A Lightweight Multiscale Fusion Framework for Traffic Vehicle Detection From Satellite Remote Sensing, UAV, and CCTV Imagery

Abstract

Vehicle detection from heterogeneous traffic imagery is essential for large-scale traffic monitoring. However, satellite remote sensing, uncrewed aerial vehicle (UAV), and closed-circuit television (CCTV) images differ substantially in spatial resolution, viewing geometry, illumination conditions, and vehicle scale, making unified cross-source detection challenging, especially for small targets and unseen source-domain distributions. To address these issues, this study proposes a lightweight source-conditioned multiscale detection framework for vehicle detection from independent satellite, UAV, and CCTV image domains. The framework is centered on a source-aware multiscale awareness fusion module (SA-MS-AFM), which performs source-expert reweighting from a single input image and supports masked-expert training for incomplete expert availability. SwiftFormer with efficient additive attention (EAA) is adopted for efficient local–global representation, while Shiftwise convolution enhances local structural cues with low overhead. Content-aware reassembly of features (CARAFEs) is introduced in the neck to reduce spatial-detail loss during top-down feature reconstruction, and a four-scale prediction head is designed to accommodate large vehicle-scale variations. Experiments on a pooled cross-source dataset of 6805 images show that the proposed model achieves an $F_{1}$ -score of 0.885, mAP@0.5 of 0.886, and mAP@0.5:0.95 of 0.513 while maintaining 41.53 frames per second (FPS) on the tested RTX 4090 GPU. Ablation, expert-masking, three-scale/four-scale tradeoff, and zero-shot external generalization tests further demonstrate the effectiveness and transferability of the proposed framework under heterogeneous traffic imaging conditions.