Context Acquisition Based on Load Sensing

Abstract

Load sensing is a mature and robust technology widely applied in process control. In this paper we consider the use of load sensing in everyday environments as an approach to acquisition of contextual information in ubiquitous computing applications. Since weight is an intrinsic property of all physical objects, load sensing is an intriguing concept on the physical-virtual boundary, enabling the inclusive use of arbitrary objects in ubiquitous applications. In this paper we aim to demonstrate that load sensing is a versatile source of contextual information. Using a series of illustrative experiments we show that using load sensing techniques we can obtain not just weight information, but object position and interaction events on a given surface. We describe the incorporation of load-sensing in the furniture and the floor of a living laboratory environment, and report on a number of applications that use context information derived from load sensing.

Full text

Context Acquisition Based on Load Sensing
Albrecht Schmidt, Martin Strohbach, Kristof van Laerhoven,
Adrian Friday, Hans-W. Gellersen
Computing Department
Lancaster University
Lancaster, LA1 4YR, UK
{albrecht, strohbach, kristof, adrian, hwg}@comp.lancs.ac.uk
http://ubicomp.lancs.ac.uk
Abstract. Load sensing is a mature and robust technology widely applied in
process control. In this paper we consider the use of load sensing in everyday
environments as an approach to acquisition of contextual information in
ubiquitous computing applications. Since weight is an intrinsic property of all
physical objects, load sensing is an intriguing concept on the physical-virtual
boundary, enabling the inclusive use of arbitrary objects in ubiquitous
applications. In this paper we aim to demonstrate that load sensing is a versatile
source of contextual information. Using a series of illustrative experiments we
show that using load sensing techniques we can obtain not just weight
information, but object position and interaction events on a given surface. We
describe the incorporation of load-sensing in the furniture and the floor of a
living laboratory environment, and report on a number of applications that use
context information derived from load sensing.
1 Introduction
Load sensing is widely applied in process control to measure the weight of goods, to
monitor the strain on structures, and to gauge filling levels of containers. However, in
ubiquitous computing applications, we find that despite the research community’s
strong interest in context-awareness and sentient computing, load sensing has not
received much attention as a potential technology for the capture of contextual
information. In this paper we aim to redress this balance by demonstrating that load
sensing can be a versatile source of context information for everyday environments;
just as location technologies are a source for much more than absolute position, we
show that load sensing technologies can provide rich contextual information, beyond
mere load or weight.
Context-aware computing aims to embody information about the situation in which
people interact with computational processes [5,8,15]. The notion of sentient
computing stresses how sensors can be used to obtain such contextual information [2].
So far, ubiquitous computing research has primarily embraced location sensing for
this purpose, as location information usually provides access to rich context beyond
position, for instance on available infrastructure and co-located people [15,17].

Computer vision is a similarly generic approach, as visual scenes lend themselves to
extract various kinds of context, for instance identity and location of people and
detection of certain activity [6,11]. Other sensing techniques including load sensing
have been used to obtain context in ubiquitous computing, but they are generally
considered as too specific for wider application.
Load sensing though has some properties that appear to be well suited for context
acquisition in ubiquitous computing. Firstly, gravitational force applies to all physical
things, giving them weight and in principle rendering them detectable by load sensors.
Secondly, changes in weight distribution in everyday settings can be assumed to be
closely coupled with human interaction in a physical environment. Thirdly, load
sensing lends itself to an event-based model in which applications respond to changes
in observed load and do not have to monitor load continuously. Finally, load sensing
is a mature and robust technology that lends itself to unobtrusive augmentation of
practically any surface. Moreover, load sensing can be realized at very low cost; as
evidenced by the use of strain gauges in kitchen and bathroom scales.
This paper investigates the utility of load sensing for context acquisition. As an
initial contribution, we identify three context primitives that can be extracted from
sensory observations on load-sensitive surfaces: weight, position, and type of
interaction. Weight is an obvious contextual parameter that can and has been used for
instance as key to identify objects. Load sensing for acquisition of object position and
interaction events though has not been considered in context-aware computing before.
We address this in our main contribution, a series of illustrative experiments that
demonstrate high accuracy object positioning, and classification of interactions such
as objects being placed, removed, or knocked over. The experiments explicitly
consider conditions in everyday environments, such as pre-loading of surfaces with a
multitude of objects. In the latter part of our paper we describe the implementation of
load-sensing in the furniture and the floor of a small living laboratory environment.
We also report briefly on a few applications facilitated in this environment,
demonstrating various uses of load-sensing based context.
2 Related Work
Previous research has considered observation of the forces applied on surfaces to infer
information on activity. More specifically, Addlesee et al. have built an Active Floor
in an office environment, measuring ground reaction force to identify and track
people [1]. In contrast to the load-sensitive surfaces considered in our work, they use
a segmented surface, i.e. a floor composed of load-sensitive tiles. The same approach
is used in the SmartFloor, implemented by Orr and Abowd in the Aware Home
project [13]. Ground reaction forces are studied in more detail in bio-mechanics
however based on special purpose equipment such as the Kistler force plate [9].
While the early Active Floor work used sensor analysis to extract identity, more
recent work has extended this to extract activity-related context, for instance whether
a person is jumping or lifting a heavy object [7]. Paradiso presented an augmented
carpet that facilitates tracking of movement, however designed as input device, for
instance for artistic performance, rather than as infrastructure for context acquisition

