AIRFLOW: An End-to-End Spatial–Channel Cross Attention Framework for Real-Time Passenger Detection and Tracking

Real-time passenger detection and tracking in transportation environments presents significant challenges due to complex backgrounds, varying lighting conditions, dense crowds, and severe occlusions. This paper proposes AIRFLOW, an end-to-end framework built upon RT-DETR that incorporates two novel components: the Spatial-Channel Cross Attention (SPCA) mechanism and the Cross-scale Guided Feature Module (CGFM). SPCA enhances feature representation by modeling bidirectional spatial and channel dependencies simultaneously through explicit cross-attention computation with 8 attention heads, enabling robust detection of occluded passengers. CGFM effectively fuses multi-scale features through semantic guidance to preserve critical spatial and contextual information. Experiments on three public benchmarks and one self-constructed AirportPassenger Dataset demonstrate that AIRFLOWachieves 89.7%, 86.9%, 88.2%, and 82.3% mAP@0.5 respectively, while maintaining 52 FPS inference speed. The framework provides an effective solution for intelligent transportation systems.