His interests include robotics, computer vision, embedded systems, control and networking. He has worked on robotic systems for mining, aerial and underwater applications.
Talk 1: "Introduction to Robotic Vision"
This talk will motivate the use of vision as a sensor robotics by reviewing alternative robotic sensors and the biological prevalence of the sense of vision.
Tutorial 2: "Robotics and Machine Vision using MATLAB"
This tutorial will introduce the class to the fundamentals of robotic vision, the extraction of features (blobs, lines, points) from images which can be used for robot control. From finding points we will then look at how images are formed, how images vary as a function of camera pose, how we can reconstruct scene geometry from multiple camera views and how we can use image information to drive a robot.
If you have a recent version of MATLAB on your computer you will be able to follow along in a "hands on" fashion. The software is available for free from www.petercorke.com
Graduated from Ecole Nationale Supérieure de Mécanique, Nantes, in 1987;
received the Ph.D. degree and "Habilitation à Diriger des Recherches" in Computer Science
from the University of Rennes, France, in 1990 and 1998 respectively.
Since 1990, he is with Inria in Rennes, France. His current position is Senior Research Scientist ("Directeur de Recherche"). He is the head of the Lagadic group, which is a common group to Inria Rennes Bretagne Atlantique and Irisa.
Research interests include visual servoing, active vision, robotics, and computer vision.
Tutorial 1: "Visual Servoing"
Visual servoing consists in using the data provided by a vision sensor
for controlling the motions of a dynamic system.
The tutorial will present the main basic modeling aspects of visual
servoing, as well as an overview of its applications in visual manipulation, localization, navigation and medical robotics.
Professor of Robot Vision at the Department of Computing, Imperial College London, and lead the Robot Vision Research Group. His main research has been in computer vision and robotics, specifically in SLAM (Simultaneous Localisation and Mapping) using vision, with a particular emphasis on methods that work in real-time with a single camera which can be classed as Real-Time Structure from Motion (Single Camera SLAM or MonoSLAM, vision-based, monocular, bearing-only SLAM, in full 3D and without inertial sensing). This technology can provide low-cost real-time localisation for domestic robots, humanoid robots, wearable sensors, game interfaces and/or other devices.
Currently, his main interests are in improving the performance in terms of dynamics, scale and detail level of real-time visual localisation and mapping.
Tutorial 3: "Real-Time Visual SLAM"
In robotics, augmented reality and other applications, an
understanding of the 3D configuration of a scene and a sensor-carrying
body's movement through it must be obtained in real-time if
interaction is to be enabled --- this is the problem known as
Simultaneous Localisation and Mapping (SLAM). In this tutorial, we
will concentrate on practical methods for SLAM with a single camera as
the only sensor. We will start with some history on early visual SLAM methods, and look at now well-established methods based on point features, filtering and non-linear optimisation. We will then move on to cover recent algorithms which enable live dense estimation of scene surface geometry, using real-time parallel implementation of dense reconstruction techniques, for both standard and depth cameras. Finally we will look in depth at the real-time tracking component of SLAM systems, analysing how performance can be improved with dense, whole image tracking and potentially high frame-rate cameras.
Graduated from TUM (Technische Universitaet Muenchen) in 2009. There he took part of the Intelligent Autonomous Systems research group, Computer Science Department.
He has been an International Fellow Researcher at the Artificial Intelligence Center (AIC), Stanford Research Institute (SRI) International.
Research Scientist at Willow Garage from 2009 until 2012. He has been working on 3D perception and creating PCL.
Visiting Lecturer at Stanford University since 2009, teaching 3D image processing.
President and CEO at Open Perception, Inc since 2012.
Tutorial 4: "Open Source 3D Processing with the Point Cloud Library"
For the past few years, research fields such as computer vision and robotics have seen an explosion in activity, which led to a tremendous number of research papers produced. Understanding the scientific impact of each individual publication and attempting to reproduce the results by implementing the algorithms described in it, is therefore becoming a very time consuming task and takes away from a researcher's time to actually build upon previous work and advance the research field. This is an area where open source code projects have a significant impact and aid researchers and engineers to validate the claims modern scientific papers propose, as well as help bootstrap new code bases and experiments faster and more efficiently, without the need to reinvent the wheel every time.
In less than 1.5 years, the Point Cloud Library (PCL) project has become a de facto standard for open source 3D processing. PCL has gathered together more than 500 contributors and developers worldwide from over 100 research institutes, commercial companies, and government labs, all serving a community of many thousands of users, and working together to build a stable, robust, and efficient open source set of C++ libraries for 3D point cloud perception and visualization. In this tutorial we will go through a wide area of subjects starting with the basics of 3D processing, and ending up with advances topics such as object recognition, registration, segmentation, SLAM, etc -- with each topic having its own set of code examples built using PCL.
He received a Doctoral degree in Electrical Engineering (Universdad de Chile, 2010), and a M.Sc. in Applied Mathemathics (Ecole Normale Superieure (ENS) de Cachan, France, 2006).
He have been a research fellow at the NDRC, Kyutech, Japan (2009-2010), and an associated researcher at the Fraunhofer IPK-Institute, Berlin, Germany (2004-2005).
During 2011 he was a visiting foreing researcher at the Kyushu Institute of Technology, Fukuoka, Japan.
Currently he is Postdoctoral Fondecyt Researcher at the Advanced Mining Technoology Center (AMTC), Universidad de Chile.
He has a long experience in the development of statistical-based face and object detection systems. He has relevant publications in this area, one of them in the IEEE RAM.
Research interest: computer vision, machine learning, object detection and recognition, and face analysis.