[14]. His work is based on a different sensing technology, using a grid of piezoelectric
wires under the carpet instead of load cells.
Active Floor and Smart Floor are all focused on floor-level sensing of people’s
identity, position or movement. In contrast, we are exploring more pervasive
augmentation of surfaces in everyday environment including specifically tables as
highly interactive spaces. More pervasive augmentation provides access to more fine-
grained information on activity, for instance based on identification and tracking of
objects.
An interesting example for object identification based on load sensing has been
used in the i-Land project for implementation of the Passage mechanism [10]. Passage
allows a user to temporarily link a virtual object to a physical item placed on a load-
sensitive surface (called bridge) which has high precision scales embedded. In the
system, the link is simply maintained by reference to the weight of the physical item,
enabling seemingly physical transport to another bridge where the virtual object can
be recalled. The Passage mechanism represents a very specific use of load sensing. In
contrast we seek to establish load sensing as versatile source for context.
2.1 Context Primitives obtained from Load Sensors
Our approach is to augment many surfaces from floors to tables, shelves, and possibly
even smaller surface units, with the intention of integrating these surfaces for
developing a more versatile load sensing platform for context capture. This is not
driven by a single application, but rather developed bottom-up by considering the
context primitives that can be obtained from load sensing, and the construction of
higher-level context capture techniques. We identify these primitives as:
• Weight: the load registered on a surface and attributed as weight to objects on
the surface.
• Position: the spatial location at which the weight (i.e. gravitation force) occurs.
• Interaction: the shape of the load signal over time resulting from interaction.
Weight is the most obvious contextual primitive that can be used for the
identification of objects; we may identify single instances of objects or entire classes
of similar objects from their intrinsic property of weight. Konomi’s passage
mechanism is an example for how weight might be used to identify objects.
Position has been considered in the Active Floor and Smart Floor; by dividing the
surface into a number of tiles some positional information is available by considering
each tile’s sensor ID. We have observed that more fine-grained positional information
can be obtained by utilising the distribution of forces on a single surface. We examine
this issue further in section 4.
The last principle we consider in this paper is interaction; this context can be
obtained by analyzing the output of the load sensors over time. One example where
the load sensor trend has been used to great effect is in the Active Floor, where the
interaction context is used to analyze differences in footfall. On a more simplistic
level, simple changes in output such as load increase and decrease may provide useful
context on object placement and removal, as examined in more detail in section 5.

3 Determining 2-D Position of Objects on Surfaces
In this section we demonstrate how the 2-D position of an object can be detected on a
surface using load sensing. We begin by describing the basic algorithms and
demonstrate the feasibility of our technique. The configuration of our experiment is
intended to recreate a typical surface that might be found in an everyday environment.
For the experiment we placed a table-top on four industrial load cells each of
which can detect forces of up to 500N. The load cells are placed at the corners of the
table-top and each connected to a commercial signal conditioning unit. The
conditioning units are in turn fed to a standard 16-bit Analog to Digital Converter
(ADC) which links to the serial line of the PC. The load cells are driven with 10V,
resulting in a output ranging from 0 to 40mV which is amplified by the conditioning
unit to 0 to 2.5V and sampled by ADC. As these components are normally used for
scales, the sampling frequency is necessarily rather low (each load cell can be read 4
times a second). The resolution of load sensing in th is setup is approximately 16g. To
minimize the influence of bit-errors in the ADC, our driver software only calculates
the position of objects between 100g and 100Kg (objects less than 100g are ignored).
3.1 Algorithm to Determine the 2D Position of an Object on a Surface
To detect the position of an object we calculate the centre of pressure on the surface
based on the load measured at each corner of the table. The overall force on the
surface introduced by an object placed on the surface at x,y is denoted by Fx. The
setup assumes static forces, so the sum of all 4 load cells F1, F2, F3, and F4 is equal to
Fx, see equation (1) and figure 1. If there is already an object on the table-top that is
represented by F0x that can be measured by the forces on each load cell F01, F02, F03,
and F04, see equation (2), then we need to incorporate these into our algorithm for
computing the position of a new object. The algorithm for identifying the position of a
new object on a pre-loaded surface is described by equation (3) and (4) below.
4321 FFFFFx+++= (1)
4321 00000 FFFFF x+++= (2)
)0(
)0()0(3322
max
xx FF
FFFF
xx −
−
+
−
= (3)
)0(
)0()0(4433
max
xx FF
FFFF
yy −
−
+
−
= (4)
The voltage level measured at each load cell is linearly dependent on the force
applied to it. In the experiment the measured voltage level at the load cells were used
directly without prior conversion into a value of the force applied. Similarly we store
the existing preload, F01, F02, F03, and F04 as raw voltage levels. We take all values

