Abstract

Artificial Intelligence explores the means for computers to think like a human. In this era of technological advancements, the unobstructed method for monitoring activities in daily life is a challenging task. Most of the recent research successfully achieved good performance at the high computational cost. With high computational costs, it is not possible to deploy action recognition methods in real-time scenarios. This research work aims to develop an action classification approach for video surveillance with scalable extension for deployment on edge devices. The input video is preprocessed extensively to extract the maximum information about the actions with all variants, which enables the network to even recognize the actions from an unseen environment. The Hybrid Video Classification (HVC) is the combination of CNN and GRU that are used for feature selection and preserving the preceding information, and in the end, SoftMax activations are used to classify the action labels. We have used HMDB51 and UT-Interaction datasets to train our proposed model. The model is tested in real-time on low end laptop. The experimentation shows that the proposed work recognizes the actions with better accuracy at a low computation cost.