Metro-to-park last-mile walking: A video- and machine-learning-based analysis of objective environmental features,...

Perceptions of metro-to-park first/last-mile (FLM) walking environments may shape people's travel intention. However, existing FLM studies have rarely considered urban parks as destinations or examined how dynamic, multisensory environmental perceptions shape travel intention. This study develops a video- and machine-learning-based framework to examine how objective environmental features and subjective perceptions are associated with walking intention in metro-to-park last-mile contexts. Eleven 60-s real-world walking videos were used in an online perception experiment involving 370 participants. Visual streetscape features and traffic-noise exposure were extracted using DeepLabV3+ semantic segmentation and a modified ECAPA-TDNN sound-classification model. XGBoost was used for prediction, and SHAP was applied to interpret feature importance, nonlinear patterns, and interaction effects. The results show that subjective environmental perceptions contributed more strongly to walking intention predictions than objective visual features. Pedestrian infrastructure quality was the dominant contributor, followed by landscape attractiveness, while traffic noise showed a negative and threshold-like contribution. Gender-stratified results further indicated that female participants were more sensitive to landscape attractiveness. Stratified analyses by gender and age revealed slight heterogeneity. Females were more sensitive to landscape attractiveness, whereas older adults placed greater emphasis on traffic order and perceived safety. Interaction analyses further revealed complex relationships among subjective perceptual variables. These findings highlight the value of multisensory perception-based assessments and provide practical implications for managers to improve metro-park walking corridors through perception-oriented environmental management.