Christiane Wiebel, Petros Georgiadis, Martina Hasenjäger, "On the role of eye and head movements for walk transitions in real world urban scenes ", 46th European Conference on Visual Perception (ECVP), 2024.

Human gaze behavior plays a significant role for successful goal-directed locomotion. Yet, it has been rarely employed in multimodal models for predicting real-world human walk behavior. In this study, we set out to investigate its potential for improving the prediction of upcoming walk mode transitions in real-world urban scenes. We use a publicly available data set including IMU motion data and gaze data from the Pupil Labs Invisible eye tracker. 20 participants completed 3 laps of an urban outdoor walking track including 3 types of walk modes: level walking, stairs (up, down) and ramps (up, down). In line with previous work, we found that participants direct their gaze (vertical eye and head angle) more strongly towards the ground during more challenging transitions including stairs. Moreover, participants adjust their gaze behavior prior to adjusting their gait behavior. Based on these results, we trained a random forest classifier for predicting walk mode transitions based on either gaze or gait parameters or on the combination of the two. Results show that more challenging transitions are easier to predict and that combining gaze and gait parameters leads to the most reliable results. However, gaze parameters had a greater impact on classification accuracies compared to gait parameters in almost all scenarios. While accuracies for correctly predicting a transition drops as a function of the forecasting horizon (1 to 4 steps ahead), we still find an average classification accuracy of 60% for predicting a transition from walking to either stairs up or down, 4 steps ahead in time. Taken together, our results suggest that gaze behavior changes in anticipation of an upcoming walk transition and as a function of the expected challenge for balance control. Consequently, we show that it can significantly improve the prediction of walk mode transitions for real-world gait behavior.