from the ADC and average them over the last 4 values. The size of the surface is fixed
in this experiment, xmax is 135cm and ymax is 75cm. The calculation for the position is
rounded to centimetres.
3.2 Proof of Concept: Position on the Surface
To carry out the experiments we developed a program in Visual Basic that reads
periodically from the ADC, calculates the point of pressure, and visualizes the result.
In the program the user can manually reset the preload by pressing a button, storing
the currently measured forces into F01, F02, F03, and F04.
In the first experiment, the position of objects on the table-top is detected. Before
each object was placed on the surface, we measured and stored the preload F01, F02,
F03, and F04 resulting from the table top itself. We then selected six well distributed
positions and marked them on the surface. We measured the x and y coordinates of
each of the positions manually and recorded them for later comparison. Next we
systematically placed one object at a time onto the selected points. After we put each
the object down and the values stabilised, we measured the load F1, F2, F3, and F4, and
calculated its position. This sequence was repeated for all six objects.
The objects we chose to measure were everyday items of varying weights, and
included a 500ml water bottle (about 530g) and a two litre water jar (about 2kg). We
regard the centre of the object as its position.
The results of the experiment show that it is possible to achieve an accuracy of
about 2% of the surface length in each dimension. In our tests this corresponds to less
than 4cm in the y axis 2cm in x axis. The difference between the positions obtained
from the weight calculations during repetitions of the experiment with the same set of
objects is even smaller: the maximum difference in both axis was 1cm, indicating that
we can achieve higher resolution by calibrating the system – if desired by the
application.
Force Fx
at (x,y)
Force F1
at (0,0)
Force F3
at (xmax,ymax)
Force F4
at (0,ymax)
Force F2
at (xmax,0)
Fig. 1. Forces on a surface used to determine the 2-D position of objects.

3.3 Real Environment: Position on a Tablecloth covered Surface with Preload
To simulate a more realistic environment, we repeated the same experiment with a
number of different objects already placed on the table. The table top was covered
with a tablecloth. Five objects were randomly distributed on the surface before the
preload values were acquired. These objects were: a TV-set, a book, a 2 litre water
bottle, a magazine and a cable reel, totalling an overall weight of about 34kg. Figure 2
illustrates this setup.
The initial set of measurement tests were repeated, the objects now being added on
top of the TV-set, on the magazine, and the other four on the tablecloth.
The results we obtained were similar to those obtained in the original experiment with
the empty table. In most cases we achieve a slightly better resolution because the
additional weight provides stability to the system.
This experiment shows that using load sensing we can determine the static position
of an object on a surface irrespective of objects already on the surface or between the
surface and the new object. Furthermore, our results indicate that it is possible to
reliably obtain centimetre level accuracy. The success of such an experiment also lead
us to postulate that our approach for object detection is deployable in non -lab
environments and is especially well suited to ubiquitous computing settings.
4 Recognizing Interaction on a Load Sensing Surface
As the world is not static and we interact with objects, place them on surfaces and
remove them again it is of interest to recognize events that relate to these actions. In
the second series of experiments, we explore how interaction resulting from events on
a load sensitive surface can be recognized with a simple algorithmic approach.
The setup used for the experiment is a wooden table-top, 80x80cm, resting on 4
industrial load cells. Each of the load cells can handle a maximum load of 20N. The
weight of the table-top is about 1kg. For this setup we built a hardware which
amplifies the output signal of the load cells by 220, resulting in an output voltage of
between 0 and 4.4V. Each of the 4 signals is sampled at 250Hz with 10 bit resolution
using a PIC16F876
microcontroller. The
microcontroller is connected
to a PC via a serial line. The
objects used to generate the
events in this experiment are
a 500ml water bottle (520g)
and a book (~200g). Detailed
information and the
schematics, software, and
datasets are available on the
project web page [12].
Fig. 2. The experimental setup; objects are stationary on
a table while the position of an added object is detected.

