Martina Hasenjäger and Christiane Wiebel, "On Utilizing Gaze Behavior to Predict Movement Transitions During Natural Human Walking on Different Terrains", PLoS One, vol. 20, no. 10, pp. e0334093, 2025.

Understanding and predicting human walk behavior is an important prerequisite for a proper design of physical assist robot control. One challenge for such systems is the accurate and timely prediction of walk transitions. To improve models based on gait behavior only, prior work has investigated the effect of exploiting visual sensor data. Only few works have included human visual behavior, even though gaze plays a significant role for successful goal-directed locomotive behavior and, therefore, may have the potential to significantly improve predictions of changes in walk modes. In this study, we investigate the potential of using estimates of human gaze behavior for improving walk transition models based on a publicly available real-world data set including IMU motion data from an Xsens motion suit and gaze data from the Pupil Labs Invisible mobile eye tracker. 20 participants completed two outdoor walking tracks including three different types of walk modes: level walking, stairs (up, down) and ramps (up, down). As a first step, we analyzed whether we would find changes in gait and gaze parameters as a function of changes in walk mode. We hypothesized that a change in viewing behavior may be detected earlier than a corresponding change in gait parameters. To analyze the proportion of gaze behavior directed towards the ground, we investigated the vertical deviation in head angle as a first proxy. Results suggest that gait parameters, e.g. average step length, change up to two steps before a transition phase. Average changes in vertical head angle precede those up to three steps. Effects are more pronounced in transitions involving stairs than in those involving ramps. Finally, we show that the time horizon for predicting walk transitions can be considerably extended by including estimates of human gaze behavior in a multimodal model of gait behavior.