Transcript Retrieving Large-Scale High Density Video Target Tracks

Chris Cummings

Traffic cameras recording targets and
retrieving them
 Cameras track targets and the data needs to be
recorded, but how are you supposed to store
physical data, let alone go through it?



Mapping and querying something of objects
defined in geometric space.
Normally handle a point, line, or polygon.
3D objects can also be represented (this case)





Sensors are running.
Video-analytic software extract target tracks
from the video in real time.
Geometric rules are drawn on maps by using
the geo-browser* to retrieve targets.
SQL query is submitted to the spatial
database.
Spatial calculations are done and the tracks
of targets are received.

Frames are recorded and saved as image
sequences in JPEG format with roughly 3
frames per second





Spatial database divides space into subspace
then indexes each of the subspaces.
The MS SQL server divides the space to 4
levels of grids.
These levels are divided to 3 grid densities.
Objects are then associated to cells which are
touched as it traverses the grid levels.
These cells become the points for the spatial
database to be indexed.


Querying space is used to search area with
geometric shapes that when combined make
up the entire search space perfectly.
Cells that are touched by objects are
compared to past cells and adjacent cells.

Grids can be broken down to degrees of detail.
 They are calculated by dividing the site size with grid densities



High = small area in meters but finer detail.
Lower = larger area with less detail.
Greater detail not always best.
 Takes up a lot of space
 Takes a lot of computing time to go through the indexes
 6.6 x 6 centimeters accuracy. Not tracked to centimeters so is a
waste.
Configuration of a grid can be set to HHHH, LLLL, and
everything in between.
 Configuration is based on the situation of the enviorment






HHHH is 8.4 times as large as LLLL due to
tracking cells.
To reduce the size of the index, reduce the
size of the data.
Tracklets are the points of data.
For every 1,000,000 tracklets insertion time is
measured as they are put into the database.
To insert 1,000,000 tracklets it takes 1227
seconds. If the data is bigger, this can take
even longer.



Simulations were done using 12 and 18
sensors, putting in data.
Highest insertion capacity if inserted in batch.
Many sensors putting in 100 targets per
insertion gave the best results.



Used like a regular trip wire would be.
The tripwire is defined as a point on all of the
camera sensors.
When the point is crossed an alarm is raised
and all the tracklets are found by finding all
which intersect the tripwire area.




Regions defined by users
These regions can be monitored and
watched.
With this, only the specific area you want will
be tracked which can save on time and space.
Removes unnecessary points.
 BUT what if those points could be used and is
overlooked?

Cache saves time.
 Cache is used to save time so the disk does not have to be
read.

Can get information from disk
 Ran from start

Can get information from cache
 Ran a second time
 Much faster
In running time, SQL sever can read from cache and
disk so it is very good at performance.
 AOI does better then Tripwire because of index sizes.

 Tripwire is larger.





Which do you think takes up more space?
This . (point)
Or -------- (line)
You guessed it! The line!
Because of this, when available points are
used and saved instead of lines.

There are other properties that are used for
targets that can be combined to make a
search fast.
 Speed, Size, Classification, Color

Obviously searching for coupled things would
result in a faster search.
 Speed + Size > Speed

Filter breaks down spatial objects, then the
properties are taken into account. This trims
down on the search time.
Tables need to be broken down for easy access.
If they were not, then it would take much longer
to access all the data.
 Table partitions allow for transactions to be
searched in specific areas instead of the whole
database.
 EX time is 30 days which is broken into 3 10 days
with all the attributes of time. This makes for a
faster search.
 The user sees none of this and it is all still 1 table





The database isn’t 1 central location.
Portioned tables are scattered all over.
A computer may hold a small area and
another may hold more.
This may not make sense to a small region
but if this system is for a city, this is a good
concept.

Questions?