No feature-based vision system can work unless good features can be identified and tracked from frame to frame. Although tracking itself is by and large a solved problem, selecting features that can be tracked well and correspond to physical points in the world is still hard. We propose a feature selection criterion that is optimal by construction because it is based on how the tracker works, and a feature monitoring method that can detect occlusions, disocclusions, and features that do not correspond to points in the world. These methods are based on a new tracking algorithm that extends previous Newton-Raphson style search methods to work under affine image transformations. We test performance with several simulations and experiments.

This paper presents a method for registering images at different magnijications (scales) by treating the prob-lem not only as one of scaling the image coordinates but also as one inherently involving multiresolution informa-tion. While some existing methods for multiresolution reg-istration do consider the way the resolution (scale) affects the image, they often consider the image one scale at a time, using geometric properties within that scale. Others use multiscale information, usually to produce more robust re-sults, but only to register same-magni3cation images (e.g., stereo). A scale trace is the set of values that a single pixel takes on as magnijication decreases and the effective point spread function correspondingly increases. As such, scale traces capture information across multiple scales. This pa-per presents a method that uses correlation of scale traces through multiresolution images to find correspondences be-tween images of the same scene at differing magnijications. Because the method relies solely on the multiscale proper-ties of the images themselves and not on spatial correspon-dence, it is less sensitive to discrete sampling of potential scale factors. Sample results demonstrate that the method is able to accurately identify both the relative position and relative magnijication (scale) of the two images. 1

This paper proposes a new method for recovering nonrigid motion and structure of clouds under affine constraints using time-varying cloud images obtained from meteorological satellites. This problem is challenging not only due to the correspondence problem but also due to the lack of depth cues in the 2D cloud images (scaled orthographic projection). In this paper, affine motion is chosen as a suitable model for small local cloud motion. However, the cloud motion model can also be extended and not necessarily be affine. Analysis of intensity images is first used to find the candidates of correspondence for each point. Then affine motion model is utilized to find the structure (cloud-top height), motion parameters, and point correspondences. Extensive analysis has been done on defining appropriate constraints on depth that are necessary for achieving converge. Finally, experimental results on time-varying images of hurricane Luis, captured by GOES-9 satellite in geosynchronous orbit are...

This paper presents a unified approach to adaptive target detection and tracking. The unifying concept is “coherent motion energy”, a measure of the extent to which a single motion dominates local spatiotemporal structure. There are three major components to the approach. First, a multiresolution analysis of coherent motion energy is used to detect salient dynamic targets. Second, a robust affine transformation estimator is used to recover frame-to-frame target motion across regions of interest defined by coherent motion. Third, a method of template adaptation based on coherent motion weighted goodness of match is used to drive automatic template update. Empirical evaluation of the approach shows the contribution of the various components and documents strong performance of the integrated whole. 1.