Wireless Body Sensor Enhanced Tracking for Extended InHome Care

Ani Nahapetian, Sabah Chaudhry, Foad Dabiri, Tammara Massey, Hyduke Noshadi, Majid
Sarrafzadeh
Computer Science Department UCLA
{ani, dabiri, tmassey, hyduke, majid}@cs.ucla.edu

ABSTRACT
In this paper, we present a medical embedded tracking
system that harnesses local, efficient sensor data to trigger
and enhance the accuracy and efficiency of tracking
applications, specifically targeted to medical and nursing
applications. The project involves enhancing the highly
parallelizable vision applications with low-level sensor data
that are continuously being gathered by various lightweight
devices. We leverage the systems hierarchical structure to
trigger cameras equipped with reconfigurable devices to
carry out various types of machine vision for medical
sensing and monitoring. The system is composed of
various heterogeneous embedded devices, with various
computation capabilities and power costs, and hence the
system is arranged in a hierarchical fashion in order to take
advantage of the low cost devices to wirelessly trigger the
operation and the actuation of the more capable but higher
cost vision systems. The experimentation provide
substantive data to confirm that the coordination of the two
types of sensors can have a large potential to decrease
cycles of execution on the FPGA, and hence allow for
power savings or the addition of tailoring of the vision
applications.

Keywords
Medical Embedded Sensing Systems, Machine Vision,
Extended In-Home Care.

1. INTRODUCTION
The increase of wellness healthcare programs and patient
management emphasizes the reality that patient information
can be readily and continually made available for
monitoring of patient health. The current proliferation of
broadband wireless services, along with more powerful and
convenient handheld devices, is helping introduce real-time
monitoring and guidance for a wide array of patients. Lowcost sensors and wireless systems can now create a
constantly vigilant and pervasive monitoring capability at
home and in conventional point-of-care environments (e.g.,
primary care physician offices, outpatient clinics,
rehabilitation centers). A variety of devices and sensors
exist, which can be unobtrusively incorporated into patient

life and which can then wirelessly transmit data to

caregivers and healthcare professionals.
In our system, we target the development of a wireless
medical embedded sensing infrastructure for the purpose of
extending in-home care for the elderly. The continuation of
normal living patterns for the elderly can have a large
impact on the quality of their lives and the financial burden
of their care, as the transition to a nursing home facility is
often a one-way process.
The purpose of this system is not to target any specific
medical condition, but instead to develop a useful
infrastructure that can be used for various conditions. The
system we present harnesses local, efficient sensor data to
trigger and enhance the accuracy and efficiency of tracking
applications. There exists a range of medical conditions
that can be robustly addressed with this type of system.
Specifically, we have considered Alzheimers disease early
onset detection and disorientation monitoring. Alzheimers
disease onset detection is often a difficult task since the
disease is one where patients have good days and bad days,
and with only the occasional visit to the physician there can
be a long delay between disease onset and proper
diagnosis. Additionally, the data that patients provide is
qualitative and possibly even inaccurate. Our system aims
to aid in this diagnosis dilemma, by capturing information
about patient disorientation.
Additionally after the
diagnosis phase, patient disorientation can be monitored
and caregivers can be alerted to these instances.
Additionally, we have considered the use of gait
monitoring in the elderly to study the cause of falls as well
as prevent their occurrence [20][18]. Simple falls can have a
huge impact on the lives of the elderly, since they often
result in serious and life-threatening health condition and
hospitalization. Gate monitoring can be used to determine
the common causes of falls in the elderly, and even work to
prevent them from happening if immediate aid is available.
Additionally, it can be used to alert caregivers, which can
allow patients to receive faster care and more specific
diagnosis. Researchers at Virginia Tech have looked into
the determination of the possible causes of falls in the
elderly [8]. In contrast with our system which is intended
for continual and nondisruptive use with a limited number
of sensors with a small form factor placed on or near the

body, their system is quite large and cumbersome and is

not intended for everyday use.
Our system is composed of med nodes, which are small
constrained wireless sensing devices and video sensors
equipped with FPGAs. The med nodes, which are in this
case equipped with accelerometers and placed on the knee
of the patient, send low level data measurements to the
video sensor systems, which triggers them to adapt the
operation of their tracking application. The overall system
carries our tracking and data inference based on its sensor
data. The combination of the different types of sensors
allows for enhanced sensing, error prevention, and
increased computation efficiency.
In the remainder of this paper, we present the related work
and the broad implications of research in this area. We also
present in detail the system infrastructure and components,
with a special focus on the vision application that has been
developed for the FPGA for this system. Finally we present
the experimental results that show the benefits of fusing
two different types of sensor data in a hierarchical fashion.

