An Activity Recognition Framework for Overlapping Activities using Transfer Learning

Abstract

Activity recognition is gaining popularity with the increase in digital content. In video data, there is a lot of information hidden that needs to be explored. Human Activity Recognition (HAR) in video streams applies to many areas, such as video surveillance, patient health monitoring, and behavior analytics. Variation in the environment, view-point changes, occlusion, and illumination are some main challenges in HAR. Among other challenges, there is also a similar activity or overlapping activity issue that has not been explored much in past. Resolving overlapping activities classes issues can be a major contribution to overall Human Activity Recognition. Hand-crafted methods and traditional Machine Learning methods were extensively explored in past. Recently, many Deep Learning-based methods are achieved high accuracy. Convolutional Neural Network (CNN) and 3D CNN methods outperform other methods. In this paper, we proposed a Transfer Learning-based Human Activity Recognition (TLHAR) for video data streams. We used VGG16 and InceptionV3, two pre-trained CNN models, and utilized their prior training knowledge for efficient activity recognition. The purposed system outperformed existing activity recognition methods and showed state-of-the-art accuracy and less computational cost requirements than other techniques by taking the benefits of Transfer Learning.