HotOS IX Paper
[HotOS IX Program Index]
Sensing User Intention and Context for Energy Management
Sensing User Intention and Context for Energy Management
Angela B. Dalton and Carla S. Ellis
Department of Computer Science
Durham, NC, 27708
Sensors are emerging as a key area of interest in operating systems
research, with a main focus on sensor networks. Turning the relationship
around, we propose the use of low-power sensors as tools for improving
OS-based energy management. Using sensors to detect user intent and
physical context we can more directly match system I/O behavior to user
needs. FaceOff is a prototype display power management system designed as a
test bed and proof-of-concept.
Managing energy as a resource is key to the future ubiquity of mobile
computing systems. Reducing power consumption is also a major challenge
in the design of mobile systems that extends beyond advances in battery
technology and low-power circuit design. Energy efficient
computer systems have broad environmental and economic
implications [4,5,16]. This position paper
focuses on using sensors to leverage physical context and user intent to
reduce a system's energy consumption. We illustrate this idea with a case
study on managing the display.
System level energy management approaches are currently tied almost
exclusively to process execution. We believe there is ample
opportunity for reducing a system's energy consumption by more directly
matching the system's I/O behavior to the user's own behavior. Consider the
display, a component that presents unique difficulties for power management
and typically represents the largest power consumer after the
CPU [9,15]. The display exists solely for the
purpose of user interaction and therefore it is only necessary when someone
is looking at it. There are many times when a user may turn his attention away
from a computer display, perhaps to answer a phone call or get a cup of
coffee. There are also scenarios in which the display is only used briefly
or not at all for a particular application. For example, someone using a
computing device to play music may only interact with the device to select
a song. Each time a song is selected, the user interaction would cause a
timeout-based display power management scheme to reinitialize the
timer. Similarly, someone may use a Personal Digital Assistant as a travel
alarm clock. When the alarm sounds, there may be no need to look at the
display, only to press a button in acknowledgement and turn off the
alarm. However, the PDA display is turned on by the user pressing a button
and would remain on for a timeout period. In these cases, turning off the
display immediately or never turning it on rather than waiting for a
timeout period would reduce energy consumption.
On the other hand the conventional timeout scheme which is based on lack of
user input may be too aggressive for some applications. A user reading an
electronic book or examining a web page with complex content
might experience the annoying behavior of the display timing out. The same
problem exists for a user watching a video or an automated slide show.
These are situations where the user interaction is dependent on the
display but is not tied to user interface input events.
In the next section we introduce our case study of a sensing system used for
power management. We then describe the architecture of the system and the
design of our prototype. In Section 4 we present energy
measurements of our prototype to justify its potential. This is followed
by a section speculating on additional roles of sensors. Section 7
outlines future work while section 6 discusses related work.
As a case study we evaluate a power management method that uses a web cam mounted to the display of a laptop as a sensor. The camera periodically captures images and a face detection algorithm determines the presence or absence of a user looking at the
display. Our research is an initial investigation into using sensors
combined with computer vision techniques to enhance display power management.
There were several open questions we hoped to address with our evaluation. First, computer vision in general is extremely computationally intensive. Are there optimizations we can use based on the constraints of the specific problem that will make the computing costs small enough to justify? Similarly, can the optimized algorithm run
quickly enough to appear seamless to the user? Can this method produce a
measurable reduction of energy consumption in the system even after
accounting for the added computing energy costs and the energy consumed by
the camera? Finally, are there specific applications to which this method
is particularly well-suited and what are the situations when it is not
We are designing a display power management system called FaceOff as a
proof-of-concept and a test bed for taking energy measurements as well as
obtaining more subjective user feedback. The FaceOff architecture is
simple and leaves significant room for optimizations to maximize energy
The FaceOff design consists, on a high level, of three main components:
image capture, face detection, and display power state
control (see Figure 1). The program periodically wakes up and calls
the image capture component. The image capture mechanism obtains a still
image from a camera and sends the image to the face detector for analysis.
Figure 1: A diagram showing the components of the FaceOff Display Power
The face detector processes the image using an optimized algorithm and
returns the Boolean value of true if a face is detected and false if no
face is detected. The display power is controlled using the ACPI
(www.acpi.info) interface to change the video device power state. When no
face is detected the program sets the device power state of the video to a
3 Initial Prototype
We are building the FaceOff prototype on an IBM T21 Thinkpad running Red
Hat Linux. The camera is a color Logitech QuickCam 3000 web cam that
connects via USB to the laptop with an average measured power consumption
of 1.5W. The display power states are defined in the ACPI specification and
supported by both the laptop hardware and the operating system. On this
laptop, the display consumes approximately 8.5W.
