Robust human action recognition scheme based on high-level feature fusion

This paper presents our research on the human action recognition which employs different low-level local and spatio-temporal descriptors. The motivation is that these descriptors emphasize different aspects of actions. We investigate a generic approach applied to different periodic and non-periodic actions in the same framework defined by Weizmann and KTH datasets. So, we explore the notion of self-similarity descriptor over time. Then, non-linear chi (2) kernel-based Support Vector Machines are used to perform classification. Individual actions are modeled independently. Finally, classifier outputs are fused using our proposed neural network based on Evidence theory method, trying to improve the classification rate by pushing classifiers into an optimized structure. Experimental results report the efficiency and the significant improvement of the proposed scheme.