Title: Reliable Real-Time SLAM within Search and Rescue Scenarios

Authors: Cristiano Premebida, Gonçalo Monteiro, Paulo Peixoto, and Urbano Nunes

Speaker: Urbano Nunes

Affiliation: Institute for Systems and Robotics, University of Coimbra, Portugal

Abstract:
Intelligent vehicles need reliable information about the environment in order to operate with total safety. In this context, the robust detection of dynamic obstacles plays an
important role. Several sensorial modalities have been exploited to address this issue. An overview of most recent developments in this field with special emphasis on methods based on laser range scanners and vision will be presented.
In this paper we describe a multi-modal system for Multi-Target Detection and Tracking, combining range data and vision. A classifier based on Gaussian Mixture Models (GMM) is used to distinct the obstacles categories (pedestrians, tree trunks/posts, and cars) in a semi-structured outdoor environment based on 2D range data. The overall system robustness is increased by using a vision module to improve the detection and classification results. A learning algorithm, combining Haar-like features and AdaBoost, is used to differentiate from different possible targets categories and to increase the system degree of confidence on the target classification task