Micro-network-based deep convolutional neural network for human activity recognition from realistic and multi-view visual data

摘要

In the recent past, deep convolutional neural network (DCNN) has been used in majority of state-of-the-art methods due to its remarkable performance in number of computer vision applications. However, DCNN are computationally expensive and requires more resources as well as computational time. Also, deeper architectures are prone to overfitting problem, while small-size dataset is used. To address these limitations, we propose a simple and computationally efficient deep convolutional neural network (DCNN) architecture based on the concept multiscale processing for human activity recognition. We increased the width and depth of the network by carefully crafting the design of network, which results in improved utilization of computational resources. First, we designed a small micro-network with varying receptive field size convolutional kernels (1×documentclass[12pt]{minimal} usepackage{amsmath} usepackage{wasysym} usepackage{amsfonts} usepackage{amssymb} usepackage{amsbsy} usepackage{mathrsfs} usepackage{upgreek} setlength{oddsidemargin}{-69pt} begin{document}$$times$$end{document}1, 3×documentclass[12pt]{minimal} usepackage{amsmath} usepackage{wasysym} usepackage{amsfonts} usepackage{amssymb} usepackage{amsbsy} usepackage{mathrsfs} usepackage{upgreek} setlength{oddsidemargin}{-69pt} begin{document}$$times$$end{document}3, and 5×documentclass[12pt]{minimal} usepackage{amsmath} usepackage{wasysym} usepackage{amsfonts} usepackage{amssymb} usepackage{amsbsy} usepackage{mathrsfs} usepackage{upgreek} setlength{oddsidemargin}{-69pt} begin{document}$$times$$end{document}5) for extraction of unique discriminative information of human objects having variations in object size, pose, orientation, and view. Then, the proposed DCNN architecture is designed by stacking repeated building blocks of small micro-networks with same topology. Here, we factorize the larger convolutional operation in stack of smaller convolutional operations to make the network computationally efficient. The softmax classifier is used for activity classification. Advantage of the proposed architecture over standard deep architectures is its computational efficiency and flexibility to use with both small as well as large size datasets. To evaluate the effectiveness of the proposed architecture, several extensive experiments are conducted by using publically available datasets, namely UCF sports, IXMAS, YouTube, TV-HI, HMDB51, and UCF101 datasets. The activity recognition results have shown outperformance of the proposed method over other existing state-of-the-art methods.