4.1 Algorithm to Detect Interaction Events on a Surface
To detect the events we use an algorithm which considers the last 500ms using a
sliding window and the sum of load (Fx). At the selected sampling frequency, this
equates to the last 125 sample values, denoted by Fx(t),…, Fx(t-124). At each sample
interval we run the analysis algorithm described by the formulae shown in table 1.
The expressions have been selected to be as simple as possible to facilitate easy
implementation on a microcontroller 1 while at the same time yielding reliable event
detection and differentiation.
In the following experiment we investigate whether or not it is possible to
recognize different events based on the features extracted from the data gathered.
Here we concentrate on the most important primitives: putting objects down onto the
surface and removing objects from the surface. We also included a further event:
knocking an object over which is already on the surface. The following rules are used
to decide which event has taken place.
1 In the later implementation the features were adjusted so that the division could be replaced
by a shift operation.
25
)(
)100)..(124(
∑
−−=
=ttj
x
s
jF
A
As is the average value of the first 25 values in
the window that is currently processed. This is
used as an indicator for the overall load on the
surface before the interaction.
75
)(
)25)..(101(
∑
−−=
=ttj
x
m
jF
A
Am is the average value of the middle 75
values in the window that is currently
processed. This value is required for the
calculation of Dm.
25
)(
)24)..((
∑
−=
=ttj
x
e
jF
A
Ae is the average value of the last 25 values in
the window that is currently processed. This is
used as an indicator for the overall load on the
surface after the interaction.
25
)(
)100)..(124(
∑
−−= −
=ttj
sx
s
AjF
D
Ds is an indicator for the change in the signal
during the first 25 samples. If no interaction
takes place Ds is close to 0.
75
)(
)25)..(101(
∑
−−= −
=ttj
mx
m
AjF
D
Dm is an indicator for the change in the signal
during the middle 75 samples.
25
)(
)24)..((
∑
−= −
=ttj
ex
e
AjF
D
De is an indicator for the change in the signal
during the last 25 samples.
Table 1. Formulae calculated to detect interaction events.

• Putting an object on the surface.
This is characterized by an increase in the overall load. In other words, before the
event the overall load is smaller than after the event (As+δ< Ae). The threshold of
weights is denoted by δ. Assuming that once the object has been put down on a
surface it remains stable, it can be seen that Ds is close to zero (Ds<ε). In the
middle of the interaction the change to the signal is greater (the moment the object
hits the surface) than later (when the object is already on the surface), stated as
(Dm> De).
• Removing an object from the surface.
This is inverse of placing an object on the surface, so the overall load is reducing
(As>Ae+δ). To begin with the signal is stable (Ds<ε) and the change during the
interaction is greater than at the end (Dm> De).
• Knocking an object over.
When an object is knocked over, this results in a large change in the middle of the
interaction, greater than at both the start and the end, and also greater than a set
threshold φ, (Dm>φ ∧ Dm>Ds ∧ Dm>De). As the overall weight on the surface stays
the same, As and Ae are similar (|As-Ae|<δ).
4.2 Proof of concept: Events on an Empty Surface
In this experiment the force over time is recorded while a number of different
interactions are performed. First, we place the water bottle on the table and remove it
after a few seconds. Next, the book is placed or dropped onto the table and lifted
away a few seconds later. Lastly, the bottle is put down on the surface, left there for a
few seconds, knocked over, and removed after a few seconds. The experiment is
repeated with each object ten times. Overall, this experiment results in 70 recorded
events for analysis. Using the simple algorithm described above 94% of the events
were classified correctly, 6% were missed, and no events were misclassified. A
visualization of the raw data stream is shown in figure 3.
1000
1500
2000
2500
Time
Load
E1 E2 E3 E4
Fig. 3. The graph shows the load change recorded over time. An object is placed on the surface at
position E1 and E4. At E2 an object is knocked over and at E3 the object is removed from the
surface.

