CSM-DETR: construction site monitoring via Mamba-Enhanced detection transformer for UAV aerial imagery

Unmanned Aerial Vehicles (UAVs) offer significant advantages for construction site monitoring through flexible deployment and high-resolution imagery. However, existing vision-based detection methods face three critical technical gaps: (1) CNN-based detectors rely on local receptive fields with limited global context modeling, which is insufficient for disambiguating small construction objects from cluttered backgrounds; (2) transformer-based detectors capture global dependencies but incur quadratic computational complexity O (n²), making them impractical for high-resolution UAV imagery; and (3) conventional multi-scale fusion strategies inadequately bridge the semantic gap between low-level spatial details and high-level semantic features, leading to degraded performance under extreme scale variations. To address these limitations, we propose CSM-DETR, a novel detection transformer specifically designed for UAV-based construction monitoring. Our framework adopts the MobileMamba as backbone to achieve linear computational complexity O (n) while capturing long-range spatial dependencies, and incorporates the Hierarchical Local-Aware Fusion (HLAF) mechanism for adaptive multi-scale feature aggregation. Furthermore, we propose three key innovations: (1) a Dual-Attention Spatial Integration (DASI) module enhancing multi-scale spatial feature representation through parallel local and global attention streams; (2) a Cross-Scale Deformable Fusion (CSDF) module enabling flexible cross-scale feature interaction through deformable sampling; and (3) a Scale-Aware Composite Loss (SAC Loss) providing scale-aware supervision for challenging small objects. We construct a comprehensive benchmark dataset named UAV-CSM47, containing 15, 860 high-resolution aerial images with 47 construction-related object categories. Extensive experiments demonstrate that CSM-DETR achieves state-of-the-art performance with 91. 8% mAP@0. 5 and 73. 6% mAP@0. 5: 0. 95, outperforming YOLOv13-L by 3. 3 percentage points and Co-DETR by 2. 7 percentage points while maintaining competitive inference speed at 38 FPS on an NVIDIA RTX 3090 GPU. Ablation studies validate each component’s effectiveness, and cross-domain evaluation confirms strong generalization capability. The proposed system provides a practical solution for automated construction site monitoring with broad applications in safety supervision, progress tracking, and resource management.