Evolving Visual Tracking using Genetic Programming

The lure of Genetic Programming is the promise of automatically generating “programs”. I demonstrate the use of my C++ based genetic programming framework to evolve programs for visual tracking. In particular I evolve programs using both “Minimal Simulations” and real data to detect the position and velocity of objects in motion. For each experiment several different parameters sets are used, mainly testing the effect of different mutation and recombination rates. It was generally found that higher mutation rates gave better results, but this apparent improvement was not always statistically significant. I also successfully demonstrate the transfer of two programs, evolved by Minimal Simulation, to an application that provides them with live data from a webcam. This application displays the programs output overlaid on the input image and displayed. The programs are seen to perform well at detecting the on-screen position and velocity of single objects moving against a relatively uniform background the task for which they were evolved.