270 likes | 362 Views
Image Registration. Vlad Magdin York University December 7, 2007. Motivation. Two sources for difference between frames Object motion Camera motion Motion is a powerful cue for detection Can detect moving objects if camera motion is eliminated via frame differencing. Motivation.
E N D
Image Registration • Vlad Magdin • York University • December 7, 2007
Motivation • Two sources for difference between frames • Object motion • Camera motion • Motion is a powerful cue for detection • Can detect moving objects if camera motion is eliminated via frame differencing
Motivation Video Frame F(t) Video Frame F(t+1)
Without Registration Frame Difference {F(t+1)-F(t)}
With Registration Frame Difference {F(t+1)-F(t)}
Registration Approaches • Global - directly compute transformation parameters • e.g. Direct Optimization • Feature-based - compute transformation parameters from a set of matching points • e.g. Optic Flow (Lucas Kanade)
Implemented Methods • Search methods: • Direct Optimization • SIFT-based Feature Registration • Lucas-Kanade Optic Flow • Transformation types • Affine • Projective • 2nd Order
Transformation Types • Affine transformation - 6 parameters
Transformation Types • Projective transformation - 8 parameters
Transformation Types • 2nd Order transformation - 8 parameters
Transformation of Choice • 2nd Order transformation - 8 parameters
Error Measure Frame Difference {F(t+1)-F(t)}
Optimization Approach • Algorithm • Given two video frames • Use an optimization routine to iteratively compute a transformation matrix based on an error measure (SSD)
SIFT-Based Method • SIFT (Scale Invariant Feature Transform, Lowe 2004) • Invariant to scale, rotation, illumination change • Feature Selection Algorithm • Laplacian Pyramid - compute a multi-resolution representation of an image • Keypoints - locate extrema in the Laplacian Pyramid • SIFT Descriptors - compute histograms of gradients around each keypoint
Optic Flow Approach • Algorithm • Estimate displacement vectors of pixels from spatial and temporal image gradients • Compute a transformation matrix by thinking of each pixel as a feature
Optic Flow Approach • Algorithm • Estimate displacement vectors of windows from spatial and temporal image gradients • Compute a transformation matrix by thinking of each window as a feature
Optic Flow Approach • For each window, compute: • Ix, Iy : spatial gradientsIt : temporal gradient • v = Ge
Optic Flow Approach • For each window, compute: • Not all windows are kept: • Windows that lead to ill-posed solutions to “v = Ge” (2nd eigenvalue of G is too small) • Windows that lead to large errors • Windows that fall within boundaries of better windows
Qualitative Comparison Video Frame F(t) Video Frame F(t+1)
Final Comments • Optimization Approach+ relatively good performance- poor speed- initial parameter guess is important • Feature-based Approach+ (relatively) fast- can sometimes select moving objects as features • Optic Flow Approach+ (relatively) fast+ best performance