VOOKA

Multi-Channel Video Classification

VOOKA extends the POOKA application domain to real-time video exploitation. The system can exploit spectral, spatial and temporal cues to identify features of interest.

Here's the original POOKA system applied to video data to demonstrate feasibility. On the left is the raw 3 color, Quick Cam data. On the right is the output from the POOKA system. Praise for the stunning video footage should be directed to ISIS team member Neal Harvey! We applied POOKA on a frame-by-frame basis. So for this problem we do not exploit temporal information. The problem was to identify green vehicles (compared to white). The green vehicle in the first frame was used for training data. See if you can find the second, partly obscured vehicle before POOKA does!

More Technical Information

To edxtend into the temporal domain, we have investigated several feed-forward (FIR) and feedback (IIR) architectures. The 5 local memory banks, available on the Firebird Reconfigurable Computer, limit our filters to functions of 3 temporal frames.

Synthetic experiments

Here are some experiments we used to test whether VOOKA would use both spatial and temporal cues.

TRAINING SEQUENCE

Input/Output

POOKA solves the problem by using a low-pass velocity filter, with some directional smoothing?

TEST SEQUENCES

We change the direction of the block, and reduce the size.

InputOutput

Note, there are some discrepancies in how we produced the mpg?s. That is, you may be seeing reflected/rotated differences between input and output?. But you will get the idea..

We change the direction again.

We move the block even faster

We try a stationary block