and hibernation. Power aware computation on the

lightweight systems was achieved by selecting modes of
operation that were most efficient for a task. Several
power management techniques for lightweight embedded
systems are discussed in [2], but they are only explored on
a theoretical level.
Similarly, dynamic power optimality has been achieved in
sensor networks through the use of rotating cluster heads
[4]. These systems minimize power utilization using a
distributed approach to maximize the entire lifetime of the
system [4][3]. Our work specifically reduces power
consumption on a lightweight embedded system through
using context sensitive information to improve the
algorithm for a lightweight medical application.
Additionally, other systems have also explored using
context to reduce power consumption in passive or groggy
wakeup schemes that use state-based sensing to achieve
better power utilization [5][1][7]. Our work differs in that
it specifically looks at minimizing power utilization by
improving the accuracy of the vision algorithm.

2. RELATED WORK

3. BROAD IMPLICATIONS

There have been two related projects that have explored a

specific aspect of the problem that we have addressed. The
GALORE Project developed a network of cameras that
collaborate and reconfigure their FPGAs to meet the realtime needs of the vision applications, but with the
adaptability offered by loading different computationally
intensive applications onto the FPGA [10][11][15]. The
coordination of services and computational reconfiguration
were explored in a tiered system of motes and cameras
[13][14][16].

The larger vision of our system is to personalize care for

the elderly through early detection and diagnosis of
conditions and the extension of in-home care. A key
requirement, however, is the pervasive yet unobtrusive
incorporation of medical care into their everyday life.
There is a growing, aging population as shown in Figure 1,
taken from the United Nations world population prospects.
There is an urgent need to address the medical conditions
of this growing segment of the population. Systems that
address the quality of life issues in addition to the strictly
medical issues will have a much larger potential to impact
lives. Additionally, there is large potential for cost savings,
since these embedded sensing devices will be more cost
efficient than constant healthcare professional or family
member monitoring.

Additionally, work has been done on habitat monitoring

that has employed a tiered structure for its processing
[17][19].

The Assisted Cognition Project [21] specifically explored

the aiding of Alzheimers patients, using AI systems to
support and enhance the independence and quality of life of
patients. The goal of the Assisted Cognition project was to
develop a novel computer system that would enhance the
quality of life of people suffering from Alzheimers disease
and similar cognitive disorders. Assisted Cognition systems
use ubiquitous computing and artificial intelligence
technology to replace some of the memory and problemsolving abilities that have been lost by an Alzheimers
patient. Two concrete examples of the Assisted Cognition
systems are an Activity Compass that helps reduce spatial
disorientation both inside and outside the home, and an
Adaptive Prompter that helps patients carry out multi-step
everyday tasks.
Additionally, there has been work done in the area of
power minimization for lightweight embedded systems. [6]
examined the power usage of lightweight mobile systems
during various modes, data transmission, data collection,

A key challenge involves the fusing of the data from

heterogeneous embedded sensing devices from both body
area networks and environmental networks, and when the
data is both physiological and psychological in nature.
The increased and standardized availability of this data can
have a profound impact on patient care. Key predictors can
be used to monitor patient status, and in turn allow for the
extension of in-home care of patients, thus increasing their
quality of life and decreasing the financial burden of their
care.

window size for which the image processing algorithms are

run.
The application tracks a person through a space, while
matching the tracking range to the level of mobility of the
person. If there is low mobility, which is most often the
case, then the algorithm on the FPGA works more
efficiently doing processing on only a part of the frame.
When there is high mobility, the algorithm is forced to
increase the range of pixels on which the processing is
performed, resulting in a decreased level of efficiency.
The advantage of this approach is that it allows for the
elimination of unnecessary computation and even memory
usage, thus allowing for other processing or even the
shutting down of system components.
Figure 1. Growing Aging Population [Source: United
Nations, World Population Prospects]
However, aside from the challenge of making the sensor
devices feasible and usable, continuous monitoring of
various parameters introduces key problems with data
management. Due to system constraints, the large amounts
of data cannot be transmitted to a physician or even a base
station. Instead local and adaptable data filtering and
clustering have the potential to handle the large amount of
data captured, stored, and transmitted. Techniques can
analyze the data to infer key health measures and also can
adaptively probe or discard the medical data. Another
challenge that is specific to multiple sensor type monitoring
is the need to determine the correlations and the
dominances between the various medical parameters.
However, with the data management techniques we
propose, this sort of analysis can be done by the system,
specifically for each patient, and can even be made
adaptable over time to match the changing condition of the
patient.

