vtrack − compute the trajectory of image features in an image set

vtrack if=infile of=outfile pf=pfile [gf=outgfile] [sz=pxysize] [xs=pxsize] [ys=pysize] [h=hsearch] [v=vsearch] [-norm | -sumdif | -sqdif] [-fast] [-lin | -grid | -vel] [-accel] [grid=] [rings=] [extra=] [-img] [-d] [-v] [-N]

Vtrack is a tracking program which determines the trajectories of a number of image features through a set of consecutive frames of an image sequence. Initial locations of the local image regions are specified in the file pfile. A location search is made in each new frame to match the feature image region identified in the previous frame. Image features are square image patches centered on the location pixel coordinates. The default size of a patch is 9x9 pixels. The search space in the next frame is constrained to a local region; the default region size is 30x30 pixels. There are two aspects to finding the location of a feature (image patch) in the next image frame: the patch matching criterion and the order in which locations are searched in the next image (the search algorithm). Vtrack implements a number of different matching metrics and search techniques. The matching metric is specified by the one of [-norm | -sumdif | -sqdif], the default metric is sum of squared differences. The search technique is specified by one of [-lin | -grid | -vel], the default is linear exhaustive search.

Vtrack can generate three output files. The output file specified by the of= parameter is an annotated image set in which the location of each feature is marked with a + sign, the boundary of the feature in the previous frame is marked with a square and the search area is marked with a larger square. The second output format is a vector file (suitable viewing with v3d) specified with the gf= parameter. In this file the temporal trajectories of each feature are stored.

The input images must be of type unsigned byte. When a feature is not located by a match maximum within the search window it is not tracked further.

if=

Specifies the input file which must be a VisionX image set

of=

Specifies the name of the annotated image set output file.

pf=

specifies an (ASCII) text file which contains the number of features to track and their initial x,y locations in the following format:

N

x1 y1

: :

xN yN

xp= yp=

can be used to specify a single point for tracking as an alternative to pf=.

gf=gfile

specifies the vector output file which contains the trajectories of the features.

-N

causes the patch to be loaded only once and NOT updated. This is useful when interframe variations are less than one pixel.

sz=

is the size of a square correlation patch matrix. The default size is 9 (i.e., 9x9).

xs= ys=

is the x and y size of a correlation patch matrix. The default size is 9 (i.e., 9x9).

h=

specifies the horizontal search size; i.e., the number of pixels to search horizontally. The default is 30.

v=

specifies the vertical search size; i.e., the number of pixels to search vertically. The default is 30.

-img

use image coordinates; i.e. with the coordinate origin in the top left rather than bottom left corner.

The correlation matching algorithm is specified by one of the following :

-sqdif

sum of squared differences (equivalent to mse) (default)

-sumdif

sum of absolute differences

-norm

normalized correlation

-fast

may be specified to invoke a faster (optimized) algorithm for the matching methods shown above.

The search algorithm is specified by one of the following :

-lin

linear exhaustive search (default)

-grid

grid pattern search

-vel

velocity prediction

-accel

acceleration prediction

Note that -accel and -vel my be specified simultaneously to provide velocity and acceleration prediction.

grid=

specifies the grid distance (default is 2).

rings=

specifies the number of search rings for velocity and acceleration prediction (default is 0).

extra=

specifies the number of extra moves for velocity and acceleration prediction (default is 0).

-d

specifies that debugging information is to be printed.

-v

specifies verbose mode.

A.P. Reeves, J. Janecek, B.P. Kelley, R. Wu