Adaptive Multi-Dimensional Convolution with Sparse Attention Mechanism for Traffic Volume Prediction Using High-Resolution Aerial Images

This work developed a new predictive system using an innovative deep learning approach for traffic management using high-resolution aerial images. The proposed method leverages a novel 3D Transformer-based You Only Look Once v9 (3D-TYolov9) model for the detection and tracking of vehicles from converted 3D image frames (feature 1, f1). The developed 3D-TYolov9 is defined as a system that likely combines the established You Only Look Once v9 (YOLOv9) object detection framework with extensions for 3D data processing, specifically incorporating a Transformer-based module for enhanced feature analysis and tracking within 3D space. Subsequently, 2D features reflecting real-time traffic behavior, such as speed and position (f2), are extracted. To enhance accuracy further, diverse external 1D contextual features (f3), including weather conditions and event disruptions, are integrated. These multi-dimensional features (f1, f2, and f3) are then processed by a Multi-Dimensional Convolution with Sparse Attention-based Dilated Densenet (MDC-SADDNet) for robust traffic forecasting. This architecture effectively captures complex patterns, enabling significant improvements in future traffic management, smart routing, and urban mobility planning. Performance validation confirms the developed model's effectiveness over existing methods.