SFHA-DET: An Efficient Real-Time Object Detection Framework for Low-Altitude Economic Operations

The low-altitude economy, defined as airspace at or below 1,000 metres, has become a key strategic growth driver with applications spanning urban air mobility (UAM), autonomous drone delivery, and infrastructure inspection. These operations require robust real-time object detection of ground-level targets from aerial viewpoints, yet existing methods face fundamental limitations: severe scale variations across operational altitudes (20-500 m), complex urban background interference, sparse small object detection of distant vehicles and pedestrians, and stringent real-time requirements for on-board analytics. To address these challenges, this paper presents SFHA-DET, a detection framework designed for resource-constrained UAV platforms in low-altitude scenarios. The framework consists of three innovations: (1) the CMSCNet backbone network based on Cross-scale Spatial-Frequency Collaborative Bottleneck (CSFCB) modules, which achieves lightweight multi-scale feature extraction through gating mechanisms and spatial-frequency collaborative processing; (2) the Hierarchical Bidirectional Feature Fusion Network (HBFFN) with Multi-Resolution Adaptive Integration Modules (MRAIM) for complementary cross-altitude feature fusion; and (3) the Dynamic Hierarchical Feature Interaction Network (DHFIN) with ERAM, CFEM, and PSFU sub-modules for sparsified attention computation and multi-scale feature optimization. Extensive evaluation on two representative aerial datasets-VisDrone2019 (urban traffic monitoring) and UAVDT (vehicle detection)-demonstrates that SFHA-DET achieves 29.8% AP at 63 FPS with 44.8 GFLOPs and 14.42M parameters. These computational requirements are designed to fall within the published computational envelope of representative edge accelerators such as the NVIDIA Jetson Orin Nano, substantially reducing hardware demands compared to Transformer-based solutions that require desktop-grade GPU platforms. However, we emphasise that all reported inference measurements are obtained on an RTX 3090 desktop GPU; direct on-device benchmarking on embedded UAV platforms is left as immediate future work, and the present framework is best characterised as edge-oriented in design rather than verified for edge deployment. The proposed framework provides a computationally efficient method for foundational perception tasks in emerging low-altitude applications.