Segmentación y parametrización de líneas en datos láser 2D basado en agrupamiento por desplazamiento de media

Abstract

This paper presents a robust algorithm that is implemented for segmentation and characterization of traces obtained through a sweep process performed by a laser sensor. The process yields polar parameters that define segments of straight lines, which describe the scanning environment. A Mean-Shift Clustering strategy that uses the average of laser scanning points set in an orient-able ellipse is proposed as an estimate of the density gradient of points within the window. Grouping is achieved by sliding this ellipse into areas of space where the density of points is high, and it is redirected towards the direction of greater data dispersion. Each grouped set of points is processed by a modified RANSAC (Random Sample and Consensus) algorithm. This method involves the construction of model assumptions from minimal data subsets chosen at random and evaluates their validity supported by the whole of data, while the associated probability densities are updated. The parameters of the detected segments are estimated by a TLS (Total Least Squares) regression, which minimizes the sum of squared differences between the function and the data. The algorithm is evaluated in indoor environments using mobile robot platform Pioneer 3DX (equipped with a SICK laser sensor), obtaining satisfactory results in terms of compactness and error parameters of the lines detected. Likewise, tests were conducted using simulated data with constant density (where the classic MSC algorithm performs poorly), achieving significant improvements in segmentation and line parameterization.

Keywords

• principal component analysis
• regression analysis
• measurement by laser beam
• clustering methods