Advanced

An Approach to a Complete Recognition Module for Legged Robots in the RoboCup Domain

Motallebipour, Hossein LU (2010) In Lund : Department of Computer Science, Faculty of Engineering, LTH, Lund University EDA920 20041
Department of Computer Science
Abstract
A need for autonomous electro-mechanical machines performing tasks that can save human efforts and even be deployed in hazardous environments along with the research in the field of human mind and body has led to new fields often called AI Robotics. To provide a mean for stimulating students and researchers around the world for promotion of this new field, an annual world championship, Robocup, has been organized with start in 1997. During these competitions, the state-of-the-art technical achievements within Sensor Systems, Mechanics, Artificial Intelligence, Machine Learning and many other related fields are presented by the participants. The final goal is to create autonomous humanoids that can win a soccer match against humans by 2050.... (More)
A need for autonomous electro-mechanical machines performing tasks that can save human efforts and even be deployed in hazardous environments along with the research in the field of human mind and body has led to new fields often called AI Robotics. To provide a mean for stimulating students and researchers around the world for promotion of this new field, an annual world championship, Robocup, has been organized with start in 1997. During these competitions, the state-of-the-art technical achievements within Sensor Systems, Mechanics, Artificial Intelligence, Machine Learning and many other related fields are presented by the participants. The final goal is to create autonomous humanoids that can win a soccer match against humans by 2050.

Computer Vision and Robotics are tightly bound together as vision is the most important part of the sensorimotor system enabling a robot to act in a dynamic environment. This includes recognizing landmarks for localization, obstacles for collision avoidance, and objects for manipulation. To go one step further in this direction, the challenge within Sony AIBO quadruped ERS-7 class for year 2004 was to recognize objects in different lighting conditions and to recognize and remember the position of newly detected landmarks. For this class of robots, the playground was a green carpet, surrounded by white walls. Two nets and four beacons in four different two-color combinations were the only permanent landmarks.

The object recognition module was decided only to be designed for rendering the task of recognition of all predetermined as well as unknown landmarks on and around the field given a set of image frames containing shapes presented as color blobs. Most importantly, the module should be designed in a way that it could recognize the regular objects independent of the lighting conditions and calculate the angle and distance to each one as well as the confidence for the calculated results. The outcome was thereafter to be saved in a common space for the other modules to use and act upon.

As the most challenging object to recognize and estimate the distance and view angle to is the ball, the step by step approache to a successful solution is presented in the report. The next subject mostly discussed is how to recognize the nets, where several factors affect the precision and confidence. Other subjects as filtering the information, how to consider the pose of the head, and how to use a voting system for raising the confidence are discussed thoroughly. Also some pros and cons of each method are looked at. For the reasons discussed throughout the text, the recognition was primarily based on color verification and not the shape of the blobs. The module was implemented in Matlab for unit test and to verify the theories and later on implemented, assembled, and compiled in C++ for the final run.

Although individual tests in the lab environment showed on a successful module in recognizing objects on and around the field, lack of a consistent color segmentation table resulted in a poor quality of recognition in variable lighting conditions during the competitions. Nevertheless, the module is reusable as the methods are as general as possible to fit different platforms with diverse frame rates and dimensions.

Regardless of the outcome, the thesis work presented in this report lays a ground for a vision module that can perform the task of recognition of known and unknown objects. In particular, it reports on the methodology used for recognition of objects with estimation of their distances and angles to the AIBO, for which a software vision module is implemented. The contribution of this thesis is the methodology for detection of the peripheral of the ball blob, using a voting system, an attempt to merge blobs, specially for finding the rectangular shapes and correct estimation of their sizes, and an investigation about using rule based verification of the objects, methods that had not been used before by any teams in previous years. (Less)
Please use this url to cite or link to this publication:
author
Motallebipour, Hossein LU
supervisor
organization
course
EDA920 20041
year
type
H2 - Master's Degree (Two Years)
subject
keywords
Robocup, Robotics, Computer Vision, Legged Robots, Computer Science
publication/series
Lund : Department of Computer Science, Faculty of Engineering, LTH, Lund University
report number
LU-CS-EX: 2010-33
ISSN
1650-2884
language
English
id
3129335
date added to LUP
2012-10-30 23:16:59
date last changed
2015-02-10 11:17:46
@misc{3129335,
  abstract     = {A need for autonomous electro-mechanical machines performing tasks that can save human efforts and even be deployed in hazardous environments along with the research in the field of human mind and body has led to new fields often called AI Robotics. To provide a mean for stimulating students and researchers around the world for promotion of this new field, an annual world championship, Robocup, has been organized with start in 1997. During these competitions, the state-of-the-art technical achievements within Sensor Systems, Mechanics, Artificial Intelligence, Machine Learning and many other related fields are presented by the participants. The final goal is to create autonomous humanoids that can win a soccer match against humans by 2050.

Computer Vision and Robotics are tightly bound together as vision is the most important part of the sensorimotor system enabling a robot to act in a dynamic environment. This includes recognizing landmarks for localization, obstacles for collision avoidance, and objects for manipulation. To go one step further in this direction, the challenge within Sony AIBO quadruped ERS-7 class for year 2004 was to recognize objects in different lighting conditions and to recognize and remember the position of newly detected landmarks. For this class of robots, the playground was a green carpet, surrounded by white walls. Two nets and four beacons in four different two-color combinations were the only permanent landmarks.

The object recognition module was decided only to be designed for rendering the task of recognition of all predetermined as well as unknown landmarks on and around the field given a set of image frames containing shapes presented as color blobs. Most importantly, the module should be designed in a way that it could recognize the regular objects independent of the lighting conditions and calculate the angle and distance to each one as well as the confidence for the calculated results. The outcome was thereafter to be saved in a common space for the other modules to use and act upon.

As the most challenging object to recognize and estimate the distance and view angle to is the ball, the step by step approache to a successful solution is presented in the report. The next subject mostly discussed is how to recognize the nets, where several factors affect the precision and confidence. Other subjects as filtering the information, how to consider the pose of the head, and how to use a voting system for raising the confidence are discussed thoroughly. Also some pros and cons of each method are looked at. For the reasons discussed throughout the text, the recognition was primarily based on color verification and not the shape of the blobs. The module was implemented in Matlab for unit test and to verify the theories and later on implemented, assembled, and compiled in C++ for the final run.

Although individual tests in the lab environment showed on a successful module in recognizing objects on and around the field, lack of a consistent color segmentation table resulted in a poor quality of recognition in variable lighting conditions during the competitions. Nevertheless, the module is reusable as the methods are as general as possible to fit different platforms with diverse frame rates and dimensions.
 
Regardless of the outcome, the thesis work presented in this report lays a ground for a vision module that can perform the task of recognition of known and unknown objects. In particular, it reports on the methodology used for recognition of objects with estimation of their distances and angles to the AIBO, for which a software vision module is implemented. The contribution of this thesis is the methodology for detection of the peripheral of the ball blob, using a voting system, an attempt to merge blobs, specially for finding the rectangular shapes and correct estimation of their sizes, and an investigation about using rule based verification of the objects, methods that had not been used before by any teams in previous years.},
  author       = {Motallebipour, Hossein},
  issn         = {1650-2884},
  keyword      = {Robocup,Robotics,Computer Vision,Legged Robots,Computer Science},
  language     = {eng},
  note         = {Student Paper},
  series       = {Lund : Department of Computer Science, Faculty of Engineering, LTH, Lund University},
  title        = {An Approach to a Complete Recognition Module for Legged Robots in the RoboCup Domain},
  year         = {2010},
}