PADSPower Aware Distributed Systems

Algorithms

USC Information Sciences InstituteBrian Schott, Bob Parker, Ron Riley

UCLAMani Srivastava

Rockwell Science CenterCharles Chien

Algorithm DevelopmentOUTLINE

Multi-resolution acoustic beam-forming for tracking and cueing.

Wireless video conferencing over sensor network.

Surveillance video processing.

Laplacian-Pyramid Los-less Image Compaction for distributed ATR.

Algorithm DevelopmentAcoustic Beam Forming

Accomplishments Obtained baseline tracking algorithms and acoustic data, for one and two

targets, from the Army Research Laboratory. Key algorithms modified by Alice Wang at MIT to run on 16-bit StrongARM. Distributed data and algorithms on CD-ROMs to PACC developers.

ARL does not want these algorithms and data distributed to the public.

Progress Developing lower power version of algorithms that do not require FFT.

Plans Measure the power and error in line of bearing resulting from varying:

Number of microphones Length of signal Resolution of signal Algorithm

Investigate using single microphone on a number of nodes for collaborative processing.

Apply algorithms to other data sets.

ARL Remote NettedAcoustic Detection System

Baseline algorithms and acoustic/seismic signal database with 1 and 2 targets provided by the Army Research Laboratory.

Remote surveillance for monitoring / targeting enemy vehicles. Microphone array with ~ 8’ diameter (geometry not restricted). Local processing at the sensor arrays for real-time tracking. Inexpensive, passive and non-line of sight capabilities.

Courtesy of N. Srour, Army Research Lab

Acoustic Sensor Array

ARL Tracking

Each cluster sends its LOB estimate to the end-user where tracking is done

LOB 1

LOB3

LOB2

Beamforming Algorithm

2-sec of each channel is Fourier Transformed with Hamming window.

Frequency domain beamforming algorithm estimates target bearing along 12 beams.

Delay-sum BF delays each signal by a time specified by mic array geometry for coincidence on a given beam.

Code developed in C++ at ARL.

Alice Wang (MIT) converted key subroutines to integer math for implementation on StrongARM processor.

ISI-East is developing an algorithm to shift the signals in the time domain requiring interpolation rather than FFT. This is expected to have little impact on results but require significantly less power.

Algorithm DevelopmentWireless Video

Accomplishments Adapted MBONE open-source multicast video tool,

VIC, to support USB cameras on Laptop PCs over wireless network.

Demonstrated at 29 Palms data collection, with enthusiastic response.

Improved exposure/white balance compensation between video frames for enhanced compression.

Progress Nearly completed adaptation of VIC and Linux

drivers to support 16-bit PCMCIA cameras enabling video collection on iPAQ handheld PC.

Plans Adapt VIC and Linux drivers to support Mercury

sleeve with built-in camera for iPAQ handheld PCs.

Wireless Surveillance and Video Conferencing

Live video display demonstrated at 29 Palms using multicast MBONE video tools over wireless network.

We’re very interested in integrating some video capability into fielded sensor node for future experiments.

Compaq iPAQ H3600

206 MHz StrongARM processor

16 MB persistent flash

32/64 MB RAM

Serial/USB ports

Expansion BUS 1, 2-slot PCMCIA sleeves

VGA out, microdrive, networking, etc.

CF sleeve Custom sleeves

320x240 display with 4096 colors

Rechargeable batteries (Li-ion)

X

iPAQ handheld PC provides user interface to sensor network.

Linux on iPAQ

Compaq supports Linux through http://www.handhelds.org All iPAQ hardware now

supported under Linux ISI-East contributed support for

suspend/resume

iPKG (Itsy Package Management System) Embedded Linux packaging and

distribution system Developed at ISI-East Now an integrated part of

standard handhelds.org Linux Distributions.

BackPAQ Research Sleeve

Digital Camera

32 MB Flash RAM

2 PCMCIA Sleeves

Accelerometer

Programmable Logic, (FPGA)

Digital IO Port

http://crl.research.compaq.com/projects/mercury/

Algorithm Development Video Surveillance

Accomplishments Improved exposure/white balance compensation between video frames for

enhanced change-detection.

Progress Improving continuous update to background image for long term operation.

Plans Create a baseline of surveillance video relevant to the application for testing

algorithms for power-aware change-detection, compression, …. Compare power efficiency of various available lossy compression codecs.

Video SurveillanceBackground Segmentation

Collect image based on schedule or trigger.

Determine and invert changes in background from reference due to lighting and sensor artifacts.

