In short, feature-based methods would likely be preferable to scan matching if they were able to offer the same robustness and broad applicability to different environments. Because of the advantages of feature-based SLAM solutions, the extraction of features from LIDAR data has been extensively explored.As pointed out before, STA-9090 classical feature detectors rely on prior knowledge of environments. Researchers reconstruct an environment with lines and curves based on previously collected environmental information and assumptions, then features are extracted from the lines and curves. For example, a specific line fitting algorithms [23,24] will be carefully tuned to the characteristics of an environment (mainly on the contour size and the error threshold).
After lines are re-constructed, the features with stable positions, for example, midpoints of lines or intersection points of lines, will be extracted as features. These classical feature detectors are easy to implement and have excellent performance Inhibitors,Modulators,Libraries in correspondingly target environments, but they are not widely applicable to other types of environment.Recent work on feature detectors focuses on addressing these problems. The curvature estimation based feature extractor [2] tries to fit various environments with curves [2]. The B-spline based extractor [25] represents the world with B-splines; although it is generally applicable, the segmentation of laser data, the selection of control points, and the feature representation in the data association process are still areas of active research [25].
Zlot and Bosse [26] propose a number of heuristics for identifying stable keypoints in LIDAR data. Their methods begin with clustering connected Inhibitors,Modulators,Libraries components and then either (1) computing the centroids of each segment, (2) computing the curvature of each segment, or (3) iteratively Inhibitors,Modulators,Libraries Inhibitors,Modulators,Libraries computing a locally-weighted mean position until it converges. Our approach replaces these three mechanisms with a single method. Zlot and Bosse additionally investigate descriptor algorithms, which significantly simplify data association tasks. These descriptor methods could also be applied to our detector.In this paper, we describe a general purpose feature detection algorithm that Carfilzomib generates highly repeatable and stable features in virtually any environment, as shown in Figure 1.
Our approach builds upon methods used in image processing, where the need for selleckbio robust feature detectors has driven the development of a wide variety of approaches. In particular, we show how the Kanade-Tomasi [27], a variant of the Harris corner detector [28] can be applied to LIDAR data. At the same time, we also studied the characteristics of the features extracted using our method, including uncertainties and feature descriptors.Figure 1.Multi-scale feature extraction from LIDAR data.