AAMPL: Accelerometer Augmented Mobile Phone
Localization
Andrew Ofstad Emmett Nicholas Rick Szcodronski Romit Roy Choudhury
Dept. of Electrical and Computer Engineering
Duke University
{andrew.ofstad, emmett.nicholas, ras18, romit}@ee.duke.edu
ABSTRACT
A variety of mobile phone applications are on the rise, many
of which utilize physical location to express the context of
information. This paper argues that physical location alone,
unless remarkably precise, may not be sufficient to express
this context. Even slight localization errors may cause a
mobile phone to be placed in a grocery store, as opposed to
its actual location in an adjacent coffee shop. Applications
such as location specific advertisements, can get affected.
This paper proposes accelerometer augmented mobile phone
localization (AAMPL), a system that uses accelerometer sig-
natures to place mobile phones in the right context. Early
evaluation on Nokia N95 phones shows that AAMPL can
correct locations derived from Google Maps. We believe
that AAMPL can be extended to additional sensors (like
light and sound) to further aid GPS-free localization.
Categories and Subject Descriptors
C.3 [Special-Purpose and Application-Based Systems]:
Real-time and embedded systems; C.2.4 [Computer Com-
munication Networks]: Network Protocols
General Terms
Design, Experimentation, Measurement, Human Factors
Keywords
Mobile phones, accelerometers, localization, energy, sensing
1. INTRODUCTION
The proliferation of mobile phones is motivating a va-
riety of pervasive, context-aware, social applications. Ex-
amples include Micro-Blog [1], MetroSense [2], Place-Its [3],
PeopleNet [4], MyExperience [5], and several others. Many
of these applications exploit the location of the user as a
primary indicator of context. While most applications as-
sume GPS based localization, recent investigations are be-
ginning to expose several tradeoffs when using GPS. Poor
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
MELT’08, September 19, 2008, San Francisco, California, USA.
Copyright 2008 ACM 978-1-60558-189-7/08/09 ...$5.00.
indoor coverage with GPS, alongside its high energy con-
sumption [1], warrants alternate localization methods. Some
alternates, proposed by Place Lab and others [6, 7], utilize
WiFi/GSM based fingerprinting and triangulation to local-
ize mobile phones. While the improvements in coverage and
energy are encouraging, they arise at the expense of higher
localization error ranging from 50 to 500m. Such error mar-
gins may exceed the tolerance thresholds of several appli-
cations. More importantly, even if these error bounds are
reduced to a few meters, they may still be insufficient to
capture the user’s context. We present our argument next.
Consider an example application – Micro-Blog [1]. One
feature of Micro-Blog is that location-specific queries are
geocast to mobile phones present at the corresponding lo-
cation. For example, an Internet user may query about the
“availability of free WiFi” at a particular coffee shop. If
this query reaches phones in the adjacent grocery store, the
replies to this query may be inapplicable. Existing localiza-
tion schemes, even with accuracy of few meters, may not
be able to avoid this. The physical separation between two
phones may be small (few meters), and yet they may be in
logically different contexts (opposite sides of the wall sep-
arating the coffee shop and the grocery store). We argue
that localization needs to be performed across two domains,
namely physical and logical. This paper presents a frame-
work, AAMPL, that accepts the approximate physical loca-
tion of a mobile phone, and augments it with context-aware
logical localization. The main idea is described as follows.
Modern phones are equipped with a large number of sen-
sors, including cameras, microphones, accelerometers, and
health-monitors. These sensors are natural candidates for
sensing the context in which a user is situated. Automatic
access to such context information can be exploited towards
localization. For instance, a user’s movement (derived from
the phone’s accelerometer) may be effective for predicting
whether the user is in a coffee shop or a grocery store. Since
geographical localization can narrow down the choices to a
few nearby contexts, accelerometer readings may be effective
in selecting the correct one from among them. Hence, the
“WiFi availability” query can be correctly guided to phones
in the appropriate coffee shop. While classifying accelerome-
ter signatures is one axis of augmentation, one may envisage
multi-dimensional context sensing through light and noise
signatures. This paper develops a framework for combining
physical and logical localization via real-time classification
of accelerometer readings. Evaluation on Nokia N95 phones
shows that AAMPL was able to correct physical locations
derived from phone GPS and Google Maps.

