驾驶中的视觉运动协调：低维表征揭示自然驾驶行为维度随情境自适应降低
Visuomotor coordination on the road: Low-dimensional representations reveal adaptive, context-dependent reductions in the dimensionality of natural driving behavior
基于 284 名驾驶员在沉浸式虚拟驾驶中经历 10 个危险路况事件的眼动、头部与车辆运动数据，研究通过主成分分析构建了跨被试的共同时间分辨行为状态空间。整个驾驶过程中少数成分即可解释大部分行为方差，且在关键事件前后有效维度可靠下降，表明感知与运动变量之间的协调更紧密。低维期间，眼、头、转向与车头朝向的贡献按任务需求向主导维度重组，手动驾驶者之间的被试间相似性更高且部分稳定。这种重组使不同驾驶条件的行为更易区分，仅用前两个成分即可较好分类驾驶模式。结果提示自然驾驶依赖共享的低维视觉运动协调策略，并被情境需求灵活重塑。
来源：PLOS Computational Biology · journals.plos.org
by John Madrid-Carvajal, Shadi Derakhshan, Peter König
The brain’s ability to transform complex, high-dimensional sensory and motor inputs into coordinated, goal-directed behavior remains a central challenge in neuroscience research. Current research suggests that behavior is generated through patterns within a low-dimensional structure. What visuomotor patterns underlie complex, unconstrained behaviors such as driving, and how stereotypical they are, remain poorly understood. Driving involves complex perception-action interactions, which we hypothesize are specific to the driver’s role and adapt to task demands. Here, we unfold these dynamics using an immersive virtual drive that recreates ten hazardous on-road events. We collected eye, head, and vehicle movements from participants assigned to either a manual or an automated driving condition. From the behavioral movements of 284 drivers, we constructed a common time-resolved behavioral state space based on collective variance, computed via principal component analysis across participants. The obtained low-dimensional representation thus captures generalized visuomotor coordination strategies. We then characterized the evolving low-dimensional structure with cosine similarity and various measures of effective dimensionality. We tested whether leading components contained condition-specific information using discriminant analysis. Throughout the entire drive, few components accounted for most of the variance in behavior, and effective dimensionality decreased reliably around critical events, indicating tighter coordination between perceptual and motor variables. During periods of low effective dimensionality, the contributions of eye, head, steering, and vehicle heading reorganized toward leading dimensions in a task-dependent manner. Across events, between-subject similarity was partially stable and stronger among manual drivers. This reorganization also made behaviors more distinct across driving conditions, enabling better classification of driving modes using only the first two components. These results show that naturalistic driving is supported by shared low-dimensional visuomotor coordination strategies that are flexibly reshaped by contextual demands, and provide a principled behavioral framework for linking neural manifolds, human driver models, and the design of adaptive automated vehicles.