Human action recognition with sparse geometric features

Abstract

In recent years, human action recognition has become a key issue in all intelligence visual surveillance systems, and it is also an important area of research in computer vision. In this paper, a novel approach is presented for human action recognition. Associated sparse geometric features are extracted based on the second generation Bandelet transformation. Aiming at the latest development of Bandelet transformation, the statistical features were proposed to represent each frame of a video and information from geometric flow to characterise the image context, especially for human body shapes and inner information. Features are selected with AdaBoost algorithm, and test experiments are carried on AdaBoost and Support Vector Machines (SVM). This method is easier to be implemented in the real world, without the complex probabilistic modelling of motion patterns and no need background subtraction. Experimental results demonstrate that the proposed approach can achieve satisfying performance on the Weizmann action dataset and KTH dataset, and exhibits considerable robustness to the body’s size change, because of its inherent strictness in describing human action with geometric flows.

Keywords:

• Sparse Geometry
• Feature selection
• AdaBoost
• Multi-class classification
• Human actions
• SVM

Acknowledgements

This work was supported by the National Natural Science Foundation of China (Grant No. 61075041).