Object Detection in Point Clouds Using Conformal Geometric Algebra

Aksel Sveier,Olav Egeland,Lars Tingelstad,Adam Leon Kleppe

doi:10.1007/s00006-017-0759-1

Aksel Sveier, Olav Egeland + Show 2 more

Open Access

https://doi.org/10.1007/s00006-017-0759-1

Copy DOI

Abstract

This paper presents an approach for detecting primitive geometric objects in point clouds captured from 3D cameras. Primitive objects are objects that are well defined with parameters and mathematical relations, such as lines, spheres and ellipsoids. RANSAC, a robust parameter estimator that classifies and neglects outliers, is used for object detection. The primitives considered are modeled, filtered and fitted using the conformal model of geometric algebra. Methods for detecting planes, spheres and cylinders are suggested. Least squares fitting of spheres and planes to point data are done analytically with conformal geometric algebra, while a cylinder is fitted by defining a nonlinear cost function which is optimized using a nonlinear least squares solver. Furthermore, the suggested object detection scheme is combined with an octree sampling strategy that results in fast detection of multiple primitive objects in point clouds.

Highlights

There has been an increase in the use of 3D cameras in robotic vision applications due to the availability of commercial products with high accuracy
The depth information from 3D cameras can be represented as point clouds, which is a set of points in Euclidean space given by the x, y, z coordinates for each point
Noise/outliers were added to the Object Detection in Point Clouds Using conformal geometric algebra (CGA) Table 1

Summary

Introduction

There has been an increase in the use of 3D cameras in robotic vision applications due to the availability of commercial products with high accuracy. The advantage of 3D cameras compared to 2D cameras is the additional depth information, which can provide information about size and position of objects in a scene, where a scene is the environment captured by the camera. The depth information from 3D cameras can be represented as point clouds, which is a set of points in Euclidean space given by the x, y, z coordinates for each point. Object detection in point clouds can be difficult due to noise, outliers and complexity in the data.

Methods

Results

Conclusion