4. SYSTEM INFRASTRUCTURE
4.1 Overall System Functionality
Generally, our medical embedded tracking system
harnesses local, efficient sensor data to trigger and enhance
the accuracy and efficiency of a machine vision tracking
applications. Med nodes, which are equipped with
constrained computational devices and sensors, here
specifically accelerometers, wirelessly transfer the sensor
data to the mote that is connected to the FPGA. The data is
then passed to the FPGA through a UART port. The FPGA
uses the sensor readings to determine variance of the
acceleration and hence determine the sensing range in the
frame. We developed a color detection based vision
sensing and tracking application. Based on the variance of
the accelerometer readings, the application adjusts the

The system is used to demonstrate the effectiveness of the

data fusion of physiological and psychological data. In this
case, the physiological data comes from the med nodes
accelerometer, and the psychological data comes from the
vision sensor, which is able to determine disorientation and
other similar psychological states. Additionally, the system
provides evidence for the effectiveness of a hierarchical
structure that can be broadly applied in similar
heterogeneous embedded sensing systems.
The project involved the development of the VHDL code
for the tracking application. Additionally, we coordinated
the wireless communication between the Berkeley Mica2
dote motes in the med nodes with the FPGA connected to
the vision sensor. Finally, the system required the
extraction and the wireless communication of the
accelerometer values from the med node devices. In the
future, we intend to enhance the system by doing more
node level data processing to decrease the burden of
communicating the data wirelessly.
In the remainder of this section we will present all the
systems components and their interaction with other
components.

4.2 Med Node

Med nodes are stand-alone components equipped with
processing units and batteries. They support various types
of sensors for obtaining physiological readings from the
human body. They can be adapted to suite a range of
applications and can be adapted to perform multiple types
of operations. Furthermore, they support a variety of
analog and digital sensors.
The med node architecture consists of an accelerometer
sensor, sensor adaptation board, Crossbow Mica2Dot
processing element developed by UC Berkeley, and a
power supply (coin cell battery). Figure 1 illustrates the
med node we used in our system.

algorithms are inherently parallelizable, since similar

processing must be done on various individual or groups of
pixels. Therefore, hardware implementation for these
algorithms enables a significant increase in system operation
speed which facilitates the real-time application needs. Also,
since FPGAs are reconfigurable architectures, these
algorithms can be updated/modified through reconfiguring
of FPGAs without redeployment of the system. In our
proposed system we have used VHDL as the high level
description language.

4.4 Video Sensor and Vision Applications

Figure 2. Med Node Package
We connected an ultra compact three-axis linear
accelerometer to the med nodes. The LIS3L02AL is a lowpower 3-axis linear capacitive accelerometer that includes a
sensing element and an IC interface that is able to take the
information from the sensing element and to provide an
analog signal.
The analog output is then fed to ADC channels on
Mica2Dot (analog to digital converter) and read by the
Atmega128 processor on the Mica2Dot board.
The LIS3L02AL recommended frequency range spans
from DC up to 2.0 KHz, which in our application is much
larger than our value. The sensor board is responsible for
power supply decoupling and noise reduction. We omitted
the final stage buffers in our system since the input
resistance of the ADC channels on the Mica2Dot is high
enough to lower the loading effect.
The period which we sample the acceleration values is set
to 50ms which yields an approximately 20Hz sampling
frequency. The sampling frequency can be changed
adaptively by reconfiguring the device wirelessly.

4.3 Reconfigurable Component: FPGA

In addition to med nodes, our system is composed of video
sensors equipped with FPGAs, on which we developed our
tracking application. We implemented the algorithm on a
Xilinx Virtex-II pro FPGA board, which includes two IBM
PowerPC microprocessors. The PowerPC architecture is a
64-bit architecture with a 32-bit subset and a 32-bit data
interface with the on-chip memory (OCM) [9].
FPGAs or field programmable gated array has been used as
an alternative for digital components in a variety of
systems. The FPGA architecture consists of programmable
logic, reconfigurable interconnects and input/output
devices. The programmable logic components can be
configured to have the functionality of basic logic
components, such as ANDs and ORs, as well as more
complex combinational functions. FPGAs also have
memory elements such as flip flops and memory blocks.
The FPGAs enable parallelism to achieve high performance
with moderate power consumption. Image processing