4.3 Real Environment: Events on Tablecloth covered Surface with Preload
The experiment was repeated on a table that was covered with a tablecloth. On the
table were also 4 static objects, a notebook computer (about 2.2kg), a book (about
500g), a newspaper (about 200g) and a water bottle (about 520g). The same sets of
interactions were performed. About half of the interactions were made on the
tablecloth and the remaining half on top of static objects. Out of the 70 events
recorded, the algorithm could classify 96% correctly, 4% were missed.
4.4 Discussion on Event Recognition
As seen from the results of our experiments, it is feasible to detect basic interaction
events with a high degree of probability, even when using simple algorithms. We can
also observe that covering the surface has a negligible effect on the data recorded.
Furthermore, having other objects on the surface of the table, vibrations were
deadened increasing the recognition performance.
We believe that it may be possible to use similar techniques to identify further
events, such as moving an object on a surface, sliding an object over the surface and
touching the surface. The datasets of the experiments reported here and also for other
experiments are available on our project web page [11].
The observation that preload has little influence on detecting the events or
identifying an object’s position, suggests that it is feasible to implement a system that
can dynamically track objects that are added, moved, or removed from surfaces in
everyday environments.
5 Incorporating a Weight Lab into a ‘Living Laboratory’
The results obtained by our experiments on using the intrinsic weight of objects to
calculate position and detect events and activities on a given surface, suggest that a
larger scale implementation would provide us with a rich source of contextual
information. In order to prove this we have designed and implemented a “weight lab”
as an integral part of our living lab area. In the weight lab, all surfaces are load
sensitive and equipped with networked data acquisition units. Currently, the weight
lab is comprised of four ‘load sensitive’ artefacts: the floor, two tables and a shelf.
Other, non-load sensitive artefacts can be arranged on these surfaces. The
arrangement of these artefacts is not fixed; all components can be moved to create
new experimental configurations. We have designed the space to be used as a
common room, as depicted in figure 8 and 9.
In order to make the artefacts ‘aware’ of the load placed upon them and also of the
position of an object or a subject, the surface must be augmented with load cells.
Different load cells have been selected depending on estimates for the overall
operating load for each surface, together with some practical engineering and
aesthetic considerations. We have also tried to use cheap and easily available
consumer load cells taken from kitchen and bathroom scales where appropriate.

Dedicated hardware
has been developed to
drive the load cells and
facilitate the data
acquisition. The current
prototype offers both
wired RS-232 and
wireless communication
interfaces to our own
custom data analysis
software.
5.1 Weight Floor
The floor is constructed
from a wooden structure
of 240cm by 180cm. The
surface of the floor is
mounted on three
supporting timber beams
resulting in an overall
height of approximately 9cm. At each corner of the floor a single load cell is mounted
into the supporting timber, see figure 4. The whole floor rests entirely on the four load
cells and the remaining structure is in contact with the conventional load bearing floor
of the building.
An estimate of the typical load in our environment is: 2 people (70kg each), 2
armchairs (15kg each), a coffee table (10kg), a shelf (20kg), and the weight of the
floor (80kg) resulting in 280kg.
Due to the structure and anticipated loading of the floor, we selected S-load cells
each with a capacity of 1000N. In this range the measurements have a guaranteed
accuracy and the cells continue to be overload safe to 2000N (in this range the load
cells still give a reading and are not damaged, but their accuracy is reduced). As there
are 4 load cells, the total weight that can be put on the floor (including the weight of
the floor, itself about 80kg) is 800kg or 400kg with well defined accuracy. The
anticipated load of 280kg - even if such a load is not evenly distributed – is still well
within the specification.
The four load cells are connected to the data acquisition hardware described later
in this section. The floor system incorporates an ‘auto-tare’ mechanism which allows
us to discount the position of stationary objects and improve our estimation of the
position of new objects introduced into the environment. More specifically, whenever
the load is considered stable this preload is stored (F01, F02, F03, and F04) then
factored in to successive calculations to determine the point of pressure of further
objects, e.g. the position of someone is walking on the floor. Using this mechanism,
furniture can be added to the floor and automatically included in the position
calculation once the floor is not occupied for several seconds. Stability is assumed if
the values of Ds, Dm, De. are close to zero for more than 5 seconds and the preload
Fig. 4. The floor when it was installed in upright position (left).
Enlarged view of the load cell embedded into the floor (right).

values are updated. The position algorithm, given in section 3 is used for calculating
the point of pressure. Note that in case of a single person the position of the person is
determined, whereas in case of more people, the point of pressure is the centre of
mass of the group. Although not yielding their individual positions, such information
may still help us determine the locus of activity.
The floor currently recognises three types of event: no interaction, people moving
and stationary occupation (e.g. someone is sitting in an armchair or standing in front
of the pin board). These events can be recognized with an accuracy of over 99%. The
position of a single person moving in a space already populated with furniture and
covered with a rug can be acquired with approximately 10cm accuracy.
5.2 Weight tables and shelves
We embedded load sensing technology in several pieces of furniture: a small coffee
table, a larger dining table, and a shelf/drawer unit. Both tables are constructed using
load cells installed between the table top and the frame, such that th e table top rests on
a load cell at each corner (see figure 5).
The coffee table can measure a maximum load of 8kg (including the 1.5kg table
top). Loads over 8kg are prevented by a mechanical overload protector which also
suspends data acquisition. The low overall weight capacity is acceptable given the
Fig. 5. Coffee table (top) and dining table equipped with load cells (bottom). Close ups of the
load cells and how they are fixed (right).