Our FaceOff prototype consists of a loop that captures an image from the
camera once a second. The image is saved to a file that is
processed by the face detection module. An obvious optimization to the
prototype is to eliminate use of the disk for storing the image
acquired from the camera. However, at this point we are logging
information to the disk for later analysis. Currently, the face detection
module consists of a skin color detector that looks for a large central
area of skin color in the image. Skin color detection was selected as a
fast and fairly simple method for the initial prototype, however more
accurate fast face detection methods exist and are part of our longer term
plans for the FaceOff system. We are integrating ACPI based power
state control into the prototype.
In order to evaluate the potential for our display power management method
to achieve energy savings, we examine usage
scenarios that should benefit from this approach. We measure the energy characteristics of the system assuming a
best case face detector and compare that to the energy characteristics of
the default timeout-based power management scheme on the same system with a
typical timeout of five minutes.
Current measurements were taken from a multimeter on the laptop's power supply with the
laptop's battery removed to eliminate changing effects. Voltage
measurements were taken for one case to verify that the voltage remains
constant throughout the experiment.
In this section we present three scenarios: a large network transfer, a
long, computationally intense process, and playing an mp3 song. The
scenarios were selected because they offer a comparison between FaceOff and
the default display power management scheme in which the timeout
intuitively does not capture the user's behavior well. It is likely that a
user might turn away from the computer after initiating a large file
transfer, beginning a large compilation, or starting to play a song. The
FaceOff system immediately turns off the display, however it continues to
perform the image polling and analysis, turning on the display when the
user returns. The default system will not turn the display off until the
timeout period expires, but it does not have the disadvantage of the
FaceOff system overhead.
The first set of experiments measured the energy consumption of the laptop
during a large network transfer. The transfers were performed using a
wireless network adapter on an internal network with no other concurrent
traffic. The measurements were taken assuming the best case in which the
user would initiate the transfer and look away, returning as soon as the
transfer completed. This application represents the case where FaceOff
overhead is not expected to affect performance. We measured the total
energy consumed during the tests as well as the time the network transfer
took to complete. The FaceOff technique used an average of 29.5% less
energy than the default, showing a significant
improvement. Table 1 shows a comparison of the energy and
time characteristics. Figure 2 shows traces of the power
over time for one run each of the network transfer with and without FaceOff.
Figure 2: Power traces for large network transfer.
The second experiment measured the energy consumption of a laptop
performing a computationally intensive task. In this case, the task was to
compile the Linux kernel. No other programs were running on the machine
except in the case of the FaceOff measurement the FaceOff prototype
was running. This captures the competition for resources imposed by
FaceOff. Again, we assumed the best case of the user initiating the
compilation, leaving and returning immediately upon completion. The
FaceOff technique resulted in an average power savings of nearly 12%. The
results of the experiment reflect the fact that the default timeout-based
power management system turned off the display close to halfway through the
compilation, reducing overall energy consumption.
Figure 3 shows sample traces of the power over time
for the compilation process with and without FaceOff. The increase in
completion time using each method for both experiments was insignificant.
Figure 3: Power traces for Linux kernel compilation.
The third experiment, playing an mp3 song, was primarily a validation that
FaceOff would cause no noticeable effect on the playback of the song. In
addition, such a scenario highlights one in which the default timeout
mechanism will never cause the display to turn off. The song we used in
the experiment lasted 4 minutes and 11 seconds, and played with no
noticeable effect when the FaceOff prototype was running. The average
energy used in the default case was 4,714 Joules, versus 3,403 Joules with
FaceOff, a 28% energy savings.
While the experiments we have presented provide a basis for our argument of
using context awareness and user intent for power management, we believe
that technological trends provide further weight in our favor. The web cam we
used for the prototype system requires more power than we would anticipate
a camera in any production version of the system to need. Low power, tiny
CMOS cameras are now available that can be embedded into computer systems and
consume as little as 20mW maximum power[7,11]. Compared with the
1.5W average power consumption we measured for the prototype's web cam,
clearly the overhead can be much lower.
