Feature Extraction for Pose Estimation. A Comparison between Synthetic and RealIR Imagery

摘要

This research addressed the problem of pose estimation of three-dimensionalobjects given their two-dimensional IR imagery and corresponding synthetic (computer-generated) IR imagery. Features and techniques were investigated to find those which may be extendable from computer models to real-world IR imagery. GTSIG and SCNGEN were used to create the synthetic imagery. Silhouette and outline shape moments were explored as optimum features for the comparison. Employing back-propagation with momentum as the training paradigm, a two-hidden-layer neural network was able to determine the base-plane orientation of the synthetic imagery to within 7.5 degrees with better than 90% accuracy. (No conclusive results were obtained from comparison with real-world IR imagery.) Additionally, the use of object hot spots relative to object height-to-width ratio is briefly discussed as an alternative feature/technique. (Author)