An efficient multi-view based activity recognition system for video surveillance using Random Forest

Abstract

Vision-based human activity recognition is an emerging field and have been actively carried out in computer vision and artificial intelligence area. However, human activity recognition in a multi-view environment is a challenging problem to solve, the appearance of a human activity varies dynamically, depending on camera viewpoints. This paper presents a novel and proficient framework for multi-view activity recognition approach based on Maximum Intensity Block Code (MIBC) of successive frame difference. The experimare carried out using West Virginia University (WVU) multi-view activity dataset and the extracted MIBC features are used to train Random Forest for classification. The experimental results exhibit the accuracies and effectiveness of the proposed method for multi-view human activity recognition in order to conquer the viewpoint dependency. The main contribution of this paper is the application of Random Forests classifier to the problem of multiview activity recognition in surveillance videos, based only on human motion.