Conference PaperPDF Available

BeaStreamer-v0.1: a new platform for Multi-Sensors Data Acquisition in Wearable Computing Applications


Abstract and Figures

In this paper, we present BeaStreamer-v0.1, a new wearable computing platform designed fusing the Beagleboard hardware platform and the GStreamer software platform. The device has been designed for monitoring a variety of day-to-day activities and to be used as a 24/24h digital personal assistant.BeaStreamer-v0.1 can acquire data collected from multiple sensors in controlled and uncontrolled environments. The benefits of using BeaStreamer-v0.1 are multiple. First, the small size of the Beagleboard allows to use a really portable computer device. In addition, Beagleboard ensures laptop-like performances despite its dimensions. Using GStreamer makes managing the parameters in the acquisition of many different media types simple and allows to joint the acquisition of different types of data under a unique and compact framework. We demonstrate how the acquisition of audio, video and motion data can be easily performed by BeaStreamer-v0.1 and we point some highlights in the computational power of the system, some of them to be exploited as future lines of the work.
Content may be subject to copyright.
BeaStreamer-v0.1 : a new platform for Multi-Sensors Data
Acquisition in Wearable Computing Applications.
Pierluigi Casale, Oriol Pujoland Petia Radeva
Computer Vision Center, Campus UAB, Edifici O , Bellaterra, Spain
Dept. of Applied Mathematics and Analysis, University of Barcelona, Barcelona, Spain
In this paper, we present BeaStreamer-v0.1, a
new wearable computing platform designed fus-
ing the Beagleboard hardware platform and the
GStreamer software platform. The device has been
designed for monitoring a variety of day-to-day
activities and to be used as a 24/24h digital per-
sonal assistant.BeaStreamer-v0.1 can acquire data
collected from multiple sensors in controlled and
uncontrolled environments. The benefits of using
BeaStreamer-v0.1 are multiple. First, the small
size of the Beagleboard allows to use a really
portable computer device. In addition, Beagle-
board ensures laptop-like performances despite its
dimentions. Using GStreamer makes managing the
parameters in the adquisition of many different me-
dia types simple and allows to joint the acquisition
of different types of data under a unique and com-
pact framework. We demonstrate how the adqui-
sition of audio, video and motion data can be eas-
ily performed by BeaStreamer-v0.1 and we point
some highlights in the computational power of the
system, some of them to be exploited as future
lines of the work.
Keywords: Wearable Sensors, BeagleBoard, Social
Sensors, Multimodal Data Fusion, Pattern Recog-
nition Applications.
1 Introduction
In a recent interview to CNN, Gordon Bell (from
Microsoft Research) tells how and why he had been
recording every single event in his life over the
last decade. He carried around video equipment,
cameras and audio recorders to capture conversa-
tions, trips and any kind of experiences. Bell says
that this huge amount of data (more than 350 Gi-
gaBytes, not including the streaming audio and
video) is a replica of his biological memory. This
digitized eMemory is never forgotten. Microsoft is
working on a SenseCam [1], shown in Figure 1.
SenseCam is a camera that can be worn around the
person’s neck and automatically captures every de-
tail of daily life with photos.
Figure 1: Images of the SenseCam, from [1].
In the same direction, Intel has been working on
the “Everyday Sensing and Perception” project[2].
A team of twelve researchers has been working for
three years on the 90/90 Challenge, i.e. in build-
ing a real-time system for egocentric recognition of
handled objects that is accurate at 90% over 90%
of our days.
These examples clearly show an increasing in-
terest in developing perception-based systems ca-
pable of monitoring a variety of day-to-day ac-
tivities, both in the research community and in
the industry as well. A system being aware of
both context and activities during daily life not
just would be able to give assistance in memory-
retrieval tasks, but also for real-time assistance to
not completely self-sufficient people.
In this paper, we present the first version of
BeaStreamer, a wearable system for multi-sensors
data acquisition and analysis that, in experiences
similar to [1] and [2], could be successfully used as
a 24/24h digital personal assistant. Using wearable
devices such BeaStreamer-v0.1 opens the oppor-
tunity to define new use cases, such as healthcare
monitoring in patients rehabilitation or studying of
social behaviour of people.
The article is organized as follows. In Section
2, we will describe the system and its parts, both
hardware and software. In Section 3, we will show
some examples of data acquisition and analysis
with BeaStreamer-v0.1. Finally, we will discuss
conclusions and future works.