Table 1: Energy and Time Comparisons
|Network Transfer || FaceOff|| Timeout
| 351.3 || 348.6
Energy (J) || 4791.2 || 6795.4
| FaceOff|| Timeout
| 603.5 || 575
Energy (J) || 11023.7 || 12506.85
5 More Roles for Sensors
Although our initial measurements show that the possibility exists to save
energy using our method of power management, the method has significant
overhead taking away from the benefits. Most notably, as realized in the
initial prototype, the image capture and face detection are continuing
costs, whereas an extended idle period incurs no overhead once the timeout
The first observation is that people radiate thermal energy
and are detectible with small motion sensors. If no person is present, no face
will be detectable and therefore we do not need to capture images or run
the face detection computation. Extremely small, low power pyro-electric
sensors are available that can detect even slight human
Integrating such a sensor into our system would allow the camera to be
powered down until movement triggers the sensor. A conservative approach
designed to minimize the delay waiting for the display to turn on would
immediately turn on the display and capture an image. The face detection
method would then take over until no face is detected and no motion is
present. A similar optimization would be to use a touch sensor in the
laptop wristrest or on the edges of a handheld device. The observation in
this case is that if a user's hands are on the keyboard or holding a PDA in
a particular orientation, even with no input, the user is most likely
looking at the display. We can therefore either reduce the frequency of
image capture or eliminate it completely and keep the display on. Face
detection could also be completely suspended when there are
frequent user interface events (i.e., in a sense, merging with the traditional
While the main focus of discussion in this paper is display power
management, there are other opportunities in which sensing context could be used for system level energy management. For example,
microphones could be used to determine background noise level and possibly
adjust speaker volume. Sensors could be used to determine whether to
completely turn off the speakers if, for example, nobody is around to hear
them. Sensing 802.11 signal strength could be used to determine whether
to offload computation to a server. Remote process execution has
been shown to significantly reduce energy consumption of mobile
6 Future Work
The FaceOff prototype is a framework for applying user intent sensing to
display power management. We would like to expand its capabilities
to include motion sensing and enhance the face detection module,
possibly adding face or gaze tracking at various levels of detail. Our aim
is to evaluate what is needed to improve the accuracy of the power states
while minimizing the system overhead. We plan to experiment with the
frequency of image sampling to make the system more responsive during the
times when motion is detected and decrease or halt the image sampling when
no motion is detected. We eventually plan to incorporate a light sensor
into the prototype for determining optimal display brightness.
We also plan to use the prototype to evaluate the user's experience,
generally a qualitative measure rather than a quantitative one. For this reason
we are examining ways of quantifying the user's experience with the
prototype, for example by adding a button to indicate
annoyance at the display state changes. We would like to be able to gauge the accuracy of the prototype
in determining context.
We plan to do a comprehensive user study characterizing usage patterns
similar to the study done in . We will use the study to
provide estimates of energy savings from our display power management
technique under realistic laptop workloads. In addition to capturing usage
patterns for laptops, which we can do using our FaceOff prototype, we would
like to study other mobile devices that could benefit from our system.
7 Related Work
As power management at a software level has gained attention both in
research and industry, several standards have emerged. The first
standard was Advanced Power Management (APM), a BIOS-based power
management specification. APM provides CPU and device power management.
Device power states are transitioned based on timeouts. Problems found as APM
matured led to the development of the Advanced Configuration and Power
Interface (ACPI) Specification. In ACPI, power management decisions are
made by the operating system rather than the BIOS . Both
APM and ACPI provide an interface for changing the power state of the
display through software using Display Power Management Signaling
(DPMS). APM and ACPI provide hooks to manage the power state of the
display, however currently the only strategy for taking advantage of the
lower power states is turning off the backlight and display after a certain
period of time with no user input . The Compaq iPAQ PocketPC
has an additional method of display power management using an ambient
light sensor to allow for adaptable display brightness. The only other
novel policy ideas we found for reducing the power consumption of the
display were zoned backlighting proposed by Flinn and Satyanarayanan, a
method which presupposes hardware that is not yet
available , and the recent work on Energy-Adaptive
Display System Designs  that proposes software optimizations
called dark windows.
