Unsupervised classification of dynamic obstacles in urban environments

Authors:
Roman Katz;Juan Nieto;Eduardo Nebot
Affiliations:
Australian Centre for Field Robotics, University of Sydney, Sydney, New South Wales 2006, Australia;Australian Centre for Field Robotics, University of Sydney, Sydney, New South Wales 2006, Australia;Australian Centre for Field Robotics, University of Sydney, Sydney, New South Wales 2006, Australia
Venue:
Journal of Field Robotics
Year:
2010

Citing 0
Cited 4

2012 Special Issue: Using GNG to improve 3D feature extraction-Application to 6DoF egomotion

Neural Networks
GPGPU implementation of growing neural gas: Application to 3D scene reconstruction

Journal of Parallel and Distributed Computing
Unsupervised online learning for long-term autonomy

International Journal of Robotics Research
A robust and fast method for 6DoF motion estimation from generalized 3D data

Autonomous Robots

Quantified Score

Hi-index	0.00

Visualization

Abstract

This paper presents a solution to the problem of unsupervised classification of dynamic obstacles in urban environments. A track-based model is introduced for the integration of two-dimensional laser and vision information that provides a robust spatiotemporal synthesis of the sensed moving obstacles and forms the basis for suitable algorithms to perform unsupervised classification by clustering. This work presents various contributions in order to achieve accurate and efficient performance, initially using laser tracks for classification and then incorporating visual tracks to the model. A procedure is proposed for accurate unsupervised classification of dynamic obstacles using a laser stamp representation of the tracks. Laser data are then integrated with visual information through a single-instance visual stamp representation, which is finally extended using a multiple-instance framework to robustly deal with challenges associated with perception in real-world scenarios. The proposed algorithms are extensively validated with a simulated environment. Experiments with a research vehicle in an urban environment demonstrate the performance of the approach with real data. The experimental results reach an accuracy of more than 92% for obstacle classification, finding the clusters that correspond to the main obstacle classes in the data. © 2010 Wiley Periodicals, Inc.