Parallelizing the Condensation Algorithm for Visual Tracking Amay Champaneria MIT 6.338 - Parallel Scientific Computing Monday, May 13, 2002 http://money.lcs.mit.edu/~amayc/partracker/
The Problem: Pen-Tracking Track the tip of a pen Given a model of the pen tip, search the image for the best match Where to search? Previous approaches: Maximum correlation Kalman filter
The Condensation Algorithm Developed by Isard and Blake (IJCV 1998) Condensation = Conditional Density Propogation Estimate the future probability density by sampling from the current density and weighting the samples by some measure of their likelihood Visualization…
Visualization of Condensation
Why Parallelize? Condensation involves taking independent samples (a Monte-Carlo method) More samples -> higher accuracy, more robust tracking Allows us to track against cluttered backgrounds
Serial vs. Parallel SerialCondens() { } ParallelCondens() { } obtain_observations() predict_new_bases() calculate_weights() update() } ParallelCondens() { obtain_observations() Bcast(observations) predict_and_observe() Gather(results) calculate_weights() update() Bcast(updated_results) }
Results And higher accuracy than Kalman
Tracking Demonstration