Fast landmark-based registration via deterministic and efficient processing, some preliminary results

Abstract

AbstractΩ Project sponsored, in part, by the Department of the Navy, Office of Naval Research. Approach Based on Subgraph MatchingThe goal of this research is to pursue a technique that can perform view registration at rates approaching 10 Hz, without any user input for initial estimates. This performance goal is set to match the data rates of range cameras. It is also desired to have a method that computes the rigid transformation (translation plus rotation) in the presence of possible scale changes. Registration at these rates could permit sensor motion to be tracked in realtime. This would permit unconstrained movement of a sensor across a large scene.To achieve fast and deterministic processing, iterative, or random [6] approaches were avoided. Established methods do not typically separate the steps of determining corresponding points and determining the transform. This limits compute speed. In the new approach these steps have been kept separate, and are implemented in a non-iterative fashion. This is an important difference for the new approach. Another difference is that correspondence between the data sets is determined only for select feature points. This improves processing speed, but it does limit accuracy.Correspondence between the data sets is determined via graph matching. Graphs are formed using salient features in each range image. Graph matching is accomplished using the LeRP Algorithm [2] [4]. LeRP approximates a subgraph isomorphism via a deterministic procedure, based on the comparison of length-r paths. The LeRP algorithm yields a set of corresponding locations in the two input range images, from which the absolute orientation may be found via closed-form solution [9].A typical graph appears in Figure 1. The white segments designate an accurately matched subgraph for the scene. Approach -Salient FeaturesGraphs describe the salient features in each range image. For rapid detection, local, well isolated, peaks in the range data were used to form the image features. These are desirable because they are likely to remain in view with small shifts in scene.The significant percentage of black segments shown in Figure 1 demonstrates a lack of stability in feature detection. Stable and invariant feature detection -via fast, efficient, deterministic means -remains the greatest challenge in this new approach. While the lack of stability shown in the figure is undesirable, it was deemed an appropriate tradeoff in terms of processing speed. Other techniques employ more robust local features [10] but require more processing time. The approach taken here was to rely on the LeRP matching algorithm to determine appropriate correspondences, despite noisy features.Some reported techniques use invariant features that involve curvature, moments, or spherical harmonics [18]. These types of features react to jump discontinuities that may occur at a limb [19]. Such feature points were avoided in this new approach. Consider the occurrence of a limb where the line-of-sight of a sensor becomes tangent to a hillside. ...

Deterministic Surface Registration at 10Hz Based on Landmark Graphs With Prediction

“…Reports of accuracy and computational rate are given in Table 1, for both the landmark-graph approach and for a 'fast-ICP' method [10]. The fast-ICP method used a simple image difference, rather than point-by-point search for correspondence.…”

“…AVI-format video clips are available for download [14]. The clips contain images rendered during the testing discussed below.…”

“…Previous work by this author with landmarkgraphs restricted analyses to individual pairs of sensor images, not to image streams [10]. Stream processing is more appropriate for the continuous sensor movement.…”

“…With an image stream, prediction may be used, as in [5]. Results of the landmark-graph approach, including prediction, are superior to those previously reported by this author [10]. More information on LeRP, the graph matching technique is available in [9].…”