Activity monitoring in assisted living environments

Abstract

Deep learning has recently emerged as a promising alternative in Human Activity Recognition (HAR) scenarios in ubiquitous and wearable sensors. The HAR enables us to employ a number of applications such as forecasting activity, health care systems by monitoring their behaviour, assisting the elderly or dementia patients, and so on. Therefore, there are numerous machine learning algorithms which are used in the Human Activity Recognition(HAR) such as, support vector machine, random forests, hidden markov model and so on. The problem with the traditional machine learning model is that we need to extract features manually. It requires expert domain knowledge and a very tedious task—The primary concern with adopting deep learning is that we do not need to extract features explicitly. Problems with real-world datasets, most notably imbalanced datasets and poor data quality, continue to limit the accuracy of activity recognition. The processed data is separated into overlapping, fixed-size panes. This window gives input to the 1D CNN and extracts the features from each window, which are then provided to the LSTM network. This thesis proposed a one-dimensional Convolution neural network (or CNN) long-short term memory or LSTM (or 1dCNNLSTM) model that recognises the activity from sensor data input. The 1dCNN is used as a feature learning from processed input data sequence. And LSTM maintains the information regarding or maintains the temporal features between sequences. Following that, the output features are fed into a series of fully connected layers for final activity recognition. The model's performance has been tested on the Opportunity dataset, a very complex dataset recorded in a rich sensors environment. The proposed model performance is tremendous in Human Activity Recognition and verified in the experiment chapter. Keywords: Assisted living, CNN, LSTM, Confusion matrix, ROC curves, Activity Recognition