Registration of Images

We use a multi-resolution approach to register the images into a common coordinate frame where we assume that the transformation between the images can be accurately modeled using a projective transformation. Since the projective transformations form a group structure, the concatenation of transformations is well-defined and allows us to efficiently compute the transformations between many pairs of images.

Tracking of Objects in Images

After the images have been registered, we calculate the residual normal flow in the images and based on the local distributions of the magnitudes of the residual flow in the images, we can classify the images into indenpendently moving objects and the background. Since the normal flow also contains directional information, we propagate the distributions in time using the normal flow constraint equations.

Results:

Here are some results on airborne visual surveillance videos:

Sequences I:

Color Coding: Red channel - reference frame, green channel - warped image, blue channel - difference image
Detected objects: purple

Military vehicles at a crossroad
Tank driving along a winding road

Sequences II:

Color Coding: Blue/green channel - background, red channel - detected people , tank

People running into the woods
People running along a street
Tank driving along a winding road

 

Sequence III:

Here we show how the evidence for a pixel belong to an independently moving object is propagated along the normal flow constraint line. The pixels whose motion magnitudes are above a given threshold are colored in red.

Tracking Cars on road