CS-465 Computer Vision

Transcription

1 CS-465 Computer Vision Nazar Khan PUCIT 9. Optic Flow

2 Optic Flow Nazar Khan Computer Vision 2 / 25

3 Optic Flow Nazar Khan Computer Vision 3 / 25

4 Optic Flow Where does pixel (x, y) in frame z move to in frame z + 1? [ ] [ ] [ ] x x u y = + y v We want to find the displacement vector (u, v) T pixel. for every Input: image sequence I (x, y, z), where (x, y) specifies the location and z denotes time/frame number Goal: displacement vector field of the image structures: optic flow (u(x, y, z), v(x, y, z)) Such correspondence problems are key problems in computer vision. Nazar Khan Computer Vision 4 / 25

5 Grey Value Constancy Assumption Corresponding pixels should have the same grey value. Thus, the optic flow between frame z and z +1 should satisfy I (x + u, y + v, z + 1) = I (x, y, z) = I (x, y, z) + ui x (x, y, z) + vi y (x, y, z) + 1I z (x, y, z) I (x, y, z) = I x (x, y, z)u + I y (x, y, z)v + I z (x, y, z) 0 assuming (u, v) is a small displacement. Linearized optic flow constraint (OFC) I x u + I y v + I z = 0 where location (x, y, z) is implied. Nazar Khan Computer Vision 5 / 25

6 How good are the assumptions? We have made two assumptions 1. Gray value constancy 2. Small displacements (since we use first-order Taylor series approximation) Both assumptions are (almost) true in surprisingly many scenarios. 1. Gray values do not change much between consecutive 1 frames. 2. Objects do not move too much between consecutive frames. For large displacements, image pyramid can be used. 1 For a video recorded at 25 frames per second (fps), consecutive frames are only 1 seconds apart. 24 Nazar Khan Computer Vision 6 / 25

7 Aperture Problem Complete Flow Normal Flow No Flow When seen through an aperture, true movement cannot be determined. Only the component of movement normal to edge direction can be determined. Nazar Khan Computer Vision 7 / 25

8 Normal Flow The OFC is one equation in two unknowns (infinite solutions). Can be written as [ ] T u I + I v z = 0 Adding any flow component orthogonal to image gradient does not affect the OFC. ([ ] ) T [ ] T u +k I u I + I v z = I +k I T I v = = 0 [ ] T u I + I v z } {{ } + I z 0 Nazar Khan Computer Vision 8 / 25

9 Normal Flow I I (u, v) (u n, v n ) I I [ un v n ] ([ ] u = I ) I v I I = I z I ( [ u I I v] I + Iz = 0 ) = 1 [ ] Ix I z Ix 2 + Iy 2 I y I z Only the component of flow in the direction of the gradient I can be computed. Since gradient is normal to the edge direction, this flow vector is called the normal flow. To compute a better estimate of optic flow, we need to make some assumptions. Nazar Khan Computer Vision 9 / 25

10 Local Optic Flow Method of Lucas & Kanade Lucas & Kanade make the following assumption: Pixels around (i, j) all have the same displacement (u, v). For 3 3 neighbourhoods, this gives 9 OFCs all having the same 2 unknowns (u, v). The optimal unknown displacement minimizes the sum-squared-error E(u, v) = 1 2 N ij (I x u + I y v + I z ) 2 Nazar Khan Computer Vision 10 / 25

11 Local Optic Flow Method of Lucas & Kanade Setting E u E = 0 and v [ N ij I 2 x = 0 yields a linear system N ij I x I y N ij I x I y N ij I 2 y ] [u ] = v [ ] N ij I x I z N ij I y I z Replacing the sums by Gaussian averaging yields [ Gρ Ix 2 ] [ ] [ ] G ρ I x I y u Gρ I G ρ I x I y G ρ Iy 2 v = x I z G ρ I y I z }{{} A (1) Notice the re-appearance of the structure tensor which now serves as the system matrix. Previously, we used it for corner detection. Flow vector can be found if rank(a) = 2. Nazar Khan Computer Vision 11 / 25

12 Local Optic Flow Method of Lucas & Kanade If rank(a) = 0, no gradients exist in the neighbourhood. So no optic flow can be computed. If rank(a) = 1, gradient vectors over all pixels in the neighbourhood are identical. Only normal flow can be computed. [ ] un = 1 [ ] Ix I z v n Ix 2 + Iy 2 I y I z To save computations, avoid computing rank. trace(a) = A 11 + A 22 0 = rank(a) = 0 trace(a) 0 and det(a) = A 11 A 22 A = rank(a) = 1 Nazar Khan Computer Vision 12 / 25

13 Lucas & Kanade Algorithm Input: Frames I 1 and I 2. Parameters: 1) Noise smoothing scale σ, 2) Gradient smoothing scale ρ, 3) Thresholds τ trace and τ det. 1. Compute Gaussian derivatives at noise smoothing scale σ I x = Gσ x I 1 and I y = Gσ y I 1 2. Compute temporal derivative I z = I 2 I Compute the products I 2 x I 2 y I x I y I x I z and I y I z 4. Smooth the products at gradient smoothing scale ρ G ρ I 2 x G ρ I 2 y G ρ I x I y G ρ I x I z and G ρ I y I z and construct the linear system in (1) at every pixel. Nazar Khan Computer Vision 13 / 25

