Weak-edge detection for motion capture using a two-layer active contour model

Image edges, especially in the extracted silhouette, offered satisfied positional information for markerless vision-based motion capture. However, traditional bottom-up edge detection methods (such as canny operator) are powerless for edge detection with small luminance difference, especially when a self-occlusion happened. In this paper, we proposed a top-down edge detection method which can be viewed as a two-layer hierarchical active contour model. Given an initial curve, we embed the motion parameters into the energy function. In the first layer, we find the best global Euclidean (affine) transformation based on the Gradient Vector Flow (GVF) field. And in the second layer, we use the GVF snakes to find the local best movements of individual curve points for non-rigid deformation. The result of the first layer provided the best initial curve to the second layer. The promising experimental results for real and synthetic images reveal the effectiveness of our approach.