Human Activity Recognition based on Stacked Autoencoder with Complex Background Conditions

Abstract

Human activity recognition is one of the prime focus areas of computer vision having a range of current and evolving applications in the real-world environment such as abnormal activity recognition, pedestrian traffic with action detection, video indexing, gesture recognition, etc. The goal of this paper is to propose a human action recognition framework that can efficiently work in complex background by exploiting the stacked autoencoder principle. Due to the rapid development of artificial intelligence (AI) aided approaches of decision making, deep learning (DL) is a preferred area of research. Among several known DL approaches, the stacked autoencoder has received extensive research interest and is considered to be among the current state-of-the-art approaches. In particular as part of this work, a stacked autoencoder with three hidden layers is trained in the first stage for representation learning. In the second stage, a SoftMax layer is integrated as a final output layer for the classification of various human actions. We applied the proposed method to a publicly available human action database to evaluate its performance. The feasibility and the effectiveness of the proposed stacked autoencoder-based human action recognition framework have been demonstrated by experimental simulation in this paper.