3d scanning and modelling using kinect

5/26/2018 3D Scanning and Modelling Using Kinect

1/15

Manindra Kumar Moharana

2K10\EC\079

Piyush

2K10\EC\097Mohit Thakran

2K10\EC\083

Nitin Verma

2K10\EC\093

Guided by

Mr Alok Kumar Singh

8thsemester, April 2014

Major Project II


2/15

The goal of this project is to reconstruct 3D models ofreal objects by analysing the Kinect Fusion algorithmthat allows real-time reconstruction and rendering ofa real world scene and implementation of it.

The point clouds obtained is used to create a 3Dscanner application to produce the 3D mesh of adesired object as the output of the scene.

Introduction


3/15

Kinect was launched in North America on 4 November 2010

Kinect is a motion sensing input device .

Is used for:

Xbox 360 console

Windows PCs

Enables users to control and interact with the application/game

Kinect competes with the:

Wii Remote PlusPlayStation MovePlayStation Eye motion

Kinect


4/15

Projects a known pattern

(Speckles) in Near-Infraredlight.

CMOS IR camera observes thescene.

Calibration between theprojector and camera has to beknown.

Triangulation of each specklebetween a virtual image(pattern) and observed pattern

Kinect Working Principle


5/15

For a calibrated projector-

camera system the 2Dcorrespondence is reduced in

purely horizontal matching

Determining correspondencesbetween each row of theprojected pattern and a row ofthe rectified camera image

The relative distance between apoint in the illuminationpattern and its position in thecaptured image is inversely

related to depth

Calculation of Depth Data


6/15

The depth map is constructed by analysing a speckle patternof infrared laser light

The Kinect dramatically improves the accuracy of traditionaldepth from focus

The Kinect uses a special (astigmatic) lens with differentfocal length in x- and y- directions

A projected circle then becomes an ellipse whose orientationdepends on depth .


7/15

A stand-alone, large scale, open project for 2D/3D image

and point cloud processing

Used to work at a high level with point clouds

A set of high-technology easy-to-use modules that allowthe developer to focus on the results of imagemanipulation and not on the need to re-invent the wheel

libraries allow the developer to perform filtering, featureextraction, keypoint and model detection, registration,neighbour search, segmentation, surface reconstruction,input/output handling and visualization

Point Cloud Library (PCL)


8/15

We start with two point clouds, cloud A and B, that

contain on a planar surface the desired object placed

in different orientations Processing done through three phases:

Filtering

Alignment

Reconstruction

Finally the mesh is reconstructed

Algorithm Overview


9/15

Two clouds, A and B, are first down-sampled using a

voxel grid filter to reduce the number of points in the

cloud for further computation Cloud A and B are then cropped using a crop-box

filter, that is a high pass filter on the three spatialdimensions

A final filtering is performed to remove the noisefrom the cloud, using a radius outlier removalmethod, which removes points which do not haveenough neighbours in a sphere of chosen radiusaround them

Filtering


10/15

The two clouds need to be aligned, so that they can

be merged together

A first alignment is performed using the SampledConsensus Initial Alignment algorithm (SAC-IA)

We compute Fast Point Feature Histograms (FPFHs)for the input cloud and use these features for the

SAC-IA alignment The alignment is then refined using ICP, resulting in

an almost perfect match of the two clouds.

Alignment


11/15

Greedy triangulation algorithm is used for fast

reconstruction

It creates polygons by linking points and thustransforming them into vertices

Least Moving Square (LMS) algorithm is used tosmooth the point cloud before final reconstruction

Output is a 3D mesh of the object, which can besaved for use with other modelling software

Reconstruction


12/15

Sample Output

Cloud : The object is resting on its base. Cloud: The object is resting on one ofits sides.

The object


13/15

From the input/output module, we use methods to load

the point clouds for use with PCL from .pcd files

From the features module, we use normal estimationmethods, used both by the alignment phase andreconstruction, and feature extraction methods.

From the registration module, we use the implementationsof the SAC-IA and ICP algorithms.

Implementation Idea


14/15

From thefilters module, we use the voxel grid filter, the cropbox filter and the radius outlier removal filter.

From the surface module, we use the MLS implementation andthe greedy triangulation implementation

From the kdtree module, we use the tree searchimplementation, used for registration and triangulation

From the visualization module, we use the PCLVisualizer classto visualize our result mesh.


15/15

Thank You

3d scanning and modelling using kinect

Documents

d scanning

applicationgame kinect

kinect fusion algorithmthat

d mesh of adesired object

point clouds

d models ofreal objects

surface reconstruction

computation cloud