Closest Point of Approach detection
Each individual sensor consolidates the sensor readings into a single closes
point of approach time stamp for each vehicle that is detected. This
is done through a combination of high-pass, and low-pass filtering of the
raw magnetic readings.
Vehicle Velocity Estimation
The collection of closest point of approach time stamps are used to calculate
the velocity of the object moving through the sensor field.
The time stamps are combined with a position estimation vector which contains
the approximate location of each of the sensor nodes. A linear regression
is the performed on the data to fit a constant velocity trajectory to the
The slope of the regression line is proportional to the velocity of
the car and the intercept of the line is the time that the car passed.
Sensor Location Update
In order to update the location of the individual sensors, a the linear
regression is used to determine the relative positioning of the individual
sensor nodes. Over time, the position estimation vector will be updated
to reflect reality. The initial position estimation vector is set
to place each node 10 feet apart -- the average spacing of motes dropped
from the plane.
Depicted is the adjustment of positions for motes 4 and 5 in
the sensor network. Each node is simply nudged in the direction that
would reduce the difference between the measured data and the linear regression.
Jason and Steve prepare the mote dropper for a launch.The
motes know the order that they will be dropped in, and use that for an
initial guess at their relative locations, which they update as they get
more track information.
It is imperative that the sensor readings from the distributed sensors
can be aligned in time with respect to each other. In order to accomplish
this, the sensor network has a distributed clocking mechanism which keeps
all node in sync with an accuracy of 1/32 of a second. This is accomplished
by having each node periodically broadcast out its version of the current
time. All nodes that hear the broadcast make sure that their clocks are
at least caught up to the time that was broadcast. If any node has
fallen behind, then it will advance its clock. This algorithm allows
the entire network to synchronize itself on the clock of the fastest node.
This mechanism can be used to achieve high levels of synchronization because
there is very little jitter in the transmission time of the update broadcasts.
Sensor Data Acquisition and Filtering
Each axis of the magnetometers are read 32 times per second. The
sensor readings are then placed into a IFR low pass filter in order to
remove noise introduced by the sampling process. This filter takes
the from of FILTER_VAL = (FILTER_VAL * 7 + READING)/8. Then a second
filter process is run to get a lower bound on the frequency of the readings.
Here, SECOND_FILTER_VAL = (SECOND_FILTER_VAL * 7 + FILTER_VAL/8. Finally,
the absolute value of the difference of these two values is filtered to
determine the rate of change in the frequency band being selected.
If this third filter value exceeds at threshold, then we assume declare
that a vechile has been detected. Here is an excerpt of the C code
that is doing the filtering.
ch->first = ch->first - (ch->first >> 3);
ch->first += ch->reading;
ch->second = ch->second - (ch-> second >> 3);
ch->second += ch->first >> 3;
ch->diff = ch->diff - (ch-> diff >> 3);
tmp = ch-> first - ch-> second ;
if(tmp < 0) tmp = -tmp;
ch-> diff += tmp;
Collaborative Signal Processing
In order to determine velocity and direction of the detected vehicles,
the individual sensors must collaborate. A magnetic anomaly alone
does not give us any velocity information. However, by combining the sensor
readings from several nodes, we are able to determine the velocity of the
Position Estimate Update
The vehicle detection information that is calculated by the sensor network
is stored inside the network until an opportunity arises to transfer that
information back to the base station. To achieve this, each node
remembers the velocity vector information that is calculated in a local
EE prom. At a later time it will report that information in response
to a query on the network. In our demonstration, the query originates
from a mote attached to the UAV as it flies over. As the plane passes
over the sensor network, it issues queries to the sensor nodes asking for
a report of the vehicles that have been detected. At this time the
sensor nodes stream the data out of their EE proms and up to the mote on
the plane. The same process is used as the plane passes over the
base station. In this case the base station issues the query while
the plane responds. I short, the plane flies over the sensor network
and picks up the data. It then flies over the base station and deposits
Code Size and Component Graph
Here is a graph of the TinyOS components that went into this application.
And we still had 80 bytes of program memory left!