Real-time trajectory analysis using stacked invariance methods

Abstract

Invariance methods are used widely in pattern recognition as a preprocessing stage before algorithms such as neural networks are applied to the problem. A pattern recognition system has to be able to recognise objects invariant to scale, translation, and rotation. Presumably the human eye implements some of these preprocessing transforms in making sense of incoming stimuli, for example, placing signals onto a log scale. This paper surveys many of the commonly used invariance methods, and assesses their performance on one particular problem, estimating the quality of human motor imitation invariant to rotation, scale and translation.