Spatial and Temporal Stability of Hierarchically Segmented Regions for use as Conjugate Pairs in the Estimation of Depth and Tracking

J.C. Woods (UK)


Identification, depth estimation, object tracking, robotic vision, image segmentation, region merging, temporal consistency, real-time.


Techniques for the selection of matching conjugate pair points used in the estimation of depth often employ interest operators, where areas of the image containing high variance are selected. Assuming a sufficient number of points can be isolated, features are compared in the two images using correlation techniques to verify a match. Features selected for matching often include corners and edges. One of the principle disadvantages of matching edges is the depth can have no meaning along an occlusion boundary where a foreground object meets the background. This will typically produce a distinct feature, but the depth could be anywhere between the near and far objects. This work proposes the use of simple segmentation techniques to produce robust feature point pairs for tracking and active vision systems, and identifies the difficulties encountered. A number of advantageous factors are identified, including the avoidance of the occlusion boundary problem, hierarchical recognition of regions during the transition down the segmentation pipeline, and abundance of statistical catagorisation.

Important Links:

Go Back