The final component of our system is the video sensor and

the image processing that is done on the captured images.
We use 260X zoom color cameras manufactured by
Kocom. This camera has a 26x optical and 10x digital
zoom. It is currently in use for a variety of applications
including for surveillance purposes. It has S-video output
that we use to connect the camera to the FPGA board.
The image processing application is based on a color
tracking procedure. To detect only moving objects, frame
subtraction is performed. Since there is not enough space to
store past frames entirely, we store only the down sampled
version, which is 1/64 of the original frame. Then, for
every sixtieth frame saved, each pixel of the current frame
is subtracted from the corresponding pixel of the saved
frame.
Once background subtraction has been completed, all
pixels RGB values are compared to the small range of
values that are used to identify the person being tracked. In
this application we equipped each person with a green
marker that was used to identify them in space under most
lighting conditions.
The accelerometer data is used to enhance the window size
where the RGB comparison is carried out. Based on the
mobility information received, the algorithm varies the
window size for which the RGB comparison is carried out.
There is some account made for error in acceleration
measurement, by adding an extra buffer region to the
window size. Additionally processing is done to account
for the mapping of three-dimensional movement to the two
dimensional frame representation.
The window is centered on the person, or rather the green
marker, that is being tracked. If there are multiple regions
of the frame to which the processing maps, then the more
prominent region is used for analysis. This can result in
some error, but it can quickly be corrected during the
subsequent frame processing.

Limited Mobility

High Mobility

area, this graph shows a quadratic decrease as a function of

diameter.

Efficiency of Sensor Enhanced Tracking

With Sensor Enhanced Tracking

Full Frame Tracking

35000

Figure 3. Summary of the adaptation of the tracking

algorithm based on accelerometer data, where the green
points highlight the position of the person over time in the
frame

5. EXPERIMENTAL RESULTS
To verify that med node data successfully aided the
operation of the tracking application, we measured the
computation cycle execution savings for various different
window sizes, as dictated by the patients mobility level.
The first graph in Figure 4 shows that by decreasing the
processing pixel area, the cycles of execution are also
decrease in a linear fashion.
Efficiency of Sensor Enhanced Tracking

25000
20000
15000
10000
5000
0
0

100

200

300

400

500

Mobility of the Patient (In Terms of the Diameter of the Frame

Converage Range)

Figure 5. Computation cycles savings as a function of

patient mobility
The results demonstrate that incorporation of the med node
data provides a large degree of computation efficiency,
which can be mapped to power savings or even the
increased system functionality, during low mobility
periods.
Additionally, the mobility information allows for error
minimization in the tracking application, verification of
inferences using redundant data, and finally in the future
the compression and filtering of the large amount of data
that is the result of continuous sensing.

35000
30000
Num ber of Cycles

Nu m b er o f Cycles

30000

25000
20000

6. CONCLUSION

15000
10000
5000
0
0

50000

100000 150000 200000 250000 300000 350000

Pixel Area

Figure 4. Computation cycles savings as the pixel area

of the processing is decreased
The second graph in Figure 5 shows a quadratic decrease in
the number of computation cycles as a function of the
mobility of the patient, presented as the diameter of
patients movement in the frame. In the case of low
mobility, such as with limited movement in place, there is 2,887
cycles/frame of computation. With high mobility, such as a brisk
walk across frame, there is 14,791 cycles/frame of computation.
Without any sort of mobility information, 33,319 cycles/frame of
computation must be carried out for each frame. As expected

from the first graph which has a linear relationship with

In this paper, we presented a medical embedded tracking

system that harnesses local, efficient sensor data to trigger
and enhance the accuracy and efficiency of a patient
tracking applications. We discussed the broad implications
of developing such systems for extending in-home care for
the elderly, by developing pervasive yet unobtrusive
heterogeneous sensor data. The project involved the
triggering by med nodes placed on the body of cameras
equipped with reconfigurable devices to carry out various
types of machine vision for medical sensing and
monitoring. The experimentation demonstrated a quadratic
decrease in computation cycles if med node sensing data
was incorporated into the tracking application.

7. REFERENCES
[1] Benbasat, A.Y. and Paradiso, J.A. Groggy Wakeup -

Automated Generation of Power-Efficient Detection

Hierarchies for Wearable Sensors, in Proceedings of
the 4th International Workshop on Wearable and

Implantable Body Sensor Networks (BSN 2007),

Aachen, Germany, March 26-28, 2007, pp. 59-64.
[2] Estrin, D. Culler, D. Pister, K. Sukhatme, G.

