frank bergschneider february 21, 2014 presented to national instruments

15
Real Time Object Recognition Using the SURF Algorithm Frank Bergschneider February 21, 2014 Presented to National Instruments

Upload: christian-rodgers

Post on 20-Jan-2018

214 views

Category:

Documents


0 download

DESCRIPTION

Object Recognition 3 General Steps 1. Interest Point Detection 2. Interest Point Description Feature Vector Extraction 3. Feature Vector Matching Between Two Images

TRANSCRIPT

Page 1: Frank Bergschneider February 21, 2014 Presented to National Instruments

Real Time Object Recognition Using the SURF Algorithm

Frank BergschneiderFebruary 21, 2014

Presented to National Instruments

Page 2: Frank Bergschneider February 21, 2014 Presented to National Instruments

IntroductionObject Recognition Real Time, Online Applications Applications: Security, Shopping, Augmented

Reality Speeded Up Robust Features Algorithm OpenCV Library

Page 3: Frank Bergschneider February 21, 2014 Presented to National Instruments

Object Recognition3 General Steps

1. Interest Point Detection2. Interest Point Description Feature Vector

Extraction3. Feature Vector Matching Between Two

Images

Page 4: Frank Bergschneider February 21, 2014 Presented to National Instruments

Object RecognitionLoad training image Detect training interest points Extract training interest point descriptorsInitialize match objectInitialize and open camera feed While (Not User Exit)

Grab video frame Detect interest pointsExtract descriptors Match query points with training points If (Matching Points > Threshold)

Compute Homography Transform Box

Draw Box on Object and Display Else Continue

End While

Page 5: Frank Bergschneider February 21, 2014 Presented to National Instruments

Interest Point Detection Approximating the determinant of the

Hessian matrixIf determinant is local max -> Interest PointHessian approx. using Integral Images and

Box FiltersBig decreases in calculation time (constant

time calcs!)

Page 6: Frank Bergschneider February 21, 2014 Presented to National Instruments

Interest Point Detection Integral Image

Hessian Matrix

Page 7: Frank Bergschneider February 21, 2014 Presented to National Instruments

Interest Point Detection Interest Points found at different scales Box Filter scaled up, instead of down

sampling image

The first octave interest points are on the left side, and the second and third octave interest points are the right side for clarity

Page 8: Frank Bergschneider February 21, 2014 Presented to National Instruments

Interest Point Detection Calculate Integral Image, ILoop For Each Point

Loop For OctaveCalculate dxx, dyy, dxy with Box FilterNormalize ResponsesCalculate DeterminantIf (Determinant > Threshold)

Store Point as Interest Point at Scale

End IfEnd Loop Octave

End Loop For Each PointSuppress Non Maximum Interest PointsInterpolate Interest Points Between OctavesOutput: Interest Points

Page 9: Frank Bergschneider February 21, 2014 Presented to National Instruments

Feature Vector Description Interest Point Characterized with Haar

WaveletsOrientation Information ExtractedSquare Area Around IP in Direction of

OrientationSpatial Intensity Distribution

Page 10: Frank Bergschneider February 21, 2014 Presented to National Instruments

Feature Vector Description

Page 11: Frank Bergschneider February 21, 2014 Presented to National Instruments

Feature Vector Description For Each Interest Point

Calculate Orientation Window Area of 20s around

PointDivide Window into 4x4

subareasFor Each Subarea

Calculate Haar Wavelet

Smooth with GaussianForm Feature Vector

End For Each SubareaStore Feature Vector

End For Each Interest PointOutput: Feature Vector

Page 12: Frank Bergschneider February 21, 2014 Presented to National Instruments

Feature Vector Matchingk-Nearest Neighbors Fast Library for Approximated Nearest

Neighbors (FLANN) 2 Nearest Neighbors Found Distance Ratio Rule for Good Matches If Number of Good Matches > Threshold

Then Object Recognized

Page 13: Frank Bergschneider February 21, 2014 Presented to National Instruments

Feature Vector Matching

Page 14: Frank Bergschneider February 21, 2014 Presented to National Instruments

ResultsKey Parameter: Min Determinant of Hessian Balance Btw Detector and Descriptor at

det(H)=4500 Best: det(H)=2000, d_Ratio=0.6, Good

Matches=8 Frames Per Second: 2.9

Page 15: Frank Bergschneider February 21, 2014 Presented to National Instruments

ConclusionSURF Algorithm implemented and test with

OpenCV Robust to Scale Change, Rotation, and Noise Sufficiently Fast for Real Time Object

Recognition Integral Image and Box Filter yield huge

performance Algorithm prime for FPGA/GPU

implementation