Ensemble Transfer Learning for Robust Human Activity Recognition from Images

Abstract

In recent years, the field of Human Activity Recognition (HAR) has witnessed a significant growth owing to the abundance of data and its practical applications in various real-world scenarios. The recognition of human activities from still images remains a challenging task due to the presence of class imbalance and limited intra-class variability. To address these issues, this work proposes an Ensemble Transfer Learning approach for image-based HAR. The proposed model employs an ensemble stacked averaging model consisting of well-known transfer learning architectures such as ResNet50V2, DenseNet169 and VGG19. The ensemble model can learn different features from different architectures, thus providing a robust recognition model. Additionally, data augmentation is employed to increase the diversity of the images in the datasets. The suggested model helps to mitigate the problems of classimbalance and the lack of intra-class variability by generating new images with different variations of the original images. The model is evaluated on two benchmark datasets for image based HAR, namely, the PPMI action dataset and the Stanford 40 Actions dataset. The results demonstrate enhanced performance compared to a few of the related research works.