Shadow Tracking Demo

1
Shadow Tracking Demo
In this demonstration we show how very simple
devices can collaborate to perform a simple
tracking task.

• To Operate
• Wave your hand over the Motes.
• Observe the screen display on the laptop.
• Interpreting the Display
• The display does not do any computations itself
  it is simply a graphical visualization of the
  radio traffic flowing among the motes.

Black dots represent a motes detection of a
shadow passing over its sensor. When a shadow is
detected, a mote emits a packet containing the
strength, time, and physical location of the
event.
Red dots represent a particular motes prediction
of the motion of your hand. The red lines
associate a specific mote with its prediction.
When a mote observes several events from its
neighbors that appear to be correlated to its own
sensor readings, it tries to fit the data it sees
to a linear model for motion, and reports a
predicted position and velocity of the tracked
object.
The black and red dots shrink over time as the
messages they represent age.
2
About the Hardware
The hardware platform used in this demo is the
Macro Mote platform developed at Berkeley
The objective of the macro mote was to provide
a scaled-up development platform for a device
that could plausibly be shrunk to the size of a
dust particle. Consequently, the mote has very
limited computational capability

• Atmel AVR 8-bit RISC processor, 4MHz
• 512 bytes RAM
• 8KB flash program memory (2 bytes/instruction)
• RFM 918 MHz radio, in OOK mode, 19200 baud
• Transmit power control for variable range
• Analog sampling of baseband signal for RSSI
• CSMA MAC implemented in software on AVR
• Analog sensor interfaces
• 8 ADC channels

These limitations introduce numerous challenges.
Because of the limited memory and communication
resources, it is impossible to keep or transmit
large amounts of historical sensor data, time
series, etc. Thus, algorithms must emphasize
on-board and in-network processing, data
aggregation, and adaptive fidelity, and must
avoid centralized computations.
3
Detection and Tracking Algorithms
The system shown in this demo is composed of
several concurrently operating algorithms

• Time Series and Long Term Average. The light
  sensor is sampled at 16Hz, and a small time
  series is kept in a ring buffer. A long-term
  average is computed by adding each new sample
  into a sum and subtracting out the previous
  average value.
• Event Detection. The minimum value in the buffer
  is computed when the minimum value is followed
  by a rise above a threshold, a packet is
  scheduled containing the time of the minimum, the
  motes location, and the value of the minimum,
  normalized against the long-term average.
• Recent Events from Neighbors. Detected events and
  events reported by neighbors are analyzed if
  they are recent and nearby, they are inserted
  into a buffer of recent events. If we have at
  least one recent detection, we search the buffer
  for the longest linear sequence, deduce a
  velocity vector, and schedule ourselves to emit a
  track event packet.
• Packet Scheduling. In order to avoid broadcast
  floods, packets are scheduled to be emitted after
  a short random delay. During that time, new data
  can override the packet in the buffer if the
  new data is better as judged by the
  application. For example
• A deeper minimum event overrides a shallower
  event
• An event detection takes precedence over a
  track prediction, but a track prediction can
  wait in a separate queue.
• A newer and longer track has precedence over a
  shorter track
• Future Work
• Recent Tracks from Neighbors. Track messages
  reported by neighbors are used to weight
  alternative track hypotheses.
• Suppression of Redundant Tracks. A track
  message from a neighbor will suppress an
  equivalent track scheduled to be sent.

4
(No Transcript)