A New Method of View-Invariant Human Activity Recognition

Abstract

We present a new method to recognize human activity under uncertain view. The activity width image which is extracted from the sequence of silhouette is used to represent the activity features. It converts the spatial and temporal pattern to a grey-level image. View difference causes the deformation of activity width images and degradation of recognition accuracy. To solve this problem, we transform the feature images to the new images in the uniform view based on the activity templates using Iterative closet point (ICP). For reducing the complexity, we choose the center line and the contour of the activity width images to register and the affine transformation is adopted to transform. The similarity matrix is used to measure the difference between the transformed feature image and the templates. We evaluate our method on the CAS IA database and the results show that our method is simple and efficient.