Distributed Image Sensing

1
Distributed Image Sensing

• Why
• Information is distributed in an environment
• A lot of that information is visual
• Information costs generally more to move than to
  process
• Energy-efficient operation demands that
  information be processed close to where it is
  gathered
• Low-power image sensors, low-power co-located
  processors make this possible. Contrast with
  traditional approach of network of webcams
  connected to a single, powerful processor
• New approaches
• Local collaboration
• Hierarchical processing
• Etc.

2
Distributed Image Sensing Whats Needed

• Hard data regarding the quantitative energy
  tradeoffs and opportunities offered by
  distributed image sensing
• Design methodologies for energy-efficient
  algorithm implementation on single or multiple
  (at the same or different level in a hierarchy),
  and possibly actuated, sensor platforms
• Re-examination of many traditional image
  processing algorithms, which in the past have
  often targeted high duty cycle, resource-rich
  platforms with very high energy consumption

3
Cyclops

• Low-power wireless image sensor platform
  leveraging cell phone CMOS imagers and works with
  Mica2 and MicaZ wireless sensor nodes
• Large gap between communication and computation
  energy efficiencies
• Energy/bit 3000 nJ/bit (802.15.4 radio)
• Energy/op 4 nJ/op (Atmel AVR ATMega128)
• Sending 1 bit takes same energy as doing 750
  instructions sending a 128x128 image (1s with
  802.15.4 radio) takes same energy as 98M
  instructions, or almost 12.5s worth of computing
• Optimize the system for low-latency communication
  of important event information
• Apply compression for image communication

Mica2 Mote
4
Double-Tier Object Tracking

• Stereoscopic vision via two cameras (LPC1LPC2)
  enables 3D location of object
• When object detected, compute HPC pan-tilt-zoom
  via epipolar geometry
• HPC takes a high resolution image of the object
• Image from HPC used for automated/manual object
  recognition
• Setup
• LPC Cyclops (70mW)
• HPC Canon VC-C4R (12W)

170?
5
Results

• Object localization error with object distance
  for 64x64 resolution

• One-tier vs Two-tier

6
Optimized JPEG Compression

• Image compression is desirable since wireless
  transmission of a bit can require several orders
  of magnitude more than a single computation
• Compression involves costs in memory, image
  quality, computation time and energy, but saves
  transmission time and energy
• These general facts are well known, but little
  experimental work exploring the specific aspects
  of these tradeoffs exist
• DCT and quantization steps have been carefully
  optimized to minimize wordlengths of fixed point
  arithmetic operators ? processing time and energy
  minimization
• Compression ratio, speed and image quality can be
  traded off by using different quantization tables
  and customizing DCT quantization precision

7
Results

• Compression ratios of 4x 6x achieved for 64x64
  and 128x128 grayscale/color images with
  acceptable image quality
• Example128x128 grayscale image
• Assuming 8 MHz processor clock speed and 19.2
  kbps transmission speed
• 802.15.4 radio 3000 nJ/bit
• ATmega128 processor 4 nJ/op

Code Size and Speed Comparisons
Energy Comparisons
RAW JPEG
Code Size 19 kb 25 kb
Computation Time 1 s 3.4 s
Transmission Time 6.8 s 1.1 s
Total Time 7.8 s 4.5 s
RAW JPEG
Computation Energy 32 mJ 109 mJ
Transmission Energy 402 mJ 77 mJ
Total Energy 434 mJ 186 mJ