14 Lucas & Kanade Algorithm 5. For every pixel, solve the linear system conditioned on the rank. if A 11 + A 22 < τ trace rank(a)=0 so no flow else if A 11 A 22 A 2 12 < τ det rank(a)=1 so normal flow else rank(a)=2 so complete optic flow Nazar Khan Computer Vision 14 / 25

15 Visualising Displacement Vectors The HSV Color Space Each color is represented by 3 values 1. Hue or shade as an angle from 0 to Saturation or strength of the color 3. Value or brightness Figure: The HSV color space. Taken from /05/munsell-hue-circle.html. Nazar Khan Computer Vision 15 / 25

16 Visualising Displacement Vectors Figure: Vector angle represented by hue/shade of color and vector magnitude represented by the saturation/strength of color. HSV color space is useful for such a mapping. H(x, y) = θ(x, y), S(x, y) = u(x, y) 2 + v(x, y) 2 and V (x, y) = constant. Nazar Khan Computer Vision 16 / 25

17 Lucas & Kanade Figure: Left to right: frame 1, frame 2, flow classification and false color visualization of optic flow vectors. For flow classification: white = optic flow, gray = normal flow and black = no flow. Integration scale was ρ = 1. Author: N. Khan (2015) Nazar Khan Computer Vision 17 / 25

18 Lucas & Kanade Figure: Left to right: frame 1, frame 2, flow classification and false color visualization of optic flow vectors. For flow classification: white = optic flow, gray = normal flow and black = no flow. Increasing the integration scale ρ to 4 fills up pixels with no flow using values from neighbouring pixels having normal or complete optic flow. Author: N. Khan (2015) Nazar Khan Computer Vision 18 / 25

19 Lucas & Kanade Summary Advantages Simple and fast method. Requires only two frames (low memory requirements). Good value for money: results often superior to more complicated approaches. Disadvantages Problems at locations where the local constancy assumption is violated: flow discontinuities and non-translatory motion (e.g. rotation). Local method that does not compute the flow field at all locations. Next we study a global method that produces dense flow fields (i.e., at every pixel). Nazar Khan Computer Vision 19 / 25

20 Variational Method of Horn & Schunck At some given time z the optic flow field is determined as minimising the function (u(x, y), v(x, y)) T of the energy functional E(u, v) = 1 (I x u + I y v + I z ) 2 +α ( u 2 + v 2) 2 }{{}}{{} x,y data term smoothness term Has a unique solution that depends continuously on the image data. Global method since optic flow at (x, y) depends on all pixels in both frames. Notation Alert! u and v are 2D arrays of the same size as the frame but they are also used to refer to a pixel location. Nazar Khan Computer Vision 20 / 25

21 Variational Method of Horn & Schunck Regularisation parameter α > 0 determines smoothness of the flow field. α 0 yields the normal flow. The larger the value of α, the smoother the flow field. Dense flow fields due to filling-in effect: At locations, where no reliable flow estimation is possible (small I ), the smoothness term dominates over the data term. This propagates data from the neighbourhood. No additional threshold parameters necessary. Nazar Khan Computer Vision 21 / 25

22 Functionals and Calculus of Variations Since u is a function, E(u, v) is a function of a function. A function of a function is also called a functional. Normal calculus can optimize functions f (x) by requiring d dx f x = 0. Functionals are optimized via calculus of variations. Optimizer of an energy functional E(u, v) = x,y F (x, y, u, v, u x, u y, v x, v y ) must satisfy the so-called Euler-Lagrange equations x F ux + y F uy F u = 0 x F vx + y F vy F v = 0 with some boundary conditions. Nazar Khan Computer Vision 22 / 25

23 Functionals and Calculus of Variations For our energy functional E(u, v), F = 1 2 (I xu + I y v + I z ) 2 + α 2 with partial derivatives F u = I x (I x u + I y v + I z ) F v = I y (I x u + I y v + I z ) ( u 2 x + u 2 y + v 2 x + v 2 y ) F ux F uy F vx F vy = αu x = αu y = αv x = αv y Nazar Khan Computer Vision 23 / 25

24 Functionals and Calculus of Variations So the Euler-Lagrange equations can be written as α(u xx + u yy ) I x (I x u + I y v + I z ) = 0 α(v xx + v yy ) I y (I x u + I y v + I z ) = 0 After writing out the first and second order derivatives α h 2 (u j u i ) I xi (I xi u i + I yi v i + I zi ) = 0 j N i α h 2 (v j v i ) I yi (I xi u i + I yi v i + I zi ) = 0 j N i where h is the grid size (usually 1). Two equations for every pixel. Can be written as a sparse but very large linear system Bx = d. Size of B will be 69GB for a image! Nazar Khan Computer Vision 24 / 25

25 Functionals and Calculus of Variations All of the above boils down to a very simple iterative scheme α ( ) h 2 j N i u (k) j I xi I yi v (k) i + I zi u (k+1) i = v (k+1) i = α h 2 N i + I 2 xi α ( ) h 2 j N i v (k) j I yi I xi u (k) i + I zi α h 2 N i + I 2 xi with k = 0, 1, 2,... and an arbitrary initialisation (e.g. zero vector). Nazar Khan Computer Vision 25 / 25