computer vision detecting the existence, pose and position of known objects within an image michael...

21
Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

Upload: natalie-hodge

Post on 17-Dec-2015

220 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

Computer Vision Detecting the existence, pose and position of known

objects within an image

Michael Horne, Philip Sterne (Supervisor)

Page 2: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

Background

• Computer vision is a diverse field– Many schools of thought

• Many different ways of achieving– Monocular– Stereoscopic (quite popular)

Page 3: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

Problem Statment

• Recognize objects within a single image– Cutting down the total information in

the images– Extracting and building up useful subset– Using that information to recognize

objects• Model Based system

– System already knows of objects

Page 4: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

System structure

• Image capture and segmentation– Pre-process image– Extract useful information

• Matching– Find links between image and objects

known to the system• Pose Solution

– Use the matches to estimate the position and orientation of the object

Page 5: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

Image Segmentation

• Feature extraction– Images have lots of information– What features are of interest?– Edges are definitive attributes of

objects– Corners can easily be matched against

Page 6: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

Image Segmentation

• Gaussian filter is applied to reduce noise– Noise adds complexity, but no useful

information.• Canny Edge detector is applied to

extract edges

Page 7: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

Image Segmentation

• Arbitrary shaped edges are converted into straight line segments.

Page 8: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

Similar Geometries

• Need image data to be represented in a similar way to the object model

• Objects are stored as wireframes

• Image data converted to a wireframe like form.

Page 9: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

Matching

• To estimate a pose we find correspondences with corners

• Classification problem– Which object to match?– Or what data to match to which objects?

(Multiple object case)• Complex problem• Random Sample Consensus

approach

Page 10: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

Matching

• Take only the minimum amount of image data needed to estimate a solution

• But do it at random• Then test the validity of the

estimation

Page 11: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

Pose Estimation

• Now using the correspondences• Algorithm based on POSIT

– A quick method– In tune with the RANSAC approach

• POSIT– Scaled orthographic projection model

Page 12: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

Pose Estimation

• POSIT – Minimum correspondences is four– Will solve regardless correctness

• Validation is necessary

Page 13: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

Pose Estimation

• Checking need to be rapid• Various levels of verification

– Object must not be skewed in making the four points of the object to the image points.

– Distance to estimated position must be minimal

• Estimations that pass undergo further checking

Page 14: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

Pose Estimation

• Next step is to project object over image– Number of matches is expanded.– All forward facing vertices are checked

for a corresponding image point• Also the geometries are verified

Page 15: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

Pose Estimation

• Goodness ratio = corner ratio * edge ratio– General means of judging the fitted

model

– The higher the ratio, the better the fit• Model fitting with the best ratio is

chosen

Page 16: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

Final

• End result

Page 17: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

Multiple objects

• After each model is fitted the points used are marked as used– Independence of objects

• The matching process restarts

Page 18: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

Results

• Tests were setup for a set of four different objects

• Varying degrees of symmetry

• Images of various poses were captured

• Varying difficulty, degenerate poses to definitive

• 85% success in estimating pose of single objects

Page 19: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

Other examples

1 2 3

Page 20: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)

Multiple Objects

• Multiple objects solved

Page 21: Computer Vision Detecting the existence, pose and position of known objects within an image Michael Horne, Philip Sterne (Supervisor)