Adaptive-grid-based Risk Modeling for Left-turn Vehicle-pedestrian Interactions from Aerial Photography Data

To address the complex pedestrian-vehicle interactions during left turns at intersections, an adaptive assessment framework based on a multi-dimensional fusion risk field is proposed. Limited observability, random motion states, and incomplete risk characterization are considered in the framework. The framework combines the YOLOv8n and DeepSORT to detect and track multiple road users from UAV aerial data. It extracts pedestrian crossing behaviors and vehicle motion features at intersections with high precision. Vehicle kinematic models and pedestrian stochastic motion models are established to describe pedestrian-vehicle motion states. A risk assessment model for left-turn scenarios is then constructed. The model integrates a blind spot field, an interaction potential field, and a field of possible future motion. An adaptive grid modeling method is further introduced to adjust spatial resolution according to risk complexity. This method improves computational efficiency while maintaining modeling accuracy in dynamic scenes. Simulation is conducted at the intersection of Jianshe Road and the East First Section of Yihuan Road in Chengdu. The model achieves a tracking accuracy of 97.30% and a localization accuracy of 0.71. Compared with existing tracking methods, it shows better stability and localization performance in dynamic pedestrian-vehicle interaction scenarios. Compared with fixed-grid methods, the adaptive grid method reduces computation time by 20.95% to 37.62%. The prediction error remains within ±2%. Compared with traditional risk assessment methods, the proposed multi-dimensional fusion risk field model performs best in conflict prediction accuracy and scenario adaptability. The proposed model enables high-precision, efficient, wide-area dynamic tracking and adaptive identification of pedestrian-vehicle interaction risks based on UAV aerial data.