Diana Cristina Albu Embedded Systems

Michael Bramberger, Andreas Doblander, Arnold Maier, Bernhard Rinner Graz University of Technology Helmut Schwabach Austrian Research Centers, Seibersdorf

Smart cameras are fun Surveillance tasks are important in todays traffic I wanted something related to the subject but not something mentioned in class

It was a Research Feature for IEEE Computer society paper in February 2006 (Volume 39, Issue 2)
It presents the design of a smart camera as a fully embedded system, with application in traffic surveillance

Smart cameras are equipped with a highperformance onboard computing and communication infrastructure, combining in a single embedded device
video sensing processing communications

Networks of embedded cameras can potentially support more complex and challenging applications:
Smart rooms Surveillance Tracking Motion analysis

From Analog to Digital Cameras: 1st generation surveillance: analog equipment (closed circuit TV cameras transmitted video signal over analog lines) 2nd generation: digital back-end components; allow real time automated analysis of incoming data 3rd generation: complete digital transformation; video converted in digital domain at the camera and transmitted via a computer network; cameras can also compress video to save bandwidth

4th generation: intelligent cameras; perform lowlevel image processing operations on the captured frames onboard to improve video compression and intelligent host efficiency
But smart cameras

however most of the processing is done at a central unit directly perform highly sophisticated video analysis video sensing video processing communication designed as reconfigurable and flexible processing nodes with self-reconfiguration, self-monitoring, and self diagnosis capabilities.

Shift from a central to a distributed control surveillance system

Increase the surveillance systems functionality, availability, and autonomy Can react autonomously to changes in the systems environment Can detected events in the monitored scenes.

A static surveillance system configuration is no longer feasible!

scalable, embedded, high-performance, multiprocessor platform consisting of a

network processor a variable number of digital signal processors (DSPs)

commercial off-the-shelf software/hardware architecture was chosen

support fast prototype development achieve flexibility and performance at a reasonable price

Sensing unit

Processing unit (PU)

Monochrome CMOS image sensor delivers images with VGA resolution at up to 30 fps transfers images via a first-in, first-out (FIFO) memory to the PU Up to 10 Texas Instruments TMS320C64x DSPs
can deliver an aggregate performance of up to 80 GIPS while keeping the power consumption low

Communication unit

PCI bus couples the DSPs and connects them to the network processor

network processor: Intel XScale IXP425 establishes the connection between the processing and communication units controls internal and external communication

currently supports two interfaces for IP-based external communication: Wired Ethernet wireless Global System for Mobile Communications/general packet radio service (GSM/GPRS)

DSP framework runs on every DSP

provides an abstraction of the hardware and communication channels supports dynamic loading and unloading of application tasks manages the DSPs on-chip and off-chip resources algorithms on different DSPs use the service management facilities to dynamically establish connections to each other the DSP framework was built on Texas Instruments DSP/BIOS operating system.

SmartCam framework - runs on the network proc

an abstraction of the DSPs to ensure the application layers platform independence application layer uses the provided communication methods to exchange information
internal messaging to the DSPs external IP-based communication

application development by high-level interfaces to DSP algorithms and the DSP frameworks functions XScale processor runs standard Linux
only customization of the Linux kernel is the DSP kernel module
processor uses it to establish the connection to the DSPs via the PCI bus

Use the smart cameras to implement a distributed intelligent video surveillance system (IVS) Partition IVS into distributed logical groups (surveillance clusters) IVS
requires an assignment of cameras to a specific cluster dynamically and autonomously maps surveillance tasks into individual cameras depending on their resources and the systems current state

Tasks are implemented onto cameras using a mobile agent system (MAS) built atop the SmartCam framework Changes in the environment trigger a task mission
Quality of Service (QoS):
parameters include frame rate, transfer delay, image resolution, and video-compression rate levels can change over time due to user interactions or changes in the monitored environment (so novel IVS systems must include dedicated QoS management mechanisms)
camera supports combined power and QoS management (PoQoS) for distributed IVS systems PoQoS dynamically configures the power and QoS level of the cameras hardware and software to adapt to user requests and changes in the environment

Power awareness

Commercial off the shelf hardware components to test and evaluate the video surveillance system

1 cam consists of:

network processor several DSPs a CMOS image sensor

Platform: IXDP425 Intel Processor: IXP425 Xscale

On-chip support:

533 MHz 256 Mbytes of external memory four PCI slots Ethernet access multiple serial ports PCI host controller Atemes network video development kits (NVDK) Each NVDK board offers 264 Mbytes of memory accessible via two different DSP external memory interfaces Texas Instruments TMS320C6416 DSPs, 600 MHz

DSP platform

PCI boards plugged into the baseboard, consist of

Eastman Kodaks monochrome sensor LM-9618 captures the images

high-dynamic range of up to 110 decibels at VGA resolution FIFO memory connects the sensor to one of the DSPs

Network processor: Linux (Kernel 2.6.8.1)

access to a broad variety of open source software modules

The SmartCam framework, executes atop Linux

ensures interoperability with the DSPs

Java also runs atop Linux, supporting platform wide applications DSP operating system: DSP/BIOS real-time operating system (DSP framework runs atop it; serves as the SmartCam frameworks counterpart on the network processor)

