Segregation of moving objects using elastic matching

Reviewer: Creed F Jones

Tracking objects in video sequences requires accurate correspondence of image features that comprise the object or objects under consideration. There is extensive literature on matching various features between frames: points, patches, line segments, and curves. The larger and richer features, such as patches and curves, exhibit less confusion but are more sensitive to interframe variations. Curves, in particular, can be used for high-performance tracking and image segregation if they can be matched in the presence of typical levels of variation in shape, differences in color and shading, and occlusion. This paper describes a proposed technique for tracking curves in a sequence of images, based on the alignment curve approach [1]. Minimization of an energy functional on the alignment curve between two subject curves will identify the optimal alignment; in its basic form, the functional penalizes both stretching and bending of either curve with respect to the other. The authors analyze the performance of this basic functional on real-world image sequences, introducing enhancements to address several shortcomings. First, significant changes in the curves?called transitions?may appear in a curve during the sequence, leading to improper matching via undue stretching. This effect is suppressed by penalizing matches that require the introduction of linear segments. Second, transitions in larger curves are handled explicitly, via analysis and matching of the combinations present. In addition, an epipolar constraint is applied, based on the assumption that object curves will approximate linear motion over small portions of the image sequence. For a stationary camera, this reduces to the assumption of a vanishing point from which the object curves will appear to emanate. This limits the possible locations of matching curves in subsequent images. Finally, color (or intensity, in the monochrome case) information is used to further reduce ambiguity in curve matching. The authors match color information by analysis of binned hue, saturation, value (HSV) histograms of regions on either side of the curves; the comparison is tolerant of image occlusion. Through the use of these successive enhancements, the authors report matching accuracies that approach 90 percent. This is a good level of performance for this problem. However, the prime motivation for this study was figure-ground segregation. The reported results show that the method provides an accurate and rich set of corresponding curves, with related trajectories, that can be identified as the foreground object. This is a well-written paper and that represents a good summary of the steps required to deliver a useful method for curve tracking and object segregation. The authors mention the possibility of additional benefits from performing multiframe curve tracking instead of all curve matching on pairs of images. This would be a very interesting follow-up to the current study. Online Computing Reviews Service