Real-time Abnormal Human Activity Detection Using 1DCNN-LSTM for 3D Skeleton Data

Over the last few years, together with the advent of deep learning and the accessibility of depth sensors, home care supervision has risen in demand. This is becoming even more evident with the ongoing COVID-19 epidemic, which has put vulnerable people like the elderly and handicapped at even a more significant health risk and created a need for online and distance-based alternatives for human interaction. As such, a lightweight method for real-time Abnormal Human Activity Detection is proposed. It uses a rolling window approach when processing streams of activities and pinpointing exactly when an abnormality occurs. A combination of an One-Dimensional Convolutional neural network (CNN) and a Long Short-Term Memory (LSTM) network is used, both overlapping in their application but having different advantages and disadvantages. The method achieves a precision of 96.6% and an accuracy of 91% on the TST Fall Dataset v2.