Activity Identification from Natural Images Using Deep CNN

Abstract

With the increasing demand for surveillance, robotics, health-care, sports play, human-computer interaction and human activity classification & recognition have become a contemporary research topic in the computer vision era. Classification of human activity from video dataset is a really challenging task in the field of computer vision due to high dimensionality, various activities, the variability of inter-class and intraclass, background behavior, actor movement and multiple signs of interaction. The convolutional neural network (CNN) in deep learning plays an active role in today’s technology for classification and recognition patterns from sequence images in the video which is superior to the conventional machine learning methods. In this paper, we propose two convolutional neural network-based deep learning models that are suitable for human activity recognition. Initially, we have applied some optimization techniques in this model that eliminate the major problem of data overfitting. The results of the proposed models show that the proposed optimization methods are increasing the CNN performance higher than the conventional two CNN models. The models trained and tested were KTH and UCF11 YouTube action dataset which contains a large amount of different human activity videos with different conditions. The proposed two models were modified from the two traditional neural network models and the results show that the output of the proposed models achieved the classification accuracy greater than the traditional convolutional neural network models.