane area, assuming that the normal vector of the plane faces an observer, calculating the ratio of the point cloud number on the other surface of the plane to the total point cloud number of the whole scene, classifying the plane smaller than the threshold value as a boundary plane, and classifying the plane larger than the threshold value as a non-boundary plane.
3. The RGB-D image based indoor scene layout estimation and target area extraction method as claimed in claim 1, wherein: the specific process of the step 2 is that,
step 2.1, segmentation based on RGI color space: converting the RGB three-channel image into a normalized RG channel and a brightness I channel, namely an RGI color space, and then performing over-segmentation on the RGI image by adopting a graph-based segmentation method;
2.2, segmentation based on the homomorphic filtered gray level image: homomorphic filtering is carried out on the RGB image, and the processed and output gray level image is subjected to over-segmentation by adopting a graph-based segmentation method;
and 2.3, segmenting the depth map after filling the cavity: filling holes in the depth map by using a global optimization coloring method, and over-segmenting the filled depth map by using a map-based segmentation method;
2.4, segmentation based on an RGB-D mixed channel: the foreground segmentation method of constraint parameter minimum segmentation is adopted, RGB image information and depth information are combined, the image of an RGB-D mixed channel is subjected to over-segmentation, and the energy formula based on the formula (2) is as follows:
Figure FDA0002660270510000021
in the formula, lambda belongs to R, v is the set of all pixel points and is the edge set between adjacent pixels, CλAs a cost function, a cost is generated when each pixel is labeled, a binary potential function VμvAs a penalty function, a penalty value is generated when similar neighboring pixels are assigned different labels.
4. The RGB-D image based indoor scene layout estimation and target area extraction method as claimed in claim 1, wherein: the specific process of the step 3 is that,
step 3.1, calculating similarity sets of all every two adjacent regions, and finding out two regions r with the maximum similarityiAnd rjMerge them into a new region rtAdding the data into the region set;
step 3.2, remove r from the similarity setiAnd rjSimilarity with its adjacent region, calculating new region rtAdding the similarity with the adjacent region into a similarity set;
and 3.3, repeating the steps 3.1-3.2 until the whole image becomes a large area, completing the hierarchical grouping of the images and acquiring target areas with all dimensions.
5. The RGB-D image based indoor scene layout estimation and target area extraction method as claimed in claim 1, wherein: the specific process of the step 4 is that,
step 4.1, as for the plane areas, the boundary plane areas are directly adopted; for each non-boundary plane, finding the boundary point of the non-boundary plane, calculating the minimum Euclidean distance between the non-boundary plane and other non-boundary planes, splicing the planes with the distance less than a threshold value, and adopting the spliced non-boundary plane area; the target located in the plane area only reserves the target area generated by the RGB image;
step 4.2, for the non-planar area, except for the target area with the small overlapping area with the non-planar area, other target areas are directly adopted;
4.3, matching a target boundary box: converting all adopted target areas into masks, matching the masks with the minimum rectangular bounding boxes including the masks one by one, then removing the fine bounding boxes, sequencing the bounding boxes according to the area size, repeatedly calculating the overlapping rate between the bounding boxes from top to bottom, filtering the bounding boxes with the overlapping rate larger than a certain threshold value, and obtaining the final bounding box of the target area.
CN201710442910.9A 2017-06-13 2017-06-13 Indoor scene layout estimation and target area extraction method based on RGB-D image Active CN107369158B (en)

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 CN107369158A (en) 2017-11-21
CN107369158B true 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)

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108742159B (en) * 2018-04-08 2021-01-15 徐育 Intelligent control device and method for water dispenser based on RGB-D camera
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
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
CN110298298B (en) * 2019-06-26 2022-03-08 北京市商汤科技开发有限公司 Target detection and target detection network training method, device and equipment
CN111611919B (en) * 2020-05-20 2022-08-16 西安交通大学苏州研究院 Road scene layout analysis method based on structured learning
CN113256776B (en) * 2021-06-21 2021-10-01 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
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

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
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

Family Cites Families (1)

* Cited by examiner, † Cited by third party
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

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
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)

* Cited by examiner, † Cited by third party
Title
Perceptual Organization and Recognition of Indoor Scenes from RGB-D Images;Saurabh Gupta等;《Computer Vision Foundation》;20131231;第564-571页 *

Also Published As

Publication number Publication date
CN107369158A (en) 2017-11-21

Similar Documents

Publication Publication Date Title
CN107369158B (en) Indoor scene layout estimation and target area extraction method based on RGB-D image
Du et al. Car detection for autonomous vehicle: LIDAR and vision fusion approach through deep learning framework
Chen et al. Moving-object detection from consecutive stereo pairs using slanted plane smoothing
CN108537239B (en) Method for detecting image saliency target
CN105261017A (en) Method for extracting regions of interest of pedestrian by using image segmentation method on the basis of road restriction
CN110610505A (en) Image segmentation method fusing depth and color information
Gupta et al. Automatic trimap generation for image matting
Wang et al. An overview of 3d object detection
CN112036231B (en) Vehicle-mounted video-based lane line and pavement indication mark detection and identification method
WO2019071976A1 (en) Panoramic image saliency detection method based on regional growth and eye movement model
EP3973507B1 (en) Segmentation for holographic images
CN113506318A (en) Three-dimensional target perception method under vehicle-mounted edge scene
Mohammadi et al. 2D/3D information fusion for building extraction from high-resolution satellite stereo images using kernel graph cuts
Chen et al. Underwater sonar image segmentation combining pixel-level and region-level information
Khandare et al. A survey paper on image segmentation with thresholding
Li et al. 3DCentripetalNet: Building height retrieval from monocular remote sensing imagery
Kozonek et al. On the fusion of camera and lidar for 3D object detection and classification
CN112037230B (en) Forest image segmentation method based on superpixels and hyper-metric profile map
Huang et al. Road scene segmentation via fusing camera and lidar data
Huang et al. Integrating visual and range data for road detection
Pankaj et al. Theoretical concepts and technical aspects on image segmentation
CN110991313B (en) Moving small target detection method and system based on background classification
Mehltretter et al. Illumination invariant dense image matching based on sparse features
Somawirata et al. Road Detection Based on Statistical Analysis
Jin et al. A dense depth estimation method using superpixels

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

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

CB02 Change of applicant information
GR01 Patent grant
GR01 Patent grant