Contours and Optical Flow: Cues for Capturing Human Motion in Videos Thomas Brox Computer Vision and Pattern Recognition Group University of Bonn Research partially funded by the German Research Foundation (DFG) Human pose tracking from video 1. Tracking of markers attached to the body + Designed to be easy to track Reliable and fast tracking Thomas Brox University of Bonn – Accuracy limited by number of markers – People may feel uncomfortable • Introduction • Segmentation • Optic Flow • Summary 2. Tracking features that naturally appear in the images • Patches (e.g. KLT, SIFT, etc.) • Contour/Silhouette • Optic flow How to extract these features reliably from the images 2 Contour and optic flow based human tracking Thomas Brox University of Bonn • Introduction • Segmentation • Optic Flow • Summary Joint work with Bodo Rosenhahn 3 Part I Object Contour Extraction Object contour extraction • Find two regions: object & background Thomas Brox University of Bonn • Introduction • Segmentation • Often: Static background background subtraction • Optimality criteria here: – Strong similarity within regions – Small boundary • Optic Flow • Summary 5 • Bayesian approach: Level set representation of contours (Dervieux-Thomasset 1979, Osher-Sethian 1988) • Introduce embedding function • Contour C represented as zero-level line of Thomas Brox University of Bonn • Introduction • Segmentation • Optic Flow • Summary 6 Courtesy of Daniel Cremers Region-based active contours (Chan-Vese 2001, Paragios-Deriche 2002) • Minimize negative logarithm: Thomas Brox University of Bonn H(x) • Introduction • Segmentation • Optic Flow • Gradient descent: • Summary H’(x) plus update of p1 and p2 7 Region statistics • 7 channels: Thomas Brox – 3 color channels (CIELAB) – 4 texture channels University of Bonn • Channels assumed to be independent • Introduction • Segmentation • Optic Flow • Summary 8 • Probability densities pij approximated by Gaussians Texture • Usually modeled by Gabor filters (Gabor 1946) Thomas Brox University of Bonn • Includes 1. Magnitude 2. Orientation 3. Scale • Introduction • Segmentation • Optic Flow • Summary • High redundancy • Sparse alternative representation feasible • Nonlinear structure tensor (Brox et al. 2006) • Region based local scale measure (Brox-Weickert 2004) 9 Sparse texture features Thomas Brox University of Bonn • Introduction • Segmentation • Optic Flow • Summary Gabor filter bank 10 Sparse representation Examples for contour extraction Thomas Brox University of Bonn • Introduction • Segmentation • Optic Flow • Summary 11 Local region statistics • Object and background usually not homogeneous Thomas Brox University of Bonn • Idea: assume them to be locally homogeneous • Introduction • Segmentation • Optic Flow • Summary 12 • Probability densities estimated by local Gaussians Introducing a shape prior • Idea: object model can serve as 3-D shape prior Thomas Brox Constrains the segmentation, unwanted solutions discarded University of Bonn • Bayesian formula: • Introduction • Segmentation • Optic Flow • Summary • Pose parameters of model unknown Two variables: contour and pose 13 Joint optimization • Simultaneously optimize contour and pose: Thomas Brox University of Bonn conventional segmentation part shape+pose constraint • Introduction • Segmentation • Optic Flow • Summary • Iterative alternating scheme: – Update contour – Update pose parameters • Related works: 2-D shape priors (Leventon et al. 2000, Cremers et al. 2002, Rousson-Paragios 2002) 14 Part II Optic Flow Optic flow based tracking Thomas Brox University of Bonn • Introduction • Segmentation Image 1 and 2, estimate flow in between Given pose at Image 1 Pose change due to optic flow Estimated pose at Image 2 • Optic Flow • Summary 16 Tracking example Thomas Brox University of Bonn • Introduction • Segmentation • Optic Flow • Summary 17 How to compute the optic flow? Thomas Brox University of Bonn • Introduction • Given: two images I(x,y,t) and I(x,y,t+1) in a sequence • Goal: displacement vector field (u,v) between these images • Variational approach: • Segmentation • Optic Flow • Summary 18 (Horn-Schunck 1981) Enhanced model (Brox et al. 2004, Papenberg et al. 2006) Original Horn-Schunck: Thomas Brox University of Bonn Robust smoothness term (Cohen 1993, Schnörr 1994) Robust data term (Black-Anandan 1996, Mémin-Pérez 1996) Gradient constancy • Introduction (Brox et al. 2004) Non-linearized constancy (Nagel-Enkelmann 1986, Alvarez et al. 2000) • Segmentation • Optic Flow • Summary 19 Spatiotemporal smoothness Final optic flow model: (Nagel 1990) Impact of each improvement Thomas Brox University of Bonn • Introduction Correct result • Segmentation • Optic Flow • Summary Nonlinear Gradient Horn-Schunck constancy constancy Spatiotemporal Robust data smoothness term Robust smoothness 8 Horn-Schunck Robust smoothness 6 Robust data term 4 20 1.78 2.44 3.5 5.97 6.36 0 Nonlinear constancy 7.17 2 Gradient constancy Spatio-temporal smoothness Accurate and robust optic flow computation Technique Nagel Thomas Brox University of Bonn • Introduction • Segmentation • Optic Flow • Summary 21 AAE 10.22° Uras et al. 8.94° Alvarez et al. 5.53° Mémin-Pérez 4.69° Brox et al. (Noisy) 4.49° Bruhn et al. 4.17° Brox et al. 1.78° Contour and optic flow based human tracking Thomas Brox University of Bonn • Introduction • Segmentation • Optic Flow • Summary Joint work with Bodo Rosenhahn 22 Summary • Contours and optic flow can be reliable features for pose tracking Thomas Brox University of Bonn • Texture, local statistics, and a shape prior are important for general contour based human motion tracking • High-end optic flow helps in case of fast motion • Introduction • Segmentation • Optic Flow What’s next? • Summary • Real-time performance • Automatic pose initialization • Prior knowledge about joint angle configurations 23 Outlook Thomas Brox University of Bonn • Introduction • Segmentation • Optic Flow • Summary Joint work with Bodo Rosenhahn 24 Backup: nonlinear structure tensor • Texture orientation can be measured with the structure tensor (second moment matrix) (Förstner-Gülch 1987, Rao-Schunck 1991, Bigün et al. 1991) Thomas Brox University of Bonn • Introduction • Gaussian smoothing nonlinear diffusion • Segmentation • Optic Flow • Summary Input image 25 Linear structure tensor Nonlinear structure tensor Backup: region based local scale measure • Estimate regions, measure their size Thomas Brox University of Bonn • Nonlinear diffusion: TV flow (Andreu et al. 2001) Input image • Introduction • Segmentation • Optic Flow • Summary • Tends to yield piecewise constant images regions • Local evolution speed inversely proportional to size of region (Steidl et al. 2004) local scale measure 26 Local scale