Number of steps by walk mode.

Human gaze behavior is crucial for successful goal-directed locomotion. In this study we explore the potential of gaze information to improve predictions of walk mode transitions in real-world urban environments which has not been investigated in great detail, yet. Using a dataset with IMU motion data and gaze data from the Pupil Labs Invisible eye tracker, twenty participants completed three laps of an urban walking track with three walk modes: level walking, stairs (up, down), and ramps (up, down). In agreement with previous findings, we found that participants directed their gaze more towards the ground during challenging transitions. They adjusted their gaze behavior up to four steps before adjusting their gait behavior. We trained a random forest classifier to predict walk mode transitions using gaze parameters, gait parameters, and both. Results showed that the more complex transitions involving stairs were easier to predict than transitions involving ramps, and combining gaze and gait parameters provided the most reliable results. Gaze parameters had a greater impact on classification accuracy than gait parameters in most scenarios. Although prediction performance, as measured by Matthews’ correlation coefficient (MCC), declined with increasing forecasting horizons (from one to four steps ahead), the model still achieved robust classification performance well above chance level (MCC = 0), with an average MCC of 0.60 when predicting transitions from level walking to stairs (either up or down) four steps in advance. The study suggests that gaze behavior changes in anticipation of walk mode transitions and the expected challenge for balance control, and has the potential to significantly improve the prediction of walk mode transitions in real-world gait behavior.人类注视行为对于成功完成有目标的移动至关重要。本研究旨在探索利用注视信息提升真实城市环境下步行模式转换预测性能的潜力，而该研究方向此前尚未得到充分细致的探究。本研究使用一套包含惯性测量单元（Inertial Measurement Unit, IMU）运动数据与瞳孔实验室（Pupil Labs）隐形眼动追踪仪采集的注视数据的数据集，招募20名参与者完成了涵盖三种步行模式的城市步道三圈行走：平地行走、楼梯（上行、下行）以及坡道（上行、下行）。与既往研究结果一致，我们发现参与者在具有挑战性的转换场景中会更多地注视地面，且会在调整步态行为前最多四个步长时就提前调整其注视行为。我们分别基于注视参数、步态参数以及二者结合训练了随机森林分类器以预测步行模式转换。实验结果显示，相较于坡道相关的转换，包含楼梯的更复杂转换更容易被预测；同时融合注视与步态参数可获得最可靠的预测结果。在大多数场景中，注视参数相较于步态参数对分类准确率的影响更大。尽管以马修斯相关系数（Matthews’ correlation coefficient, MCC）衡量的预测性能会随着预测前瞻步长（从提前1步到提前4步）的增加而下降，但模型仍取得了远高于随机猜测水平（MCC=0）的稳健分类性能；在提前4步预测平地行走转换为楼梯（上行或下行）的场景中，模型的平均MCC达到0.60。本研究表明，注视行为会针对步行模式转换以及预期的平衡控制挑战发生提前调整，该信息有望显著提升真实场景中步行行为的模式转换预测性能。