SVMDnet: A Novel Framework for Elderly Activity Recognition based on Transfer Learning

Abstract

Elderly Activity Recognition has become very crucial now-a-days because majority of elderly people are living alone and are vulnerable. Despite the fact that several researchers employ ML (machine learning) and DL (deep learning) techniques to recognize elderly actions, relatively lesser research specifically aimed on transfer learning based elderly activity recognition. Even transfer learning is not sufficient to handle the complexity levels in the HAR related problems because it is a more general approach. A novel transfer leaning based framework SVMDnet is proposed in which pre-trained deep neural network extracts essential action features and to classify actions, Support Vector Machine (SVM) is used as a classifier. The proposed model is evaluated on Stanford-40 Dataset and self-made dataset. The older volunteers over the age of 60 were recruited for the main dataset, which was compiled from their responses in a uniform environment with 10 kinds of activities. Results from SVMDnet on the two datasets shows that our model behaves well with human recognition and human-object interactions as well.