Weight and context method for action recognition using Histogram Intersection

The Bag of Words (BOW) method with spatiotemporal interest points have achieved great success in action recognition. However most of the existing BOW-based approaches treat each local feature evenly and neglect the relations between them. In this letter, we propose a novel method which computes the weights for features and evaluates the relations between them using Histogram Intersection. Additionally, spatio-temporal contextual information is adopted to describe each point. There are two advantages of the proposed method. Firstly, noisy features and discriminative features are treated differently to specified actions; Secondly, spatio-temporal relationship among local features is taken into consideration and is fully used to form the contextual representation of each point. Experiments demonstrate the effectiveness of our method.