Activity Recognition Using Particle Network

ABSTRACT Nowadays, human activity recognition is considered as an important field in machine vision. It has a lot of applications in video surveillance, person recognition and prediction. Human activity recognition is a challenging task because of its complexity and highly diversity of actions. Most of the previous methods used tracking algorithms to track the person and then recognize him. In this paper we propose a new method which can be used without tracking the person. In this method, we consider particles on frame, and then by using optical flow algorithm on these particles we obtain some specific features. Using this feature vector can help us to recognize the activity of people. Finally, we will have a feature vector for each frame of activity and by using classification methods such as Adaboost we are classifying the activities. We evaluate this method on Weizmann dataset. The proposed method in addition of having high accuracy, is independent of the need for video tracking algorithms and their constraints Keywords: Human activity recognition, Particle filter, Particle network, Mixture of Gaussian, AdaBoost.