2 BeaStreamer-v0.1
Beastreamer-v0.1 is a wearable system designed
for real-time multi-sensors data acquisition. In this
work we use it for acquiring audio, video and mo-
tion signals, but its capabilities are not restricted to
these data types. Any kind of data flow might be
acquired from the system and stored in memory. In
Figure 2(a) we show the system disassembled on a
table, showing all the components. In Figure 2(b)
we show a tester wearing the system.
The system can be easily brought in one hand or
in a little bag around the users waist. The audio
and video dataflows are acquired using a standard
low-cost webcam that can be hooked to the shirt
(a) (b)
Figure 2: (a) The BeaStreamer-v0.1 system;
(b) The BeaStreamer-v0.1 system worn by a tester.
just down the neck or at chest level. An Arduino-
based bluetooth accelerometer, can be put in the
pant pocket or in the shirt pocket. Audio and video
data are acquired via GStreamer, motion data are
acquired via bluetooth. Although at the moment,
the main functionality of the system is data acqui-
sition, the system has been designed also for data
The core of the system is based on Beagleboard,
an OMAP-based board with high computational
power. The system is equipped on-board with a 4
Gigabytes SD-Card where both operating system
and data acquired are stored. In the next sections,
we describe the hardware components, the devel-
opment environment and finally, the operating sys-
tem and the application software running on the
2.1 The Hardware Core : BeagleBoard.
The BeagleBoard (BB)[3], shown in Figure 3, is
a low-power, low-cost single-board computer pro-
duced by Texas Instruments (TI).
With open source development in mind, BB has
been developed to demonstrate the potential of TI’s
OMAP3530 system-on-chip, though not all OMAP
functionalitites are available on the board. The BB
sizes approximately 80mmx80mm and it provides
all the functionalities of a basic computer.
The OMAP3530 system-on-chip includes an
ARM Cortex-A8 CPU at 600 MHz which can run
Windows CE or Linux, a TMS320C64x+ DSP for
Figure 3: Beagleboard front view, from [3].
accelerated video and audio codecs, and an Imag-
ination Technologies PowerVR SGX530 GPU to
provide accelerated 2D and 3D rendering that sup-
ports OpenGL ES 2.0. Built-in storage and mem-
ory is provided through a Package on Package chip
that includes 256MBytes of NAND flash memory
and 256MBytes of RAM. The board carries a sin-
gle SD/MMC connector, supporting a wide vari-
ety of device such as WiFi Cards, SD/MMC Mem-
ory Cards and SDIO Cards. One interesting feature
of the OMAP3530 is the possibility of booting the
processor from SD/MMC card.
Video output is provided through separate S-
Video and HDMI connections. A 4-pin DIN con-
nector is provided to access the S-Video output
of the BeagleBoard. This is a separate output
from the OMAP processor and can contain differ-
ent video output data from what is found on the
DVI-D output. The BB is equipped with a DVI-D
connector that uses an HDMI connector. It does
not support the full HDMI interface and it is used
to provide the DVI-D interface only.
Two USB ports are present on the board. Both
ports can be used as host ports, using High Speed
USB devices conform to USB 2.0 protocol, using
a maximum of 500 mA to power the host device. If
additional power is needed or multiple devices as
mouse, keyboard and USB mass storaged devices
must be used, one USB port can be used as OTG
( On-The-Go ) port to drive an self-powered USB
hub. The USB OTG port can be also used to power
the board from a standard external USB port. If
both USB ports need to be used, there exists an
additional 5 mm power jack to power the board.
DC supply must be a regulated and clean 5 Volts
supply. The board uses up to 2 Watts of power.
Beagleboard presents on board a populated RS-
232 serial connection where a serial terminal is
present. Using the terminal, it is possible to set the
boot parameters and the size of the video buffer.
Furthermore, a 14-pins JTAG connection is present
onboard to facilitate the software development and
debugging on-board using various JTAG emula-
tors. Two stereo 3.5mm jacks for audio input and
output are provided. An option for a single 28 pin
header is provided on the board to allow the con-
nection of various expansion cards. Due to multi-
plexing, different signals can be provided on each
pin providing more that 24 actual signal accesses.