normal use of the table (e.g. putting
newspapers, magazines and cups on it)
and that it permits us a far higher
accuracy of measurement. For the larger
‘dining’ table we selected load cells with
a capacity of 500N each, resulting in an
overall capacity of 200kg (which is
appropriate as people sometimes sit or
lean on the edge of tables).
The tables and shelves use the same
data acquisition hardware as the floor,
but run different software on the
microcontrollers. The main events we
recognise are as explained in section 4.
However, with the sensitive small coffee
table we are able to track the pressure
from a finger or an object over the
surface and convert it into a mouse-like track-pad, for details see [12,16].
The shelf unit was placed on 4 load cells of the same type as the ones used for the
large table. The position of the interaction is assigned to a column of the shelf, as a
single set of load cells can not detect where in the vertical axis the interaction took
place. Additionally, some of the shelf boards within the unit are equipped with load
sensing boards. These boards are built based on cheap consumer load cells.
5.3 Hardware for Data Acquisition, Event Recognition and Communication
All of the load cells used in the experiments presented in this paper are based on
resistive technology. Put simply, each cell is a wheat stone bridge providing a
maximum output signal of 20mV at a driving voltage of 5V. The AD-converter
included in the microcontroller can measure voltages between 0 and 5V. To best
utilise this range, we amplify the output signal of the load cells by a factor of 220
using an LM324, resulting in a output signal of 0 to 4.4V (the exact values vary
slightly between the load cells). The amplified output voltage of each of the load cells
is converted into a 10 bit sample; the best resolution offered by the MCU’s internal
AD converter. Each of the four input channels is sampled at about 250Hz. The four
input values correspond to the load that is measured on each of the load cells, setting
the values for F1(t), F2(t), F3(t), and F4(t).
Events from the table and shelf unit are sent wirelessly using an RF transceiver
module (Radiometrix BIM2) that offers data rates of up to 64kbit/s. We run the
protocol at 19,200 bits/s. Events, such as putting an object down at a certain position,
removing an object, tracking over the surface or pressing down, are sent in a single
packet. Each packet comprises a preamble, followed by a start-byte, an object
identifier to determine the origin (coffee table, large table, shelf), the event type, and
the event dependent data. Finally, two bytes of 16-bit CRC are attached to ensure that
the transmitted data will be received correctly. The data acquisition unit only
Fig. 6. First generation data acquisition
and communication hardware

transmits data (there are no
acknowledgements), however,
we have found the protocol to
be very reliable at the low
transmission speed.
The hardware module used
for data acquisition (as depicted
in figure 6) also acts as a base
station that receives the events
and sends them to the host PC
via RS-232. The floor is directly
connected to the PC. The
experiments reported in the next
section are based on this
hardware.
Learning from the experience
gained by deploying our system
we have designed and built a
second hardware iteration that
offers some improvements over
our first design. The block
diagram of the newer system is
depicted in figure 7. In
particular the following improvements have been made:
• Increased AD resolution.
To avoid the influence of single bit errors in the position calculation and also to
increase the weight range of the objects that can be detected, an ADC with higher
resolution has been chosen (an external 16 bit ADS8320 analog to digital converter
with a voltage reference of 3.3V). This chip offers only one input so a multiplexer
was used.
• In order to improve the performance of the amplification we used instrumentation
amplifier (INA118 from Analog Devices) instead of Op-Amps. The INA118 offers
an amplification range of up to a factor of 10000.
• As the load on the surface changes and as it is also desirable to be able to detect
very light objects or minimal interaction, the factor for the amplification can be
selected at runtime from the microcontroller. Two amplification factors are
implemented (factor 150 and 1000) using a solid state relay that switches resistors
in the periphery of the instrumentation amplifier.
• Additional RAM.
In some of the more advanced scenarios we conceived, we realized that keeping a
history on the artefact is useful to detect more complex events or patterns of
events. Therefore we added a FRAM chip that is connected over I
2C and offers
8KByte of non-volatile RAM (FM24C64 from Ramtron).
Further technical details, full schematics, and PCB layouts together with the
software are available on our project page [11].
4-1 MUX
Load
Cell
Load
Cell
Load
Cell
Load
Cell
amp amp amp amp
AD conversion, 16Bit
ADS8320
Processing, PIC16F876
RF-Communication
Radimetrix BIM2
Serial
Communication
FRAM
64kbit
Fig. 7. The second generation of the load sensing
and communication unit.

