Action Recognition by Imprecise Hidden Markov Models
Authors: Alessandro Antonucci, Rosa, R. and Alessandro Giusti
Abstract: Hidden Markov models (HMMs) are powerful tools to capture the dynamics of a human action by providing a sufficient level of abstraction to recognise what two video sequences, depicting the same kind of action, have in common. If the sequence is short and hence only few data are available, the EM algorithm, which is generally employed to learn HMMs, might return unreliable estimates. As a possible solution to this problem, a robust version of the EM algorithm, which provides an interval-valued quantification of the HMM probabilities is provided. This takes place in an imprecise-probabilistic framework, where action recognition can be based on the (bounds of the) likelihood assigned by an imprecise HMM to the considered video sequence. Experiments show that this approach is quite effective in discriminating the hard-to-recognise sequences from the easy ones. In practice, either the recognition algorithm returns a set of action labels, which typically includes the right one, either a single answer, which is very likely to be correct, is provided.
Details: In Proceedings of the 2011 International Conference on Image Processing, Computer Vision and Pattern Recognition, IPCV 2011. CSREA Press, pp. 474-478.
A version similar to the published paper can be downloaded.