Motion feature detection using steerable flow fields

Abstract

The

A voting-based computational framework for visual motion analysis and interpretation

“…An example of using basis set methods (as steerable flow fields) is the work of Fleet et al [34]. The results are good, but the use of a gradient descent solution is heavily dependent on initial conditions and parameters governing movement in the coefficient space.…”

Robustly Estimating Changes in Image Appearance

“…To handle complex motions with concise models, Black et al [13] proposed linear parameterized models learned from training examples using principal component analysis (PCA). Similarly, Fleet et al [18] modeled motion features, such as dynamic occlusion edges, and moving bars, using linear combinations of steerable basis flow fields. These linear models constrain the interpretation of image motion, and are used in the same way as translational or affine motion models.…”

Probabilistic detection and tracking of motion discontinuities

“…This use of bottom-up information, along with the temporal prediction of Condensation, allows us to effectively sample the most interesting portions of the state-space. To initialize new states and provide a distribution over their parameters from which to sample, we use a method described by Fleet et al [8] for detecting motion discontinuities. This approach uses a robust, gradient-based optical flow method with a linear parameterized motion model.…”

“…The temporal prior is defined in terms of the posterior distribution at the previous time instant and the temporal dynamics of the discontinuity model. The initialization prior incorporates predictions from a low-level motion feature detector [8]. The posterior distribution over the parameter space, conditioned on image measurements, is typically non-Gaussian.…”

“…At the low level, there is a set of dense motion discontinuity detectors that signal the presence of potential occlusion boundaries and give estimates of their orientations and velocities. In the example here, these detectors are based on approximate linear models of motion discontinuities, the coefficients of which can be estimated with robust optical flow techniques [8]. Neighborhoods of these filter outputs provide a prior distribution over model parameters that is sampled from when initializing non-linear models at the higher level.…”