ME/CS 132: Introduction to Vision-based Robot Navigation! Low-level Image Processing" Larry Matthies"

Transcription

1 ME/CS 132: Introduction to Vision-based Robot Navigation! Low-level Image Processing" Larry Matthies" "

2 Announcements" First homework grading is done! Second homework is due next Tuesday! Third homework will be due Feb 8; combines material on outlier detection and egomotion! (Next quarter)! lhm - 2

3 Recap and Where Weʼre Going: 1 st Module (Done)" Image formation, cameras, and camera calibration! Illumination and radiometry! Sources of light sunlight, thermal emission, night sky glow! Propagation of light reflection from surfaces, attenuation in media! Cameras! Basic optics! Camera architectures! Image detectors materials, architectures, performance, for various regions of the EM spectrum! Geometric camera modeling and calibration! lhm - 3

4 Recap and Where Weʼre Going: 2 nd Module (Midway Through)" Visual motion estimation and 3-D perception! Low-level image processing! Feature detection, matching, and outlier detection! Pose estimation (egomotion) and visual odometry! Dense range imaging with stereo vision! Other range sensors, range data analysis, introduction to robots! 1-week mini-project on visual localization using a stereo camera head to match landmark points!? lhm - 4

5 Recap and Where Weʼre Going: 3 rd Module" State estimation, localization, and mapping! Introduction to estimation! Linear Kalman filter! Extended Kalman filter! Particle filters and the UKF! Simultaneous localization and mapping! 1-week mini-project on stereo-based SLAM for localization and occupancy grid mapping with ladar! lhm - 5

6 This Lecture" Point operators! Neighborhood operators! Edge detection! Image pyramids! Geometric image transformations! Reading material:! Today: Szeliski ch. 3 and sections ! Next Tuesday: Szeliski ch. 11! lhm - 6

7 Images as Functions" f(x,y) y Interpret image either as:! Continuous image f(x,y)! Discrete array f[x,y]! Images are affected by:! Noise! Nonlinear distortions of intensity and geometry! x lhm - 7

8 Image Noise" lhm - 8

9 Image Distortions" lhm - 9

10 Point Operators" Output pixel = function of one input pixel in one or more images! g(x) = h(f(x)) or g(x) = h(f 0 (x),..., f n (x)) g(i, j) = h(f(i, j)) Examples! Monadic linear: bias and gain adjustment,! g(x) = a f(x) + b Monadic nonlinear: gamma correction,! Dyadic linear: image blending,! g(x) = [f(x)] 1/ϒ g(x) = (1 - α) f 0 (x) + α f 1 (x) lhm - 10

11 Point Operators: Histogram Equalization" g(x) = h(f(x)) or g(x) = h(f 0 (x),..., f n (x)) g(i, j) = h(f(i, j)) g(x) = a f(x) + b g(x) = [f(x)] 1/ϒ g(x) = (1 - α) f 0 (x) + α f 1 (x) lhm - 11

12 Point Operators: Color Space Conversion (e.g. RGB to HSV)" HSV(i, j) = f ( RGB(i, j) ) lhm - 12

13 Application of Color Space Conversion: Simple Segmentation" (a) (b) (c) (d) lhm - 13

14 Neighborhood Operators (or Window Operators, or Local Operators)" Slide one small window of numbers over the image and compute some function, replacing the central pixel in the image with the output of the function.! lhm - 14

15 Linear Filters" h() is called the filter, or kernel, or mask; entries in h() are the filter coefficients. Abbreviated notation:! Can also be written:! Correlation! Convolution! lhm - 15

16 Some Basic Properties of Convolution" Commutative:! f * g = g * f! Associative:! f * (g * h) = (f * g) * h! Distributes over addition:! Differentiation:! Shift invariant! f * (g + h) = f * g + f * h! (f * g)ʼ = fʼ * g = f * gʼ! lhm - 16

17 Neighborhood Operators: Border Effects" When applying convolution with a KxK kernel, the result is undefined for pixels closer than K pixels to the border of the image! Options:! Warp around! Expand/Pad! Crop! K lhm - 17

18

19 Kernel: 1/9 1/9 1/9 1/9 1/9 1/9 1/9 1/9 1/9

20 Kernel:

21 Kernel: 15 x 15 matrix of value 1/225

22 1-D: 2-D: Slight abuse of notations: We ignore the normalization constant such that

23 , σ = 5 Kernel:

24 Simple Averaging Gaussian Smoothing

25 Image Noise

26 Gaussian Smoothing to Remove Noise No smoothing σ = 2 σ = 4

27 Computational Issues for Filters: Separability and Moving Averages" When a KxK filter is equivalent to a Kx1 and a 1xK filter! Reduces number of multiplies from K 2 to 2K! For box filter, moving average makes cost independent of K! lhm - 27