This header is not populated on the BB and, de-
pending on the usage scenario, it can be populated
as needed. Because of the efficient power con-
sumption, the board requires no additional cooling.
Typical usage scenario for the BB are shown
in Figure 4. BB might be considered a laptop
substitute. There are many projects using BB in
Figure 4: Tipical Usage Scenarios for Beagle-
board, from [3].
robotic applications ([4], [5]). Nevertheless, up
to now, there is no literature using BB as a wear-
able device, despite of the low dimensions of the
board. The major issue of using BB in wearable
applications is the need of a portable power supply
source. In our applications, we use an A.C. Ryan
MobiliT external USB battery at 3400mAh, allow-
ing 4hours of autonomy for the system in complete
2.2 The Motion Sensor : Arduino
Arduino ([6]) is an open-source electronics pro-
totyping platform based on flexible, easy-to-use
hardware and software. Arduino can sense the en-
vironment by receiving input from a variety of sen-
sors and can affect its surroundings by controlling
lights, motors, and other actuators. The microcon-
troller on the board is programmed using the Ar-
duino programming language and the Arduino de-
velopment environment. The boards can be built
by hand or purchased preassembled. The software
can be downloaded for free. Although Arduino was
built for artists and hobbysts, there are many peo-
ple working on real electronic interactive projects,
thanks to the rapid prototypization Arduino allows.
We prototype a Bluetooth-based accelerometer us-
ing the Arduino board, an analogic ADXL 345 ac-
celerometer and a BlueSMiRF Gold Bluetooth mo-
2.3 The Development Side : OpenEmbed-
ded + ˚
Openembedded (OE) offers a complete cross-
compiler environment and allows developers to
create complete Linux Distributions for embedded
systems. OE offers different kernels for the BB.
All kernels come with several patches and the sup-
port of the BB hardware is not perfect yet. Figure 5
summarizes the current status of hardware support,
where it is possible to see that all the features of
OMAP processors available in the BB are actually
available for the use.
Figure 5: Current Status of Hardware Support for
BB in OE, from [3].
The Linux Kernel 2.6.28r12 runs on our sys-
tem. This particular Linux Kernel has V4L2 (Video
for Linux 2) drivers, allowing to plug in the sys-
tem almost every Linux-compatible webcam. Fur-
thermore, it contains BlueZ, the official Bluetooth
protocol stack. Using this kernel version, several
problems appear when using the DSP on the board.
At the moment, all the problems related to DSP
have been officially resolved in the Linux kernel
version 2.6.29 but there exist concerning issues re-
garding the use of USB ports. For that reason, we
use the “old” kernel release leaving aside, provi-
sionally, the DSP functionalities.
The ˚
om Distribution (AD) is the Linux
Distribution running on the board. AD if a specific
Linux distribution for embedded systems. A com-
plete image of AD can be built using OE or with
an online tool [9], where it is possible to choose the
packages to be installed in the system. In the distri-
bution we build, we include a toolchain for devel-
oping source codes on board. We install the arm-
gcc compiler, arm-g++ compiler and the Python-
Numpy development environment. In addition, we
build the GStreamer and the OpenCV packages, as
we will explain in the next section.
2.4 The Software Side: GStreamer +
GStreamer is a framework for creating stream-
ing media applications. The GStreamer frame-
work is designed to make easy writing applica-
tions handling audio/video streaming. Neverthe-
less, Gstreamer is not restricted to audio and video,
and it can process any kind of data flow. One of the
the most obvious uses of GStreamer consist in us-
ing it to build a media player. GStreamer already
includes components for building a media player
that can support a very wide variety of formats,
including MP3, Ogg/Vorbis, MPEG-1/2, AVI, and
The main advantages of GStreamer are that the
software components, called plugins, can be mixed
and matched into arbitrary pipelines so that it is
possible to write complete streaming data editing
applications. Plugins can be linked and arranged
in a pipeline. The GStreamer core function is to
provide a framework for connecting plugins, for
data flow management and for media type han-
dling and negotiation. Using GStreamer, perform-
ing complex media manipulations becomes very
easy and it integrates an extensive debugging and
tracing mechanism. In BeaStreamer-v0.1, we use
a pipeline for acquiring audio and video from we-
bcam, with the possibility to encode the dataflow
with the request quality and the resolution and the
possibility to change the acquisition parameters at
run time.