5.4 Software
Several pieces of software have been developed to capture and analyse the data from
the load sensors. The microcontroller runs a basic interactive system which allows the
user to configure the board for a particular artefact via serial line. Parameters, such as
the size of the surface, sampling speed and the object identifier can be specified using
a serial terminal program. These parameters are permanently stored in the internal
EEPROM of the microcontroller. After the board is configured all events recognized
are broadcast using the RF module.
On the base station a program is running that receives all the broadcast events and
passes them to the PC. The unit offers two modes: one where all the data coming in is
streamed over the serial line, and the other, where the last 3 events are stored for each
artefact and can be requested from the PC using the artefact identifier.
We have also developed a number of programs for the PC to allow the storage and
visualisation of the data. One of these applications is depicted in the following section
in figure 8. This application visually maps events happening on the tables and on the
floor onto the plan of the lab.
6 Applying the Setup in a real-world Environment
The infrastructure described in section 5 is deployed in a part of our living lab area.
The basic floor plan and photos are presented in figure 8 and 9. On one side of the
floor there are two armchairs. In front of the armchairs there is the coffee table. On
the wall behind the floor there are two pin boards. To one side of the floor is the
larger table, next to the floor. On this table is a TV-set and people put things
temporally down (e.g. orders arrived, books, food) on the remaining part of the table.
The area is used to as a place where people come and chat, have a coffee or where
people read.
In the remainder of this section we will report on some of the applications
implemented on the load sensing infrastructure and also on the experience gained
from deploying the system.
6.1 “Don’t leave your things behind” – Application
This application shows the potential of having interconnected surfaces that can sense
weight. As the larger table is used by people to leave items (e.g. files, books, etc.)
temporally, people sometimes forget whatever they have put down when they leave.
This application reminds the user to take their objects with them when they leave the
room. The basic idea of the application was inspired by ethnographic studies reported
in [4].
Whenever the addition of an object on the large table is recognized, the weight
added to the table is stored together with the weight added on the floor (e.g. usually
the persons weight). When the person leaves the floor (the overall weight is reduced
by a certain amount) – the negative change of weight is used to check for an entry in
the stored data set. If there is an entry – the person has put something down on the

surface while entering – the software running on the PC provides an audio cue to
remind the user to take his items with him. As the weight of the object is known it is
also possible to prompt the user when the disappearing weight is not matching those
stored in the dataset.
6.2 Tracking the Position of a single Person and Estimating Activity
Using the setup described we have built an application that can track the position of a
person in the space. The position of the person is visualized, as shown in figure 8. The
Fig. 8. Weight lab viewer showing a trace on the floor layout of the lab and also some
interaction with the table (object put down).
Fig. 9. Corresponding situation to the mapping in figure 8. User is interacting with the right
pin board and then sits down in the arm chair in the corner.

position of the person is available in the system as coordinates in centimetres (x,y).
Given domain knowledge, e.g. the whereabouts of furniture, walls and devices, this
makes it possible to provide the position information in a symbolic form, such as on
chair 1, on chair 2, in front of the left pin board, in front of the right pin board, in
front of the TV, entering the floor. As suggested in [3], geometric domain knowledge
can also be used to create a higher level description of the position. The positional
information is also available via HTTP to be included in other applications. As there
is also a temporal element to the tracking data the direction and speed of movement
can be determined and offered to applications for further use.
Accumulating the tracking data over time offers a way to estimate the overall
activity in the space. A simple measure for this is to calculate the distance the person
has covered in a certain time.
6.3 Recognizing Situations
In figure 8, the tracked activity over a period of approximately 50 seconds is
illustrated. The corresponding situations are shown in figure 9. The person is entering
the floor, going to the left pin board, turning around to the right pin board, and then
walking to the upper arm chair. On his way he puts a chocolate bar on the table.
The context of sitting down on a particular chair, putting a cup down on the table,
taking a magazine from the table can all be recognized in the above setting. On a
longer time scale, it may be possible to determine patterns that relate to non-atomic
longer term activities such as drinking coffee, having lunch, reading the newspaper
given certain side conditions, such as domain specific knowledge about the location.
These situations can be described in terms of rules. For some of the situation these
rules involve constants that are specific for the space which is occupied, e.g. the
position of the pin board or where the table is. These constants are acquired either
geometrically by measuring and calculation, or by supervised learning from example
situations.
As more people enter the same space, the recognition of context is made more
complex. However as people enter one after another (even if the time between them
may be very short, e.g. less then a second) it is still possible to detect the number of
subjects. When multiple people are on the weight sensitive floor the centre of pressure
is usually between them (their aggregate centre of mass). For example when both
chairs are occupied the centre of pressure is in the middle between the chairs. To
recognize such contexts we find it useful to create the situation, log the data, and
record the analysis as vectors; new situations are then matched against these samples
to select the closest match. This can be realized using nearest neighbour matching or
more complex supervised learning methods.
7 Conclusion
In the work presented in this paper we have tested the hypothesis that load sensing
can be used in everyday environments as a versatile source of contextual information.
While it is obvious that weight information is a useful context for the identification of

