Convolutional Autoencoder for Vision-Based Human Activity Recognition

Human activity recognition (HAR) is a crucial component for many current applications, including those in the healthcare, security, and entertainment sectors. At the current state of the art, deep learning outperforms machine learning with its ability to automatically extract features. Autoencoders (AE) and convolutional neural networks (CNN) are the types of neural networks that are known for their good performance in dimensionality reduction and image classification, respectively. As most of the methods introduced for classification purposes are limited to sensor based methods. This paper mainly focuses on vision based HAR where we present a combination of AE and CNN for the classification of labeled data, in which convolutional AE (conv-AE) is utilized for two functions: dimensionality reduction and feature extraction and CNN is employed for classifying the activities. For the proposed model’s implementation, public benchmark datasets KTH and Weizmann are considered, on which we have attained a recognition rate of 96.3%, 94.89% for both, respectively. Comparative analysis is provided for the proposed model for the above-mentioned datasets.