28 Some Other (Nonlinear) Neighborhood Operators: Median Filter" Replace center pixel of KxK window with the mean value of all pixels in the window! Good for removing large noise spikes without blurring image! Original image! Gaussian filtered! Median filtered! lhm - 28

29 Some Other (Nonlinear) Neighborhood Operators: Bilateral Filter for Edge-Preserving Smoothing" 4-Jan-2011 ME/CS 132 lhm - 29

30 Some Other (Nonlinear) Neighborhood Operators: Bilateral Filter for Edge-Preserving Smoothing" Original image! 15x15 box filter! 15x15 bilateral filter! lhm - 30

31 Other Useful Neighborhood Operators: Feature Detectors (Recall Earlier Lecture)" (Implementation note: moving averages)! lhm - 31

32 Other Useful Neighborhood Operators: Image Matching (Recall Earlier Lecture)" Sum squared difference (SSD) vs. sum absolute difference (SAD)! Normalization: why? Many approaches.! lhm - 32

33 Other Useful Neighborhood Operators" Morphology! Distance transform! Connected components! lhm - 33

34 Edge Detection: What is an Edge?" Local maxima of rate of change of intensity! lhm - 34

35 Origin of Edges" surface normal discontinuity depth discontinuity surface color discontinuity illumination discontinuity Many factors! Sometimes care which factor applies; sometimes can determine that! lhm - 35

36 Image Derivatives: 1-D Case" We want to compute, at each pixel (x,y) the derivatives:! In the discrete case we could take the difference between the left and right pixels:! Convolution of the image by! Problem: Increases noise! Difference between Actual image values True difference (derivative) Twice the amount of noise as in the original image lhm - 36

37 Edges in 1-D: Effects of Noise" Where is the edge?! lhm - 37

38 Solution: Smooth First" f (x) h(x) h f x (h f ) Where is the edge? Look for peaks in! x (h f ) lhm - 38

39 Derivative Property of Convolution: Saves One Step" x (h f ) = x h f f (x) x h x h f lhm - 39

40 Laplacian of Gaussian" f (x) 2 x 2 h 2 x h f 2 4-Jan-2011 Where is the edge? Zero crossing ME/CS 132 of bottom graph! lhm - 40

41 2-D Edge Detection Filters" Laplacian of Gaussian Gaussian derivative of Gaussian is the Laplacian operator: lhm - 41

42 DOG Approximation to LOG" lhm - 42

43 The Effect of Scale on Edge Detection" larger σ larger σ Scale space (Witkin 83) lhm - 43

44 θ Edge pixels are at local maxima of gradient magnitude Gradient computed by convolution with Gaussian derivatives Gradient direction is always perpendicular to edge direction

45 Applying the Gradient Magnitude Operator" I lhm - 45

46 Different Thresholds Applied" to the Gradient Magnitude" Additional steps:! Thresholding with hysteresis! Thinning (non-maximum suppression)! Linking! lhm - 46

47 Sampling an Image" Examples of GOOD sampling! lhm - 47

48 Undersampling and Aliasing" Examples of BAD sampling -> Aliasing! lhm - 48

49 Downsampling and Aliasing" Sample every other pixel to go left to right! lhm - 49

50 Downsampling with Smoothing (Gaussian, 1 Sigma)" Sample every other pixel to go left to right! lhm - 50

51 Downsampling with Smoothing (Gaussian, 1.4 Sigma)" Sample every other pixel to go left to right! How do you know how much to smooth, with what kind of filter? Fourier transform theory.! lhm - 51

52 Intuitive Introduction to Fourier Transforms" lhm - 52

53 Intuition for Two Dimensions" dot = A measure of image content at this frequency and orientation lhm - 53

54 Multiresolution Image Representations" Known as a Gaussian Pyramid [Burt and Adelson, 1983]" Gaussian Pyramids have all sorts of applications in computer vision! lhm - 54

55 Gaussian Pyramid Construction" filter kernel Repeat! Filter! Subsample! Until minimum resolution reached! can specify desired number of levels (e.g., 3-level pyramid)! The whole pyramid is only 4/3 the size of the original image! lhm - 55

56 Computer Vision - A Modern Approach Set: Pyramids and Texture Slides by D.A. Forsyth

57 Laplacian Pyramid Construction" Given input I Construct Gaussian pyramid I G 1,..,I G n Take the difference between consecutive levels:! I L k = I G k I G k-1 Image I L k is an approximation of the Laplacian at scale number k Laplacian is a band-pass filter: Both high frequencies (edges and noise) and low frequencies (slow variations of intensity across the image)! lhm - 57

58 Computer Vision - A Modern Approach Set: Pyramids and Texture Slides by D.A. Forsyth

59 Another Example" lhm - 59

60 Geometric Image Transformations (e.g. Barrel Distortion Correction, Rectification)" lhm - 60

61 Geometric Image Transformations (e.g. Barrel Distortion Correction, Rectification)" lhm - 61