Издательство Springer, 2006, -230 pp.
Autonomous robots have the potential to revolutionize the service industry. Specialized applications such as palletizing, vacuum cleaning and tour guiding can be solved with today's technology, but next-generation universal robots will require new solutions to the challenge of operating in an unconstrained human environment. For example, imagine you suffer from arthritis so severe that your daily needs go unattended. A robotic aid could restore your independence by helping to put on your shoes, pour a drink, stack dishes, put away groceries or a host of other household tasks involving lifting, bending and carrying. Such a robot must be capable of interpreting and planning actions based on simple supervisory commands, recognizing a variety of objects despite a continuous spectrum of variation in appearance, and manipulating a dynamic environment in continuously new ways. Unlike their industrial counterparts, widespread adoption of domestic robots demands minimal reliance on application specific knowledge and high robustness to environmental changes and inevitable operational wear. Rich sensing modalities such as vision are therefore likely to play a central role in their success. This book takes steps towards the realization of domestic robots by presenting the fundamental components of a 3D model-based robot vision framework. At all stages from perception through to control, emphasis is placed on robustness to unknown environmental conditions and calibration errors.
At the lowest level of perception, stereoscopic light stripe scanning captures a dense range map of the environment. Unlike conventional structured light techniques, stereoscopic scanning exploits redundancy to robustly identify the primary light stripe reflection despite secondary reflections, cross-talk and other sources of interference. An automatic procedure to calibrate the system from the scan of an arbitrary non-planar object is also described.
At the next level of perceptual abstraction, range data is segmented and fitted with geometric primitives including planes, cylinders, cones and spheres. Central to this process is a surface type classification algorithm that characterizes the local topology of range data. The classifier presented in this book is shown to achieve significantly greater noise robustness than conventional techniques. Many classes of domestic objects can be identified as composites of extracted primitives using a graph matching technique, despite significant variations in appearance.
At the highest level of perception, model-based tracking compensates for scene dynamics and calibration errors during manipulation. Selection of salient tracking features is a challenging problem when objects and visual conditions are unknown. This book explores the use of multi-cue fusion based on edge, texture and colour cues, which achieves long-term robustness despite losing individual cues as objects and visual conditions vary. Multi-cue tracking is shown to succeed in challenging situations where single-cue trackers fail.
Finally, robust hand-eye coordination to perform useful actions requires visual feedback control of a robot manipulator. This book introduces hybrid position-based visual servoing, which fuses kinematic and visual measurements to robustly handle occlusions and provide a mechanism for on-line compensation of calibration errors, both classical problems in position-based visual servoing.
The results of extensive testing on an upper-torso humanoid robot are presented to support the framework described in this book. The culmination of the experimental work is the demonstration of two real-world domestic tasks: locating and grasping an unknown object and pouring the contents of an interactively selected cup into a bowl. Finally, the role of vision in a larger multi-sensor framework is explored through the task of identifying a cup of ethanol from among several candidates based on visual, odour and airflow sensing.
Introduction
Foundations of Visual Perception and Control
Shape Recovery Using Robust Light Stripe Scanning
3D Object Modelling and Classification
Multi-cue 3D Model-Based Object Tracking
Hybrid Position-Based Visual Servoing
System Integration and Experimental Results
Summary and Future Work
A Active Stereo Head Calibration
B Light Stripe Validation and Reconstruction
C Iterated Extended Kalman Filter
D Stereo Reconstruction Error Models
E Calibration of System Latencies
Autonomous robots have the potential to revolutionize the service industry. Specialized applications such as palletizing, vacuum cleaning and tour guiding can be solved with today's technology, but next-generation universal robots will require new solutions to the challenge of operating in an unconstrained human environment. For example, imagine you suffer from arthritis so severe that your daily needs go unattended. A robotic aid could restore your independence by helping to put on your shoes, pour a drink, stack dishes, put away groceries or a host of other household tasks involving lifting, bending and carrying. Such a robot must be capable of interpreting and planning actions based on simple supervisory commands, recognizing a variety of objects despite a continuous spectrum of variation in appearance, and manipulating a dynamic environment in continuously new ways. Unlike their industrial counterparts, widespread adoption of domestic robots demands minimal reliance on application specific knowledge and high robustness to environmental changes and inevitable operational wear. Rich sensing modalities such as vision are therefore likely to play a central role in their success. This book takes steps towards the realization of domestic robots by presenting the fundamental components of a 3D model-based robot vision framework. At all stages from perception through to control, emphasis is placed on robustness to unknown environmental conditions and calibration errors.
At the lowest level of perception, stereoscopic light stripe scanning captures a dense range map of the environment. Unlike conventional structured light techniques, stereoscopic scanning exploits redundancy to robustly identify the primary light stripe reflection despite secondary reflections, cross-talk and other sources of interference. An automatic procedure to calibrate the system from the scan of an arbitrary non-planar object is also described.
At the next level of perceptual abstraction, range data is segmented and fitted with geometric primitives including planes, cylinders, cones and spheres. Central to this process is a surface type classification algorithm that characterizes the local topology of range data. The classifier presented in this book is shown to achieve significantly greater noise robustness than conventional techniques. Many classes of domestic objects can be identified as composites of extracted primitives using a graph matching technique, despite significant variations in appearance.
At the highest level of perception, model-based tracking compensates for scene dynamics and calibration errors during manipulation. Selection of salient tracking features is a challenging problem when objects and visual conditions are unknown. This book explores the use of multi-cue fusion based on edge, texture and colour cues, which achieves long-term robustness despite losing individual cues as objects and visual conditions vary. Multi-cue tracking is shown to succeed in challenging situations where single-cue trackers fail.
Finally, robust hand-eye coordination to perform useful actions requires visual feedback control of a robot manipulator. This book introduces hybrid position-based visual servoing, which fuses kinematic and visual measurements to robustly handle occlusions and provide a mechanism for on-line compensation of calibration errors, both classical problems in position-based visual servoing.
The results of extensive testing on an upper-torso humanoid robot are presented to support the framework described in this book. The culmination of the experimental work is the demonstration of two real-world domestic tasks: locating and grasping an unknown object and pouring the contents of an interactively selected cup into a bowl. Finally, the role of vision in a larger multi-sensor framework is explored through the task of identifying a cup of ethanol from among several candidates based on visual, odour and airflow sensing.
Introduction
Foundations of Visual Perception and Control
Shape Recovery Using Robust Light Stripe Scanning
3D Object Modelling and Classification
Multi-cue 3D Model-Based Object Tracking
Hybrid Position-Based Visual Servoing
System Integration and Experimental Results
Summary and Future Work
A Active Stereo Head Calibration
B Light Stripe Validation and Reconstruction
C Iterated Extended Kalman Filter
D Stereo Reconstruction Error Models
E Calibration of System Latencies