Segment foreground based on threshold of color distance from reference.

Erode & dilate to remove speckle noise and voids.

Transmit foreground, if large enough, for overlay on static reference image.

Collect reference frame as part of sensor set up.

Compress and transmit static frame to potential viewers.

Video SurveillanceBackground Correction

Cameras employ dynamic white-balance and gain to optimize use of dynamic range of 8-bit / color images.

Changes in sun angle, clouds, and adding foreground objects can lead to changes in brightness and color of all pixels in image.

Find relative gain of background between reference and current image for each color channel. Compute histogram of ratio of current to reference pixels, and chose the most

common (mode.) Relative exposure of background should be constant, relative exposure of

foreground to background is more random and often contains fewer pixels.

Only blocks of foreground pixels (often none) are encoded and transmitted for overlay on a reference image.

Lossy compression for viewing, lossless compaction for distributed processing ATR.

Algorithm Development Laplacian Pyramid

Accomplishments Developed Laplacian-Pyramid Lossless Image Compaction for

distributed multi-resolution image processing.

Progress Making current version available to group as baseline.

Plans Create Huffman code tables from Laplacian distribution

estimate based on standard deviation of image differences. Send distribution parameters rather than transmit code table. Can use different code tables for each layer.

Laplacian PyramidHierarchical Image Coding

Compaction (lossless encoding), not lossy compression.

Scenario:

Distributed processing of objects found in video. Compact image to save power on data transmission. Avoid degradation of target recognition by lossy compression. Previous video frame poor estimate of foreground objects.

Immediate results based on low resolution image.

Quality improves with each increment in resolution.

Stop transmitting when sufficient quality attained.

Concept:

Image is successively downsampled by 2 and subtracted from previous layer to form a pyramid.

Averaging of blocks of pixels during downsampling is a gaussian blur, the difference is a Laplacian.

0 1 2 4

3 0 2 4

4 9 8 0

1 2 0 8

1 3

4 4

3

-1 0 -1 12 -1 -1 1

0 5 4 -4-3 -2 -4 4

-2 01 1

3

0 1 2 4

3 0 2 4

4 9 8 0

1 2 0 8

Laplacian Pyramid

Original Image

Image Pyramid

Differential Pulse Code ModulationIntra-frame Compaction

Use pixels on previous lines and columns to predict current

x = A + B + C

where 1 = + + .

Best predictions are typically for the coefficients

x = B + C - A

Encode difference between pixel and its predicted value with variable length encoding such as Huffman.

Predictive differencing maps histogram to Laplacian dist.

P(d) = exp(-2 |d| / ) / [2]

Better prediction leads to a more peaked histogram,

smaller standard deviation, producing more

efficient compaction.

A B

C x

Variable Length Encoding

Based on image histogram, create coding table that translates most common values to fewer bits.

To minimize entropy, maximize coding efficiency, codeword should be chosen with lengths based on prob

L(d) = -log2[P(d)], P(d) = Hist[d] / N

This produces an estimate of the minimum average length

<L> = - { P(d) log2[P(d)] }

Huffman code iteratively combines the two least probable symbols to a single symbol, appending 0 to the code word for one and 1 to the other until there is only one symbol.

Modified Huffman: combine all outliers to single symbol and append 9-bit integer difference or 8-bit original.

0.3 0.6 0.1

1 - 0

1.0 1.0 1.001 1 00

0.4 0.6 0.4

-1 0 1

- - -PCode

Sym

Laplacian PyramidMethod

Use previously transmitted lower resolution image.

Simply encoding is inefficient due to 33% more pixels.

Only 3 in 2x2 downsample block required, fourth can be reconstructed from other 3 and downsampled image. Downsample horizontally.

Save row difference on left. (9-bit) Save row average on right. (8-bit)

Downsample right column vertically Save column difference on top. (9-bit) Save column average in downsampled image. (8-bit)

Similar to Ken Knowlton’s encoding.

1 35 4

3

-2 21 4

-2 -21

Laplacian PyramidResults

Enables progressive reconstruction and achieves compaction nearly as good as intra-frame DPCM.

Use DPCM for initial frame, LP for increments in resolution.

Estimate Orig B B+C2

A+B+C3

B+C-A 3(B+C)-2A 4

LPyr

Tank 7.46 5.37 5.52 5.75 5.25 5.36 5.58

Lena 7.39 4.92 4.47 4.66 4.62 4.39 4.86

HUMV 7.28 5.13 5.11 5.39 4.82 4.81 5.36

A B

C x

Average bits per pixel