Graphics Reference
In-Depth Information
of smartphones and tablets has accelerated the pace of this trend. These mobile
devices include powerful processors, ample memory, and high-resolution cameras.
Coupled with high-level operating systems such as iOS and Android, these enable
the quick development of computer-vision-based applications. One can argue that
these devices provide a conduit for the deployment of embedded computer vision in
the consumer electronics market. However, these devices are fairly expensive, which
allow them to provide the resources computation- and memory-hungry computer
vision applications require.
Consumer robotic devices, on the other hand, face severe constraints on the cost
of computation. The mass consumer market is very price-sensitive, so the retail cost
of the robot is key for the success of the product. The consumer electronics industry
standard suggests a retail price for the product that is 3-5 times the cost of parts
(bill of materials, or BOM). In other words, for a $300 MSRP robot, the BOM
should be between $60 and 100, including all mechanical parts, electrical parts,
battery, processor, memory, motors, assembly, packaging, user manuals, and miscel-
laneous items! These strict cost constraints translate into a reduced availability of
computational resources, requiring the development of particularly lean algorithms.
Consumer robotics thus serves as an important platform for deploying embedded
computer vision applications.
Another interesting factor that distinguishes the smartphone from the robotic use
case is autonomy. Unlike in a smartphone app, the vision system in a robot must
work reliably with no user assistance, continuously and for long periods of time.
There is no opportunity for a user to correct an error or ignore a defect; if the vision
system fails, the effectiveness of the overall system is reduced.
Visual localization and mapping is attractive for low-cost robotics applications
since cameras are data rich, low power, and inexpensive. The challenge lies in design-
ing an algorithm that can efficiently extract relevant information from this high-rate
visual data stream. Despite Moore's Law, low-cost embedded platforms are still
constrained by limited processing power, memory, and storage. Many state-of-the-
art approaches to visual SLAM rely on interframe tracking, which requires high
frame rate processing. Additionally, common constraint graph SLAM methods for
agglomerating sensor information often incur computation and storage costs that
grow with time rather than with space explored. For a robot operating for extended
periods within a limited spatial area—typical of practical applications—this is an
undesirable trade-off.
This chapter describes the development of a localization system that can enable
systematic navigation of domestic robots in a household environment. The target
application is a mobile domestic robot with a price lower than $1000, and ideally
below $500. We present an approach to visual localization and mapping designed
for a low-cost robotic platform equipped with simple odometry and a single camera.
Operating primarily as a recognition engine, the visual measurement subsystem
requires only occasional, weak assumptions on processing rate, and intrinsically
provides robust loop closing when previously-mapped areas are revisited. Visual
measurements and odometry are fused in a graph representation and optimized incre-
mentally. Important novel features of this system include techniques for bounding the
Search WWH ::




Custom Search