There are several projects that involve face detection and face, gaze and
eye tracking for perceptual user unterfaces. The Smart Kiosk System uses
vision to detect potential users and decide whether the person is a good
candidate for interaction. . CAMSHIFT is a face tracker that is being used to
control games and 3D graphics by defining head movements to perform specific
actions . Another related project is a perceptual user
interface for recognizing predefined head gesture acknowledgements. The
face detection is performed by using an IBM PupilCam to locate the pupils
in the image and then uses simple image processing techniques to detect the
upper face region . A series of articles on
Attentive User Interfaces discusses several projects that use eye tracking
to design context-based user interfaces . To our knowledge
there are no other projects that are integrating face detection and power
Sensors as components of sensor networks have recently become an
interesting target domain for operating systems research (e.g.,
TinyOS ). In this position paper, we turn this around
and consider low-power sensors as tools in the service of OS-based energy
management for mobile computers. As a case study, we consider sensors
providing information from which to infer user intention and user
context as it affects energy management of the display - capturing
the direct dependency that looking at the screen
suggests a need for it to be illuminated. Intuitively, this is
attractive as a more direct indication of the user's need for display
power consumption than the keyboard and mouse input events used in traditional
Our preliminary exploration of this use of sensors to
inform the OS combines currently available technology that allows
software to switch the display power state, low-power sensors, and
face detection techniques.
We have proposed a method of reducing display power
consumption by turning the display off in the absence of a user. Face
detection, while a computationally intensive technique, can be optimized
for the simplified problem of detecting an upright, frontal face of an
approximate size indicating the presence of a user looking at the display.
For our FaceOff prototype, a web cam acquires images and the computer's own CPU
performs skin color based face detection.
Measurements of power consumption using the prototype system indicate
the promise of significant energy savings from this type of
context-based display power management scheme. Camera technology
trends indicate that cheap, very low power cameras are becoming more
readily available and could produce greater net energy savings in the
future using our technique. Furthermore, the specific task of user
detection could potentially be optimized using additional low power
sensors combined with less computationally intensive techniques to
further reduce overall energy consumption.
While the most obvious immediate benefit of our display power management
system would be extending the battery life of mobile devices, the method
could also provide the basis for energy savings on larger scale displays. One
interesting possibility is to apply a similar technique to very large
displays, using gaze tracking to determine what part of the display to turn
In this position paper, we have demonstrated that exploiting
low-power sensors to assist the OS in inferring user intention and
context for the benefit of energy management is a fruitful direction.
This work is supported in part by the National Science Foundation (ITR-0082914,CCR-0204367).
Gary R. Bradski.
Computer vision face tracking for use in a perceptual user interface.
Intel Technology Journal, 2(2):12-21, 1998.
Parts for pyroelectric infrared motion detector.
James W. Davis and Serge Vaks.
A perceptual user interface for recognizing head gesture
In Workshop on Perceptive User Interfaces 2001, Orlando, FL,
C. S. Ellis.
The Case for Higher-Level Power Management.
In Proceedings of the 7th Workshop on Hot Topics in Operating
Systems, Rio Rico, AZ, March 1999.
Jason Flinn and M. Satyanarayanan.
Energy-aware adaptation for mobile applications.
In Symposium on Operating Systems Principles (SOSP), pages
48-63, December 1999.
Jason Hill, Robert Szewczyk, Alec Woo, Seth Hollar, David Culler, and Kristofer
System architecture directions for network sensors.
In Proceedings of the International Conference on Architectural
Support for Programming Languages and Operating Systems, pages 93-104,
Wilderness Security Inc.
Bd-127 high resolution cmos camera.
Subu Iyer, Lu Luo, Robert Mayo, and Parthasarathy Ranganathan.
Energy-adaptive display system designs for future mobile
In Proceedings of The First International Conference on Mobile
Systems, Applications, and Services, May 2003.
J. Kolinski, B. Press, and A. Henroid.
Power Management History and Motivation, chapter 2, pages
Intel Press, September 2001.
Jacob Lorch and Alan J. Smith.
Software Strategies for Portable Computer Energy Management.
IEEE Personal Communications Magazine, 5(3):60-73, June 1998.
Space Television Ltd.
Color and b/w cmos board camera.
J. Rehg, M. Loughlin, and K. Waters.
Vision for a smart kiosk.
In Proc. IEEE Conf. on Computer Vision and Pattern Recognition,
pages 690-696, 1997.
Alexey Rudenko, Peter L. Reiher, Gerald J. Popek, and Geoffrey H. Kuenning.
The remote processing framework for portable computer power saving.
In Selected Areas in Cryptography, pages 365-372, 1999.
Attentive user interfaces.
Communications of the ACM, 46(3), 2003.
Gregory F. Welch.
A survey of power management techniques in mobile computing operating
Operating Systems Review, 29(4):47-56, 1995.
Heng Zeng, Xiaobo Fan, Carla Ellis, Alvin Lebeck, and Amin Vahdat.
ECOSystem: Managing energy as a first class operating system
In Tenth International Conference on Architectural Support for
Programming Languages and Operating Systems (ASPLOS X), October 2002.
File translated from TEX by TTH, version 1.50.