Action Recognition Based on Sequential 2D-CNN for Surveillance Systems

Abstract

Action recognition plays an important task in surveillance systems, robot-human interaction and autonomous, systems. However, there are many challenging problems due to varieties of shape, illumination conditions, and complex of actions. Consuming time and precision are typically the main challenges for action recognition systems. Deep neural network techniques have dramatically become the state of the art in image processing. The high capacity deep learning on the temporal action from video analysis has been impeded because of varieties of classes, similarity of actions. This paper presents a new approach based on sequential deep neural network and data augmentation for improving accuracy. First, the deep neural network is constructed using several inceptions with different parallel convolutional operations which support reducing consuming time. Second, image augmentation of training dataset generates a larger data enough for deep neural network learning. This proposed task is aimed to address the small data problem. It is utilized for enhancing capabilities of deep learning. The proposed approach was evaluated on some benchmark datasets. Experimental evaluation on public benchmark dataset shows that the proposed approach enhances performance to 89.53% accuracy. Comparison results illustrated that our proposed method reaches higher performance than almost state- of- the- art methods.