In addition, we compile on BeaStreamer-v0.1
the well-known OpenCV libraries and its Python
3 Experiments with BeaStreamer-
In this section, we show some experiment per-
formed with BeaStreamer-v0.1 to demonstrate its
capabilities. In Figure 6, we show six sequential
photos taken wearing BeaStreamer-v0.1, walking
in the street. Using GStreamer, we take photos
with a framerate of one photo/second with a res-
olution of 320x240 pixels, compressed in jpeg for-
mat. At the same time, in a separate thread, we
record a countinous audio flow from the webcam
microphone, sampled at 44100 samples/s and com-
pressed in ogg format. GStreamer allows setting
online the parameters of acquisition making simple
to change the resolution of photo and the encoding
audio quality.
In order to get an estimate of the autonomy of
the system, we record an audio/video stream com-
pressed in ogg format and receive motion data from
the bluetooth accelerometer. We are able to record
up to 4 hours of audio, video and motion data.
Using OpenCV, we setup a face detector running
on photos adquired sequentially with GStreamer.
The face detector can compute detections at a
Figure 6: A sequence of photos taken with
framerate of 5-10 frames/second depending of the
images resolution, without using DSP. An example
of faces successfully detected is shown in Figure 7.
Using images with resolution of 80x60 pixels, the
face detector can scan the image in less than 100
ms and detect faces in 200 ms.
Figure 7: Face Detector running on BeaStreamer-
Finally, in Figure 8 we show how BeaStreamer-
v0.1 receives motion data. The acceleration ana-
log values are converted in 10-bit values from Ar-
duino ADC (Analog-to-Digital Converter) at 40Hz,
stored in a buffer and sent every second via Blue-
tooth as UNICODE characters with a label show-
ing the axis. BeaStreamer-v0.1 receives the data
and stored them in a text file.
4 Conclusions and Future Works
In this paper we presented BeaStreamer-v0.1, a
new platform for multi-sensors data acquisition.
BeaStreamer-v0.1 is small and easy to bring, al-
Figure 8: Acquisition of motion data.
lowing its use in wearable computing applications,
for controlled and uncontrolled environments. We
showed that different types of data can be easily ac-
quired joining the potentiality of the Beagleboard
and GStreamer.
The Beagleboard allows to connect different types
of sensors communicating via Bluetooth or via the
principal types of communication protocols imple-
mented in the OMAP processor. GStreamer pro-
vide a framework to manage different types of data
flows using a single and coherent environment.
At the moment, just a basic face detector has been
developed on the system to demonstrate its capa-
bilities. Furthermore, the computational power of
the system will increase as soon as the DSP side of
the OMAP will be completely operative.
Finally, we consider that unifying Bluetooth and
general sensors acquisition under GStreamer will
provide a powerful and complete platform for gen-
eral multi-sensors data acquisition and analysis.
This work is partially supported by a re-
search grant from projects TIN2006-15308-
C02, TIN2009-14404-C02, FIS-PI061290 and
00018), MI 1509/2005.
Special thanks to the “Computer Vision
Group/DSPLab” of Instituto di Fisiologia Clinica,
Consiglio Nazionale di Ricerche ( CNR ), Pisa,
Italy for their help in the first stage of system
[1] Microsoft Research, Introduction to Sense-
[2] X. Ren, M. Philipose, ”Egocentric Recog-
nition of Handled Objects: Benchmark
and Analysis”, Proc. of 1th Workshop on
Egocentric Vision,
[3], Beagleboard Technical Doc-
[4] Home Brew Robotic Club, A tutorial on set-
ting up a system to do image processing with a
[5] Beaglebot, Beagle powered robot
[6] , Arduino Technical Documenta-
[7], Open-
Embedded Official Manual,
[8] The ˚
om Distribution,
om Documentation,
[9] The ˚
om Distribution, Online
om Building, http://www.angstrom-
[10] GStreamer, GStreamer
Technical Documentation ,
With the enrichment of technologies, humans want to maximize automation by reducing the manpower and time, Human Activity Recognition (HAR) has a heterogeneous broad range of significant applications such as health care, theft detection, work monitoring in an organization and detecting emergencies. Various machine learning (ML) classification algorithms are applied on publicly available HAR datasets to recognize human activities in the literature. In this work, we have identified different HAR datasets involving different levels of activities and the methods for acquisition of data. We have also done a detailed review of HAR approaches with the implementation of various ML classifiers along with specific future directions in this area.