Mobile Agent System supports autonomous operation of the surveillance tasks Each task incapsulated in a mobile agent which migrate between hosts DSP agents: a module that manages the agents integration into its environment a DSP binary representing the agents functionality an optional set of intermediate data a set of DSP resource Task allocation mechanism requires these parameters to autonomously allocate surveillance tasks to smart cameras SmartCam agents: perform status information and communication tasks are executed on the network processor and can access the DSPs dont include resource requirements or DSP binaries Additional agents provide system functionality task-allocation system System exploits mobile SmartCam agents to determine in a distributed manner how to optimally allocate surveillance tasks to the clusters SmartCams

Two identical SmartCam prototypes Integrated up to three additional PCs (Pentium III running under Linux at 1 GHz) to evaluate larger SmartCam networks Complete SmartCam framework and the MAS could execute on the PC without any modification Diet agents running under Java as the MAS and applied the JamVM Java virtual machine on the smart camera prototype
Compared the SmartCam prototypes Java performance with that of a standard PC

The results showed that the interpreter-based JamVM is about 20 times slower than the Sun Java runtime environment (JRE) 1.4.2 on the PCs
the native computing performance between a Pentium III PC and the SmartCam (XScale) differs only by a factor of two

Multicamera object-tracking application Multicamera system instantiates only a single tracker (agent) task The agent follows the tracked object migrating to the SmartCam that should next observe the object Tracking agent based on a Kanade-Lucas-Tomasi feature tracker Main advantage is its short initialization time

Applicable for multicamera object tracking by mobile agents Tracking agents control the handover process, using predefined migration regions When the tracked object enters a migration region, the tracker initiates handover to the next SmartCam Each migration region assigned to one or more possible next SmartCams Motion vectors help distinguish among several SmartCams assigned to the same migration migration region Motion vectors check whether the object moves in the correct direction A master-slave approach for the tracked object handover

Tracking agents migration between SmartCams takes up to 1 second Task-allocation systems setup timeapproximately 190 milliseconds

Keys to successful deployment of smart cameras are:

the integration of sensing, computing, and communication in a small, power-aware embedded device the availability of high-level image/video processing algorithms or libraries for the embedded target processors (the DSPs) a lightweight software framework supporting glueless intra- and intercamera communication the availability of various system-level services such as task mapping and QoS adaptation to allow autonomous and dynamic operation of the overall multicamera system

System usage:
traffic surveillance detection of stationary vehicles detection of wrong-way drivers computation of traffic statistics such as
average speed lane occupancy vehicle classification

The approach is good considering they are using off the shelf products

The JamVM seems to be slowing down tracker migration (about 1 sec)

Maybe try another virtual machine (eg. Kaffe)

The amount of memory and power dissipation are higher than the design would require it is good for testing and research but not suitable in real world situations

migration times rather long also because of the master-slave architecture - increases resource utilization because two or more trackers are active at the same time

An Integrated Visualization Of A Smart Camera Based Distributed Surveillance System - Sven Fleck, Christian Vollrath, Florian Walter, Wolfgang Straer WSI/GRIS, University Of Tubingen A Mobile Agent-based System For Dynamic Task Allocation In Clusters Of Embedded Smart Cameras - Michael Bramberger, Bernhard Rinner And Helmut Schwabach An Embedded Smart Camera On A Scalable Heterogeneous Multi-dsp System - Michael Bramberger, Bernhard Rinner And Helmut Schwabach Embedded Smart Cameras As Key Components In Reactive Sensor Systems Michael Bramberger, Bernhard Rinner And Helmut Schwabach Decentralized Object Tracking In A Network Of Embedded Smart Cameras M. Quaritsch, M. Kreuzthaler, B. Rinner, B. Strobl Autonomous Multicamera Tracking On Embedded Smart Cameras - Markus Quaritsch, Markus Kreuzthaler, Bernhard Rinner, Horst Bischof, And Bernhard Strobl

1. W. Wolf, B. Ozer, and T. Lv, Smart Cameras as Embedded Systems, Computer, Sept. 2002, pp. 48-53. 2. G.L. Foresti, C. Mahonen, and C.S. Regazzoni, Multimedia

Video-Based Surveillance Systems, Kluwer Academic Publishers,

2000. 3. M. Bramberger, B. Rinner, and H. Schwabach, A Method for Dynamic Allocation of Tasks in Clusters of Embedded Smart Cameras, Proc. Intl Conf. Systems, Man and Cybernetics, IEEE Press, 2005, pp. 2595-2600. 4. R. Steinmetz and K. Nahrstedt, Multimedia Systems, Springer, 2004. 5. A. Maier, B. Rinner, and H. Schwabach, A Hierarchical Approach for Energy-Aware Distributed Embedded Intelligent Video Surveillance, Proc. IEEE/IFIP Intl Workshop Parallel and Distributed Embedded Systems, IEEE Press, 2005, pp. 12-16. 6. J. Shi and C. Tomasi, Good Features to Track, Proc. IEEE Intl Conf. Computer Vision and Pattern Recognition, IEEE Press, 1994, pp. 593-600.