Adaptive Threshold Filtering to Reduce Noise in Elderly Activity Classification Using Bi-LSTM

Abstract

As the global population ages, there is an increasing need to provide better care and support for older individuals. Deep learning support to accurately predict elderly activities is very important to develop. This research discusses a new model integrating filtering techniques using adaptive thresholds with Bidirectional - Long Short-Term Memory (Bi-LSTM) networks. The problem of activity prediction accuracy, mainly due to noise or irrational measurements in the dataset, is solved with adaptive thresholds. Adaptive characteristics at the threshold are needed because each individual has different activity patterns. Experiments using the HAR70+ dataset describe the activity patterns of 15 elderly subjects and the gesture patterns of 7 activities. Based on body movement patterns, the elderly can be classified as using walking aids. The proposed model design obtains an accuracy of 94.71% with a loss of 0.1984.