2. RELATED WORK
We divide the related work into two following branches:
Activity Recognition: Several papers have studied ac-
tivity recognition using accelerometers. Bao and Intille showed
that it is possible to detect 9 everyday activities using 2
biaxial accelerometers mounted on a user [8]. Other re-
search has investigated augmenting accelerometer data with
on-body audio sensors to detect recurring human behaviors
[9]. Project SATIRE [10] extends these results by imple-
menting a sensor mote based architecture that takes advan-
tage of recognizable accelerometer signatures. The authors
attempt to develop accelerometer-equipped smart clothes
that could trigger alerts or record daily activities. Several
projects have explored applications of context-aware mobile
devices [11, 12, 13]. Project SenSay [14] aims to provide
a context-aware phone that adapts its state based on the
environmental and physiological changes. It uses externally
mounted light, motion, and sound sensors to provide the
contextual information. Inspired by these findings, AAMPL
leverages accelerometer signatures towards augmenting lo-
calization. Readings from the on-board accelerometer of a
mobile phone are transmitted over WiFi/GSM connections
to a central server, which then localizes the phone in real
time. The architecture is made energy-aware, permitting
AAMPL to be a deployable underlay to next generation ap-
plications.
Localization: With increasing location based applica-
tions for mobile phones, localization has been an exciting
topic of research. Since GPS is energy-hungry and has poor
accuracy in indoor and dense urban areas, other localiza-
tion techniques have been explored [6, 7]. Augmenting GPS
with WiFi and/or GSM data has been shown to aid in lo-
calization. Also, external sensors can aid in placing a user
in a specific location [12]. The advantage of AAMPL is that
it unifies GPS based localization with other sensory inputs
(accelerometer in the case of our implementation). While
GPS offers an approximate location, an accelerometer based
classifier effectively discriminates between possible logical
locations. We present the design, implementation, and the
evaluation of the system, next.
3. ARCHITECTURE AND DESIGN
Figure 1 presents the client-server architecture of AAMPL.
Without loss of generality, we describe AAMPL for localiz-
ing phones located in urban business areas (shops, stores,
malls). However, AAMPL may be equally effective in many
non-business indoor environments, such as homes, schools,
or workplaces.
3.1 Client-Server Architecture
The AAMPL client runs on a Nokia N95 smartphone, and
uses a 3G or WiFi Internet connection to communicate to
the server. The client and server operate in conjunction to
first classify the type of business a user is in, based on ac-
celerometer data. The server then uses this classification to
choose from a list of businesses near the phone’s physical
location. The main operations are as follows.
The client collects X, Y, and Z-axis accelerometer data
at one second intervals. These data points are logged in
memory over the duration of a minute, and are then passed
Figure 1: Block diagram of the AAMPL architecture
through a filtering layer that classifies each accelerometer
data point into an “action state”, which, in our implemen-
tation, is either “sitting” or “standing”. This classified ac-
celerometer data is then aggregated into three features that
can be used at the server to classify a location (details pre-
sented later). The aggregated features are included in a data
packet, along with a timestamp and the physical coordinates
derived from GPS, WiFi, or GSM. The data packet is sent
to the server via an HTTP POST request.
The server is mainly responsible for classifying the accelerom-
eter data, and using the results to refine the physical location
of the device. Logically, the server application is divided into
three different components: (1) A MySQL database stores
business and mobile activity information, keeping an up-
dated list of nearby businesses and client data logs. (2) A
classifier is responsible for classifying the business category
based on a set of recent points in the “logs” database. (3)
A controller coordinates events across the different compo-
nents, and communicates to the mobile client when neces-
sary. When a packet first arrives, it is added to the“logs” ta-
ble, which contains a list of recent accelerometer data points
and location coordinates, provided by the mobile client. The
server dynamically queries Google Maps for an updated list
of businesses near these location coordinates. The controller
then examines the database to determine whether the lat-
est accelerometer data point is within a “span,” which is
a sequence of data points with close spatial and temporal
proximity to one another. The new data point is appended
to the most recent span if appropriate. If not, a new span
is created. Note that it is assumed that all the points in a
single span come from the same business location. The con-
troller then sends all the points from the latest span to the
classifier, which uses the accelerometer features to classify
the business category of the entire span. The controller can
then use this classification to filter out the nearby businesses
in the database, only returning businesses in the given cate-
gory. The filtering module also uses other relevant pieces of
information to filter out unlikely businesses. For example,
business hours, stored in the“businesses”database table, are
used to filter out locations that are not open. This list is
returned to the phone and displayed on the screen.