Tracking Object Positions in Real-time Video using Genetic Programming


Authors: Will Smart, Mengjie Zhang
Source: GZipped PostScript (745kb); Adobe PDF (216kb)

This paper describes a new approach to the use of Genetic Programming (GP) to evolve programs for tracking objects quickly in streaming video. A small number of images, with located objects, are used as training data and GP automatically performs feature-selection on these images at the pixel level. The use of feature functions is introduced, taking a single offset argument, in contrast to the standard feature terminal approach. The features include both ``directionless'' intensity features and ``directional'' edge detection features. The fitness function rewards evolved programs that can move training points, located on a grid around an object, closer to the object. As such, a good program will also be able to update an object position from frame to frame for tracking. Two video sequences are examined, with evolved programs tracking the left-eye and forehead of a person successfully. The method is very fast, tracking a frame in six or seven milliseconds on a 2.6GHz PC.

[Up to Computer Science Technical Report Archive: Home Page]