Multiaxial Haar-Like Feature and Compact Cascaded Classifier for Versatile Recognition

A versatile recognition algorithm has been proposed to process image, sound, and 3-D acceleration signals with a common framework at low calculation cost. Firstly, a novel 1-D Haar-like feature is used to roughly extract frequency information from temporal signals. Biaxial and mean-embedded Haar-like features are proposed to extract the standard deviation and the interaxial correlation from 3-D acceleration signals. Secondly, two techniques are proposed to build a compact cascaded classifier. Redundant feature selection (RFS) incorporates the features which are already selected in previous stage classifiers to reduce the calculation cost. A dynamic look-up table (DLUT) is proposed to construct a look-up table-based weak classifier with the smallest possible number of bins. A train loss function is by globally optimized using dynamic programming. The proposed algorithm is tested experimentally on speech/nonspeech classification and human activity recognition. The proposed algorithm yields a speech/nonspeech classification performance comparable to the state-of-art method called MFCC while reducing the calculation cost by 100 times. The algorithm also achieves human activity recognition accuracy of 96.1% with calculation cost reduction of 84% compared with the state-of-art method based on C4.5 decision-tree classifier using the basic statistical features. The proposed algorithm has been employed to build the versatile recognition processor.