A discriminative representation for human action recognition

Action recognition has been standing as an active research topic over the past years. Many efforts have been made and many methods have been proposed. However, there are still some challenges such as illumination condition, viewpoint, camera motion and cluttered background. In order to tackle these challenges, a discriminative representation is proposed by discovering key information of the input data. This task can be addressed by improvements of two major components: parameterized representation and discriminative classifier. The representation is parameterized with hidden variables and can be learned from training data. And the classifier can be trained to recognize actions based on the proposed representation. The contributions of this paper are as follows: (1) a novel probabilistic representation is utilized to capture the relative significant information of low level features; (2) a novel framework is proposed by combining the parameterized representation and discriminative classifier; (3) an alternating strategy is favorable to improve the performance of action recognition by updating the representation and the classifier alternatively. Experimental results on five well-known datasets demonstrate that the proposed method significantly improves the performance in action recognition.