Machine Learning Prediction of Locomotion Intention from Walking and Gaze Data

Abstract

In many applications of human–computer interaction, a prediction of the human’s next intended action is highly valuable. To control direction and orientation of the body when walking towards a goal, a walking person relies on visual input obtained by eye and head movements. The analysis of these parameters might allow us to infer the intended goal of the walker. However, such a prediction of human locomotion intentions is a challenging task, since interactions between these parameters are nonlinear and highly dynamic. We employed machine learning models to investigate if walk and gaze data can be used for locomotor prediction. We collected training data for the models in a virtual reality experiment in which 18 participants walked freely through a virtual environment while performing various tasks (walking in a curve, avoiding obstacles and searching for a target). The recorded position, orientation- and eye-tracking data was used to train an LSTM model to predict the future position of the walker on two different time scales, short-term predictions of 50[Formula: see text]ms and long-term predictions of 2.5[Formula: see text]s. The trained LSTM model predicted free walking paths with a mean error of 5.14[Formula: see text]mm for the short-term prediction and 65.73[Formula: see text]cm for the long-term prediction. We then investigated how much the different features (direction and orientation of the head and body and direction of gaze) contributed to the prediction quality. For short-term predictions, position was the most important feature while orientation and gaze did not provide a substantial benefit. In long-term predictions, gaze and orientation of the head and body provided significant contributions. Gaze offered the greatest predictive utility in situations in which participants were walking short distances or in which participants changed their walking speed.