Towards On-Line Intensity-Based Surface Recovery from Monocular Images

In this paper, we present a method that allows simultaneous surface reconstruction and camera localization from monocular images in static scenes. The novel aspect of the method is its independence from any explicit feature detection schemes. Instead, it uses method similar to intensitybased bundle adjustment. Thus, it is better suited for 3D reconstruction of weakly textured surfaces. A number of methods with similar functionality have already been described [4, 6]. All of these methods, however, rely on some kind of feature detection schemes, such as such as SIFT [2] features, FAST corner detection [8], and so on. The basic concept of our algorithm can be summarized as follows: In traditional bundle adjustment, coordinates of 3D points that are associated with feature points are recovered from a set of 2D feature position measurements. This approach will obviously work only if a feature detecting scheme can be used at all. In our case, we do not assume that robust feature extraction is possible, and thus we do not work with 2D positions, but with image intensities. Originally, our method was inspired by a stereo disparity tracking method developed by Ramey [7]. The generalization that we are suggesting leads to an optimization problem that corresponds to intensitybased bundle-adjustment that is restricted to two frames. Thus, our solution shares some characteristics with typical bundle-adjustment algorithms [1, 9]. Our method establishes a depth map of the region of interest within a template image that has been chosen by the user. That depth map is then a function Sd(u,v) mapping a k-dimensional parameter vector d together with image coordinates (u,v) ∈ R2 to a depth value λ ∈ R at the specified coordinate. Given intrinsic camera parameters, this depth map can actually be interpreted as a 3D surface. Let