• human action recognition;
  • local search particle filter;
  • reduced integral image;
  • simple features;
  • statistics on temporal evolution;
  • support vector machines


This paper proposes a new approach to recognize human actions in 2D sequences, based on real-time visual tracking and simple feature extraction of human activities in video sequences. The proposed method emphasizes the simplicity of the strategies used, in an attempt to describe human actions as precisely as other more sophisticated (and more computationally demanding) methods in the literature. Specifically, we propose three complementary modules for the following: (a) tracking; (b) feature extraction; and (c) action recognition. The first module is based on the hybridization of a particle filter and a local search procedure and makes use of a reduced integral image to speed up the weight computation. The feature extraction module characterizes the silhouette of the tracked person by dividing it into rectangular boxes. Then, the system computes statistics on the evolution of these rectangular boxes over time. Finally, the action recognition module passes these statistics to a support vector machine to classify the actions. Experimental results show that the proposed method works in real-time, and its performance is competitive against other state-of-the-art methods.