objects, we have shown that load sensing can also be used to obtain positional
information and interaction events.
We have demonstrated that the position of objects on surfaces can be determined
with high accuracy and reliability. It has to be stressed that the algorithms are simple,
and geared toward implementation on low-cost micro-controller platforms. Further,
interaction events such as object placement on and removal from a surface can be
detected, likewise with high reliability and little computational cost.
The reported experiments also highlight the applicability in everyday settings
which we have simulated by preloading surfaces with additional objects. In order to
continue to study load sensing under living lab conditions we have augmented a
number of surfaces, including a relatively large unsegmented floor area, various tables
and shelves. The setup is an example for unobtrusive facilitation of everyday
environments with context acquisition technology. This property is especially relevant
in domains where preserving privacy is a central issue, e.g. care for elderly and
assisted living.
In our future work we will build on our basic techniques for identification,
positioning and event detection and develop algorithms that account for complex
interactions with multiple objects on surfaces, and that support object tracking across
multiple surfaces.
Acknowledgement
This research was partially funded by the Equator IRC, EPSRC GR/N15986/01
(http://www.equator.ac.uk).
References
1. Addlesee, M.D., Jones, A., Livesey, F., and Samaria, F.: ORL Active Floor. IEEE Personal
Communications, Vol.4, No 5, October 1997, pp. 35-41. IEEE, Piscataway, NJ, USA.
2. Addlesee, M.D., Curwen, R., Hodges, S., Newman, J., Steggles, P., Ward, A., Hopper, A.:
Implementing a Sentient Computing System. IEEE Computer Magazine, Vol. 34, No. 8,
August 2001, pp. 50-56.
3. Brumitt, B., Krumm, J., Meyers, B., and Shafer, S.: Ubiquitous Computing and the Role of
Geometry. IEEE Personal Communications, August 2000.
4. Crabtree, A., Hemmings, T. and Rodden, T.: Pattern-based Support for Interactive Design in
Domestic Settings. Technical Report Equator-01-016, University of Nottingham, The
School of Computer Science and IT, December 2001.
http://www.equator.ac.uk/papers/Authors/crabtree.html
5. Dey, A.K.: Providing Architectural Support for Building Context-Aware Applications.
Ph.D. thesis, December 2000, Dr. Gregory D. Abowd (advisor), College of Computing,
Georgia Institute of Technology
6. Georgia Institut of Technology. The Aware Home Research Initiative.
http://www.cc.gatech.edu/fce/ahri/
7. Headon, R. and Curwen R.: Ubiquitous Game Control. UBICOMP 2001 Workshop on
Designing Ubiquitous Computing Games. Atlanta, 2001.

8. Kidd, C., Orr, R., Abowd, G.D., Atkeson, C.G., Essa, I.A., MacIntyre, B., Mynatt, E.,
Starner, T.E., and Newstetter, W.: The Aware Home: A Living Laboratory for Ubiquitous
Computing Research. Proceedings of the Second International Workshop on Cooperative
Buildings - CoBuild'99. Position paper, October 1999.
9. Kistler force plate. http://www.kistler.com/
10.Konomi, S., Müller-Tomfelde, C., Streitz, N.: Passage: Physical Transportation of Digital
Information in Cooperative Buildings. In: Streitz, N., Siegel, J., Hartkopf, V., Konomi, S.
(Eds.), Cooperative Buildings - Integrating Information, Organizations, and Architecture.
Proceedings of the Second International Workshop (CoBuild'99). LNCS 1670. Springer:
Heidelberg. pp. 45 -54.
11.Krumm, J., Harris, S., Meyers, B., Brumitt, B., Hale, M., and Shafer S.: Multi-Camera
Multi-Person Tracking for EasyLiving. IEEE Workshop on Visual Surveillance, July 2000.
12.Lancaster University, Embedded load sensing project,
http://www.comp.lancs.ac.uk/~albrecht/load/
13.Orr, R. J. and Abowd, G. D.: The Smart Floor: A Mechanism for Natural User Identification
and Tracking. Proceedings of CHI 2000 Human Factors in Computing Systems (April 1–6,
2000, The Hague, Netherlands), ACM/SIGCHI.
14.Paradiso, J, Abler, C., Hsiao, K., Reynolds, M.: The Magic Carpet: Physical Sensing for
Immersive Environments. In Late-Breaking/Short Demonstrations of CHI'97, pp.277-278.
ACM, USA.
15.Schilit, B.N., Adams, N.I., Want, R.: Context-Aware Computing Applications. Proceedings
of the Workshop on Mobile Computing Systems and Applications, Santa Cruz, CA,
December 1994. Pages 85-90. IEEE Computer Society.
16.Schmidt A., Laerhoven, K., Strohbach, M., Gellersen, H.: Ubiquitous Interaction - Using
Surfaces in Everyday Environments as Pointing Devices. Paper submitted to 7th ERCIM
Workshop "User Interfaces For All", 2002.
17.Ward, A. J., Hopper, A.: A New Location Technique for the Active Office, IEEE Personal
Communications, vol. 4, pp. 42-47, 1997.