Investigation of Drivers' Visual Attributes in Highway Entrance Zones Utilizing Self-Organizing Mapping Neural Network

Abstract With increasing highway traffic, safety concerns at ramp entrances and confluence areas have intensified. This study analyzes drivers' eye movements in real-world settings, focusing on visual behavior in entrance curve section, merging observation section, ramp merging section, and freeway section. Metrics such as saccade time, angle, and speed were assessed alongside the complexity and task load of each section. Results show that the merging observation section exhibits the highest visual task load, with peak saccade time, angle, and speed at 100.78 ms, 35.09°, and 1.93 deg/ms, respectively, exceeding those in other sections. Through Self-Organizing Map (SOM) neural network clustering analysis, saccade time and angle were categorized into five groups: short-time small-angle, short-time wide-angle, medium-time small-angle, medium-time wide-angle, and long-time small-angle. Driver visual behavior varies distinctly across different road sections. Notably, in the merging observation section, the average eye movement-speed matching index (EMSMI) peaks at 0.1249 deg/ms, with a marked increase in categories A2 and A5. This section involves complex driving tasks, requiring frequent visual adjustments to navigate dynamic traffic conditions. In contrast, the entrance curve and freeway section predominantly exhibit categories A1 and A3, with lower EMSMI values of 0.0410 deg/ms and 0.0408 deg/ms, respectively. These sections show fewer outliers and a more concentrated distribution, indicating reduced cognitive load. This study quantitatively elucidates the link between drivers' visual behavior and road complexity, highlighting the heightened cognitive demands of complex traffic environments and their effects on visual behavior.