Connecting the Physical World with Pervasive

Networks. IEEE Pervasive Computing. Jan 2002.
pp. 59-69.
[3] Jafari, R. Dabiri,F., Sarrafzadeh, S. An Efficient

Placement and Routing Technique for Fault-tolerant

Distributed Embedded Computing. The 11th IEEE
International Conference on Embedded and Real-Time
Computing Systems and Applications (RTCSA),
August 2005, Hong Kong.
[4] Mainwaring, A., Culler, D., Polastre , J. Szewczyk, R.,

Anderson, J. Wireless Sensor Networks for Habitat

Monitoring. ACM International Workshop on
Wireless Sensor Networks and Applications (WSNA
2002). pp. 88-97.
[5] Ma., H., Paradiso, J. The FindIT Flashlight:

Responsive Tagging Based on Optically Triggered

Microprocessor Wakeup. UbiComp 2002. pp. 160167.
[6] Raghunathan, V., Schurger, C., Park, S., Srivastiva, M.

Energy-aware Wireless Microsensor Networks.

IEEE Signal Processing Magazine. March 2002. pp
40-50.
[7] Shih, E., Bahl, P., Sinclair, M. Wake on wireless: An

Event Driven Energy Saving Strategy for Battery

Operated Devices. MobiCom 2002.
[8] Gait Monitoring Project

http://www.research.vt.edu/resmag/sc2003/whyfall.ht
m
[9] Virtex-II Pro and Vertex 2 Pro X Platform FPGAs

http://direct.xilinx.com/bvdocs/publications/ds083.pdf.
[10] Soheil Ghiasi, Ani Nahapetian, Hyun J. Moon, Majid

Sarrafzadeh, "Reconfiguration in Network of

Embedded Systems: Challenges and Adaptive
Tracking Case Study", Journal of Embedded
Computing (JEC), Vol. 1, No 1, 2004.
[11] Soheil Ghiasi, Hyun J. Moon, Ani Nahapetian, Majid

Sarrafzadeh, "Collaborative and Reconfigurable

Object Tracking", Kluwer Journal of Supercomputing,
Vol. 30, No 3, pp. 213-238, December 2004.
[12] Soheil Ghiasi, Hyun J. Moon, Majid Sarrafzadeh, "A

Networked Reconfigurable System for Collaborative

Unsupervised Detection of Events", Submitted to
ACM Transactions on Embedded Computing Systems
(TECS). Soheil Ghiasi, Hyun J. Moon, Majid

Sarrafzadeh, "Collaborative and Reconfigurable

Object Tracking", International Conference on
Engineering of Reconfigurable Systems and
Algorithms (ERSA), pp. 13-20, June 2003.
[13] Soheil Ghiasi, Hyun J. Moon, Majid Sarrafzadeh,

"Improving Performance and Quality thru Hardware

Reconfiguration: Potentials and Adaptive Object
Tracking Case Study", Workshop on Embedded
Systems for Real-Time Multimedia (ESTIMedia), pp.
149-155, October 2003.
[14] Soheil Ghiasi, Karlene Nguyen, Elaheh Bozorgzadeh,

Majid Sarrafzadeh, "On Computation and Resource

Management in Networked Embedded Systems",
International Conference on Parallel and Distributed
Computing and Systems, pp. 445-451, November
2003.
[15] Soheil Ghiasi, Karlene Nguyen, Majid Sarrafzadeh,

"Profiling Accuracy-Latency Characteristics of

Collaborative Object Tracking Applications",
International Conference on Parallel and Distributed
Computing and Systems, pp. 694-701, November
2003.
[16] Ani Nahapetian, Soheil Ghiasi, Majid Sarrafzadeh,

"Task Scheduling on Heterogeneous Resources with

Heterogeneous Reconfiguration Costs", International
Conference on Parallel and Distributed Computing and
Systems, pp. 916-921, November 2003.
[17] H. Wang, D. Estrin, L. Girod, "Preprocessing in a

Tiered Sensor Network for Habitat Monitoring", in

EURASIP JASP special issue of sensor networks, Vol.
2003, No. 4, pp. 392-401, March 15, 2003.
[18] R. Vidal and Y. Ma and S. Soatto and S. Sastry, "Two-

view Multibody Structure From Motion", to appear in

Intl. Journal of Computer Vision, 2003.
[19] V. Bychkovskiy, S. Megerian, D. Estrin, M.

Potkonjak, "A Collaborative Approach to In-Place

Sensor Calibration", In Proceedings of the 2nd
International Workshop on Information Processing in
Sensor Networks (IPSN'03), volume 2634 of Lecture
Notes in Computer Science, pages 301-316. SpringerVerlag Inc., 2003.
[20] A. Bissacco, P. Saisan, S. Soatto, "Modeling Human

Gaits With Subtleties", Proceedings of the IFAC

Symposium. on System Identification, August 2003.
[21] Assisted Cognition Project

http://www.cs.washington.edu/assistcog.