Human Activity Recognition using Sensor Data based on Recurrent Neural Networks with Bidirectional Long Short-Term Memory
Ashok Kumar Pamidi Venkata, M. Prabhakar, Pradyumna Amasebail Kodgi, Hayder Muhamed Abas, K Priyanka · 2025
Over the past few years, Human Activity Recognition (HAR) systems has been widely used in various fields such as healthcare, security purpose, and activity monitoring due to its real-time capabilities. However, existing Convolutional Neural Network with Bidirectional Long Short-Term Memory (CNN-BiLSTM) model is hindered by its reliance on high-quality training data and substantial computational resources. To overcome these challenges, this research proposes a Recurrent Neural Networks with BiLSTM layers to learn time-series features of sensor data. Initially, data is obtained from UCI-HAR dataset which consist of six movements captured from smartphones of 30 volunteers aged between 19 to 48 years. The collected raw sensor data involves preprocessing includes two steps data filtering with Low-Pass Butterworth Filter (LPBF) to eliminate high-frequency noise and Min-Max scaling used to minimize impact of signal magnitude variations. After that, RNN-BiLSTM is utilized in which, RNN layers function as feature extractor by capturing spatiotemporal relationships that distinguish different activities. Next, extracted features are fed as input into BiLSTM layer to make predictions of performed action like walking, standing, lying down, and seating. Finally, the proposed RNN-BiLSTM achieved better results than existing CNN-BiLSTM in terms of accuracy (98.12%) and F1-score (87.45%) respectively.