Tutorial 5: "Strategies for Efficient Object Detection"
During the last years, there has been an explosion in object detecting research. The detection of objects is crucial for a robot in order to be able to understand the world and interact with it. Object detection basically corresponds to determining the position and scale of a particular set of object classes that may be present in an image. Common examples of this kind of problem are the detection of faces, pedestrians, cars and robots.
Object detection is a difficult problem to solve, mainly because of (1) the variability that the objects can present (deformations, poses, etc.), (2) the variability in the image formation and acquisition process (changes in illumination, rotations, translations, point of view, scale, etc.), (3) the variability in image background, (4) and the time constrains, in particular in the case of robot vision.
Several approaches have been proposed to solve this problem, however, the methods that have been able to work successfully in the case of robots, faces, pedestrians, and cars -- and in dynamic environments --, are those based on machine learning techniques. Under this paradigm, the object detection problem can be formulated as a two-class classification problem, with the classes being "object" and "non-object".
Starting from a brief introduction to the process of building an object detection system based on statistical classifier, in this tutorial some of the most successfully strategies for efficient object detection will be presented, including: (1) cascade classifiers and variants (nested/soft cascades), (2) boosted classifiers, (3) feature evaluation (Haar-like wavelets, LBP-based, HoG), and (4) efficient search. We will continue with an outline of some recent promising systems (such as part-based models, and efficient subwindow search). If time allows it, a method that deals simultaneously with the detection of multiple object-classes, by exploiting similarities of different the object-classes in an efficient way, will be presented.
Talk 2: Cooperative Simultaneous Localization and Mapping (SLAM) with Multiple Robots
A cooperative multi-robot team allows for the execution of
complex tasks that require simultaneous action or presence. In
performing many of these tasks, it is important for the team to obtain
a pose estimate for each robot, as well as a map of the operating
environment. In this talk, we will cover the formulation of the
multi-robot cooperative localization problem, and the multi-robot SLAM
problem. While the formulation of the multi-robot SLAM problem is very
similar to the single-robot case, a multi-robot system has the
advantages of inter-robot observations and communication. Vision-based
sensing is particularly useful for obtaining inter-robot observations
as robots are visually distinctive, thus facilitating the task of data
association. We will show an example of vision-based multi-robot SLAM,
and also examine the topics of decentralization, and inter-robot
Talk 3: Analysis of Multiple Views in Computer Vision
This talk will provide different
techniques that can be used to solve computer vision problems with
multiple views. The aims of the talk are: 1) to give an overview of
multi-view stereo system; 2) to show existing state-of-the-art
techniques in multiple view analysis; and 3) to present several
applications using color and X-ray imaging in which multiple view
analysis play an important role.
In this talk I will describe some of our work applying machine learning techniques to the computer vision area. In particular, I will describe applications at different levels of visual abstraction, such as algorithms able to learn new low level image descriptors, or algorithms able to recognize indoor scenes exploring objects configurations. Then, I will describe some of our recent work using contextual information to improve object recognition and to construct meaningful dictionaries of visual words.
Talk 5: Sensing and Computer Vision Techniques for Driver Assistance and Autonomous Vehicles
This presentation will show the latest developments in sensors and computer vision techniques for road segmentation, lane keeping, pedestrian detection, and driver fatigue monitoring. All these aspects are essential in the implementation of driver assistance technology for intelligent vehicles, as well as the embodiment of autonomy in future driverless cars. The talk will discuss the existing approaches, present the main results, and summarize the key challenges that remain to be solved before full vehicle autonomy is achieved in real world's complex road environments.
In current visual SLAM (Simultaneous Localisation and Mapping) methods, point-like landmarks are used for representation on maps. In this talk, I will describe a visual SLAM system based on the use of what are called rigid-body 3D landmarks. A rigid-body 3D landmark represents the 6D pose of a rigid body in space (position and orientation), and its observation gives full-pose information about a bearing camera. Each rigid-body 3D landmark is created from a set of N point-like landmarks by collapsing 3N state components into 7 state components plus a set of parameters that describe the shape of the landmark. The proposed methodology is based on SURF descriptors, but it can be extended to any SLAM system based on the use of point-like landmarks, including those generated by laser measurement.
Vision brings information to a robot that it can use to reduce the uncertainty about its environment. Usually, visual sensors have a restricted field of view and embedded computers in a robot have limited processing power. Therefore, every instant, the robot must select a portion of the potentially available visual information to focus on. More specifically, it must decide to which direction it will point its camera(s) and to which regions of the image it will devote the most computational resources. Analogously, humans are able to understand a very complex world and to perform many tasks simultaneously in this world. Probably, this ability rely, among many others, in the ability of changing very quickly the attention from one point to another in order to gather the most relevant information from the world. Back to Robotics, the Active Vision problem can be defined as that of deciding in each instant the best region to focus on. This decision is made considering how the expected perception coming from processing each region will reduce the uncertainty on the world. Information-theoretic and task-oriented approaches will be presented and compared.
(Section under construction).
Demo Chair: Dr. Mauricio Correa, AMTC, Universidad de Chile.
The Advanced Mining Technology Center (AMTC) of the Universidad de Chile has different projects in order to foster research in robotics. The main motivation is to integrate and examine a wide range of technologies to solve the associated problems. In the demo session, different projects related to robotics that are currently being developed at the University will be presented, and the methodologies being used to solve the various problems in each project will be described.
The demo session will consist of four different demonstrations: a humanoid service robot, a soccer team of two-legged robots (NAO), an autonomous vehicle, and a hexacopter. A live demonstration of each this project will be given, and the participants will be able to discuss with the researchers working in each of the projects.