IEEE Robotics & Automation Magazine - September 2011 - 112

Remote Console

Other
Sensors

5 4

4
1

Visual
Sensor

Other
Sensors

Visual
Sensor

Mobile
Robot

200
180
160
140
120
100
80
60
40
20
0

Send Ref to PC

Token Process

Generate Track

Get Robot Location

Send Border Points

Adjust Snake

Extract Obstacles

Get Foreground Frame

Robot Not in View
Robot in View

Read One Frame

(ms)

Figure 9. A communication protocol.

obtained from multiple Bloom filters, as presented in the
"Multiple Bloom Filters for Distributed Routing" section.
In the experiments, the system achieves a performance
of request-response time of less than 0.5 s, average moving
speed of about 0.3 m/s, and a maximum control error of
less than 0.2 m. The control error considered here is the
deviation of the robot from the snake, which produces a
dynamic path responding to unknown moving obstacles.
This is higher than the conventional tracking error with a
static path because of unexpected intruders and longer
control latency. The wheelchair is driven by following a
velocity profile obtained from the predictive control for
each sample period. If an intruder is detected by a visual
sensor during this period, the snake will be updated and
may produce a bigger error in the next control period. As a
result, the maximum error is dependent on the control
latency and robot speed. For behaviors requiring higher
accuracy, such as passing a door, the snake will be constrained with less flexibility in that area, and the predictive
control will automatically generate a lower speed to
guarantee a safe passage; see the generated velocity profiles
at about 2.4 and 3.1 s in Figure 8 for passing obstacles.
The time spent on individual tasks by a visual sensor
module was recorded in Figure 10. The sampling period can
be varied between 260 and 400 ms depending on whether
this visual sensor is controlling the robot. The visual sensor
in control of the robot does not increase the processing time
for reading a frame, extracting obstacles, and adjusting the
snake but increases the time for trajectory generation using
predictive control, communication token processing, and
sending the snake to other nodes. The time for extracting
the foreground also increases with more time needed to distinguish the robot from the background.

Figure 10. Time spent on tasks.

1) Control Points Exchanging Commands: The coordinates
of control points are exchanged between adjacent wireless sensors for force calculation to deform the snake as
in the "Snake as a Path Planner" section.
2) Control Token Commands: At a specific time, only a single sensor node with the token can control a robot as the
coordinator. The commands are used to broadcast the
ownership of the token periodically or initiate a token
handover procedure.
3) Control Commands: The vision sensor with the control
token will send the velocity profile obtained from the
predictive control to the robot for tracking, as shown in
the "Predictive Control for Snake Tracking" section.
4) Monitoring Commands: The commands are used to
send out current information, e.g., control points and
detected obstacles, to a remote console for the purpose
of monitoring.
5) Service Request Commands: These commands are from a
remote console, such as a PDA, to request a navigation service to a destination. The route to the destination can be
112

IEEE ROBOTICS & AUTOMATION MAGAZINE

SEPTEMBER 2011

Conclusions
A robot with limited intelligence can be navigated with distributed visual sensors. Multiple Bloom filters can be used
for efficient routing, and a snake can be used as a path
planner and an efficient network organizing mechanism.
The multiple Bloom filter is based on an EEB design,
instead of the conventional FPB design. It achieves a uniform relative error distribution and uses less memory. The
snake-based predictive control is a concise and unified
mechanism to embrace all the navigation components
from path planning, trajectory generation to motion control, dealing with kinematic and dynamic constraints. The
method is presented as a general approach for navigation
using a wireless sensor network, with limited memory
space and computational power in individual sensors.
Experiments prove that distributed cameras can be used to
enhance the mobility of service robots, such as indoor
wheelchairs for aged and disabled people.
Acknowledgment
The authors thank the anonymous reviewers and editors
for their constructive and helpful comments. This article

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2011