Comparison of RNN-LSTM and Kalman Filter Based Time Series Human Motion Prediction

Abstract Machine-human and machine-machine interaction is inevitable and needs to be considered apart from human-machine interaction. For the safety of machine interaction, the system is required to prevent any unexpected accident. This paper proposed a system that can forecast human motion to know the movement direction for one second. This research was conducted by using an RGB camera as a reliable alternative. The feature extraction process has been done by OpenPose to obtain the coordinate of human body parts in the frame. Then the coordinate data is converted to the movement data containing the distance and direction in the key points to the next frame for the input in the prediction method. This research aims for the human motion prediction using these methods to compare the performance on the human motion data and realize the prediction of human motion. Mainly, Kalman Filter shows more positive results than RNN-LSTM as the prediction method. Movement like hand gestures is more effortless than motions like hand gestures and steps to the left side. The validity of the system based on the RGB camera with unstable data has been confirmed based on the results.