A Novel Framework for Future Human Activity Prediction Using Sensor-Based Data

Abstract

The prediction of human activities has garnered significant attention, owing to their relevance in diverse applications spanning healthcare, robotics, and user-computer interaction. This paper addresses the imperative need for a comprehensive multi-step prediction system tailored to accommodate these varied domains. The proposed framework comprises three key stages. Firstly, wavelet transform (WT) is employed for data pre-processing to eliminate noise and render the data amenable for time series analysis, the second phase of the framework involves the utilization of a hybrid model combining Long Short-Term Memory (LSTM) and Convolutional 1D (CONV1D) layers, denoted as LSTM-CONV1D designed to effectively address the complexities involved in extracting relevant features and tackling data imbalance issues. The LSTM component is employed for human activity classification based on sensor data, taking into account the sequential nature of activities. Concurrently, the CONV1D component is utilized for feature extraction. In the third phase, a method is introduced for predicting future activity levels and steps. This method incorporates a function that takes the output of the trained LSTM-CONV1D model as an input sequence. The performance of the proposed model is rigorously evaluated across four benchmark datasets based on sensor data: UCI-HAR (Accuracy: 99.03%), M-health (Accuracy: 93.21%), WISDM (Accuracy: 89.43%), and PAMAP2 (Accuracy: 86.85%). Notably, these results represent the state-of-the-art performance characterized by its simplicity attributed to the utilization of a single-layer LSTM and CONV1D. Furthermore, an additional evaluation is conducted on the UCI-HAR dataset with an accuracy of 98.92, wherein the LSTM time series function iteratively generates predictions for future time steps, following the same pre-processing steps.