Full-text available
Constructing a binary tree, the Huffman algorithm introduced the method of text compression that helps to reduce the size keeping the original message of the file. Nowadays, Huffman-based algorithm assessment can be measured in two ways; one in terms of space, another is decoding speed. The requirement of memory for a text file is going to be reasonable while the time effectiveness of Huffman decoding is being more significant. Meanwhile, this research is introducing the adjacent distance array with Huffman principle as a new data structure for encoding, and decoding the Bengali text compression using transliterated English text. Since the transliterated English text accommodates to reduce the unit of symbols accordingly, we transliterated the Bengali text into English and then applied the Huffman principle with adjacent distance array. By calculating the ASCII values, adjacent distance array is used to save the distances for each adjacent symbols. Apart from the regular Huffman algorithm, a codeword has produced by traversing the whole Huffman tree for a character in case, respectively adopting the threshold value and adjacent distance array can skip the lengthy codeword and perform the decoding manner to decode estimating the distances for all adjacent symbols except traversing the whole tree. Our findings have acquired 27.54% and 20.94% compression ratios for some specimen transliterated Bengali texts, as well as accomplished a significant ratio on different corpora.
We propose a variant of Principal Component Analysis (PCA) that is suited for real-time applications. In the real-time version of the PCA problem, we maintain a window over the most recent data and project every incoming row of data into a lower-dimensional subspace, which we generate as the output of the model. The goal is to reduce the reconstruction error of the output from the input and to retain major components pertaining to previous distributions of the data. We use the reconstruction error as the termination criteria to update the eigenspace as new data arrives. We then propose two variants of this algorithm that are progressively more time efficient. To verify whether our proposed model can capture the essence of the changing distribution of large datasets in real time, we have implemented the algorithms and compared performance against carefully designed simulations that change distributions of data sources over time in a controllable manner. Furthermore, we have demonstrated that proposed algorithms can capture the changing distributions of real-life datasets by running simulations on datasets from a variety of real-time applications, e.g., localization, activity recognition, customer expenditure, and so forth. Results show that straightforward modifications to convert PCA to use a sliding window of datasets do not work because of the difficulties associated with determination of optimal window size. Instead, we propose algorithmic enhancements that rely on spectral analysis to improve dimensionality reduction. Results show that our methods can successfully capture the changing distribution of data in a real-time scenario, thus enabling real-time PCA.
Conference Paper
Recognizing objects being manipulated in hands can provide essential information about a person's activities and have far-reaching impacts on the application of vision in everyday life. The egocentric viewpoint from a wearable camera has unique advantages in recognizing handled objects, such as having a close view and seeing objects in their natural positions. We collect a comprehensive dataset and analyze the feasibilities and challenges of the egocentric recognition of handled objects. We use a lapel-worn camera and record uncompressed video streams as human subjects manipulate objects in daily activities. We use 42 day-to-day objects that vary in size, shape, color and textureness. 10 video sequences are shot for each object under different illuminations and backgrounds. We use this dataset and a SIFT-based recognition system to analyze and quantitatively characterize the main challenges in egocentric object recognition, such as motion blur and hand occlusion, along with its unique constraints, such as hand color, location prior and temporal consistency. SIFT-based recognition has an average recognition rate of 12%, and reaches 20% through enforcing temporal consistency. We use simulations to estimate the upper bound for SIFT-based recognition at 64%, the loss of accuracy due to background clutter at 20%, and that of hand occlusion at 13%. Our quantitative evaluations show that the egocentric recognition of handled objects is a challenging but feasible problem with many unique characteristics and many opportunities for future research.
A tutorial on setting up a system to do image processing with a BeagleBoard
  • Home Brew
  • Robotic Club
Home Brew Robotic Club, A tutorial on setting up a system to do image processing with a BeagleBoard., [5] Beaglebot, Beagle powered robot
Introduction to Sense- Cam
  • Microsoft Research
Microsoft Research, Introduction to Sense- Cam, us/um/cambridge/projects/sensecam/
Beagleboard Technical Documentation
  • Beagleboard
  • Org, Beagleboard Technical Documentation,
Arduino Technical Documentation , [7], Open- Embedded Official Manual
  • Arduino
  • Cc, Arduino Technical Documentation, [7], Open- Embedded Official Manual,