Video Action Recognition Using Residual Vector Quantization and Hidden Markov Models

In this paper, we discuss usage of a multi-stage Residual Vector Quantization (RVQ) strategy for human action recognition. To the best of our knowledge, this is the first reported application of multi-stage RVQ to any form of video processing. The RVQ is used to generate a single byte descriptor per image. The evolution of these descriptors is analyzed using a bank of HMM models, one per action to be classified. Correct classification is implemented by maximizing the posterior class conditional density. We report comparable classification results to state of the art methods, i.e. above 90%, while using less than half the available training set.

• Type: Journal
• Year: 2010
• Where: IPCV
• Authors: Salman Aslam, Christopher F. Barnes, Aaron F. Bobick
• Keywords: Class Conditional Density, Image Processing, IPCV 2010, Multi-stage Residual Vector, Single Byte Descriptor