Computer and Robot Vision - IEEE Xplore

2 downloads 0 Views 451KB Size Report
Jun 14, 2011 - comparison of pros and cons of the approaches difficult, which can be seen as a limitation when compared to the well-defined problems in com ...
TC SPOTLIGHT



Computer and Robot Vision By Ville Kyrki ([email protected]) and Danica Kragic

V

ision is an integral part of the state-of-the-art robotics. Visual sensing has been found invaluable in applications from industrial to service robotics helping in solving problems from localization and mapping to manipulation and teleoperation. Robot vision is not just computer vision. While robotics poses several challenges for visual processing, for example, three-dimensional (3-D) operation and performance requirements, there is an aspect in particular that makes robot vision special and distinguishes it from mainstream computer vision. While the ultimate goal of traditional computer vision is to model and understand the world using images, robot vision aims for things beyond this: to allow a robot to complete various tasks by acting and interacting with its environment. The success of a vision system in a robot is then related to the success of the task, which means that vision cannot be independently considered but instead needs to be considered as a component of a complete robotic system. Robot vision is difficult, or is it? Vision in robotics cannot be understood only as a perceptual process, which is one of the major reasons that makes robot vision difficult. In traditional computer vision, which considers only the perception, there have been significant and rapid advances once problems have been clearly defined. In areas such as stereo and 3-D reconstruction, precisely defined Digital Object Identifier 10.1109/MRA.2011.941638 Date of publication: 14 June 2011

1070-9932/11/$26.00ª2011 IEEE

computational problems promise for new practias well as well-defined cal application areas if benchmarks have fueled the reliability was insignificant advances so creased. A key way to solve that most theoretical issues this challenge is to develop © EYEWIRE of these problems can be conbenchmarking activities. The sidered to be solved. A recent exam- second significant challenge is the ple of the practical advances made current paradigm shift to cognitive through the theoretical work is the vision and robotics. Vision systems are introduction of Kinect, an inexpen- moving from systems that use measuresive projected light-based 3-D vision ments and simple sensor feedback to sensor. cognitive approaches that imply eleThe advances made in computer ments of interpretation and reasoning. vision can thus greatly benefit the robot- Enabling robots • icists using vision. However, defining to competently The strength of a similar clear problems in robotic vision perform everyday robotic approach to is much more difficult because the range manipulation acof possible solutions to a problem is tivities exceeds in cognitive vision comes wider. For example, the problem of terms of comfrom the fact that a grasping an object using vision can be plexity anything solved using different approaches, from that we have so robot can act to 3-D modeling of the object and grasp far investigated or confirm a hypothesized planning to approaches using machine successfully imlearning to learn suitable grasp points plemented in rofunctionality. directly from two-dimensional image botics and artificial • structures. intelligence at On the one hand, this makes the large. So, we pose a question to the comparison of pros and cons of the community: What are the biggest chalapproaches difficult, which can be lenges in the area of robot vision that seen as a limitation when compared would be most empowering to robotics to the well-defined problems in com- if they were solved? puter vision. On the other hand, the To operate in an everyday world, multitude of possibilities allows greater a robot may be required to find creativity in the solutions. While the objects that can fulfill certain criteria. variety is a limitation for systematical For example, a robot may look for study and benchmarking, at the end, it something to steer with or something makes it possible to successfully solve that can hold liquid. To understand complex problems through creative the semantic meanings, a robot needs solutions. to learn about its environment through Great challenges with big promises interaction with objects, observing are still underway. For example, the humans interacting with objects, or reliability of the existing vision sys- communicating with a human. For tems is currently not acceptable for example, a robot can interact with a many applications. There is a big human to build models of objects to JUNE 2011



IEEE ROBOTICS & AUTOMATION MAGAZINE



121



Figure 1. A robot detecting objects.

be manipulated, as shown in Figure 1, which shows a scene with four objects and a result of a multiobject segmentation. The strength of a robotic approach to cognitive vision comes from the fact that a robot can act to confirm a hypothesized • functionality. The success of a vision Another interesting area is system in a robot is dealing with obrelated to the success jects that change their appearance of the task. shape over • and the course of an action being applied to them. Examples include a cracked egg or hacked

ROS Topics



about the ongoing activities in the field, and serving as a contact point for researchers planning to get in to the research field. Members of the TC benefit by getting relevant information (for example, about workshops, publication venues, and potential collaboration partners) and joining discussions of the future of the field. The current chairs of the TC are Danica Kragic and Ville Kyrki. If you are interested in the activities or joining the TC, contact one of the chairs or visit the TC Web site at http://www.csc.kth.se/ danik/RAS.

• (continued from page 18)

nothing. Using this test, ROS Build 2 is about 100 times faster than the original ROS Build. Additional, significant speedup comes due to increased parallelism. On modern computers with many cores, using make’s “-j8” option can allow eight files to be compiled on eight cores all at once. Tools such as CMake can take advantage of this if they have broad visibility but will not be able to make full use of all the cores when limited to serially considering one package at a time. Originally, ROS was designed with the PR2 robot in mind, and the system has had great success across multiple robots. However, requirements change as ROS is ported to new robots, operating systems and

122

cucumber. Almost no research has been performed yet in this area. The current rapid developments in many of the action capabilities, such as manipulation skills, also fuel the research in robot vision. The technical committee (TC) on computer and robot vision aims at bringing together researchers sharing interest in robot vision and promoting discussion between computer vision and robotics communities. Activities include organizing tutorials and workshops in both computer vision and robotics conferences, editing special issues in journals, creating periodic innovation summaries, informing

IEEE ROBOTICS & AUTOMATION MAGAZINE



compilers. The out-of-source build (wherein the process of building software does not mutate the source tree at all) is an important technique for the developer and integrator and will be supported by ROS Build 2. The out-of-source build makes cleaning up extremely easy (rm -r), allows one to simultaneously compile with multiple compilers (including cross compilers), and generally makes integration tasks easier. In addition to building in a separate directory, ROS Build 2 also enables install and package targets via CMake’s utility CPack. Install targets lead to smaller Debian packages (.debs) and make it easier to package ROS on other Linux distributions such as

JUNE 2011

RedHat, since taking a more standard approach makes the job of package maintainers easier across all distributions. This in turn will make it easier for the ROS to be downloaded and installed on more systems. ROS is a community resource that helps robotics developers’ work together to be as productive as possible. The build system has been a workhorse of ROS, making it easy to extend the system and distribute the improvements. The new build system supports the community by making code sharing better: making builds more efficient and making it easier than ever to bring up ROS on a new platform.