Automated detection of complex construction scenes using a lightweight transformer-based method

Abstract

Accurate and real-time object detection in complex construction scenes from multiple viewpoints plays a crucial role in effective project management. However, this task remains limited by boundary information sharing and scene sensitivity inherent in deep features. To investigate the deep features in construction scenes and analyze method performance, SODA and VisDrone datasets, mean Average Precision (mAP) series metrics, visual inspection, Grad-CAM, and ablation studies are utilized. This paper proposes a lightweight Transformer-based detection framework named Complex Construction Scenes Transformer (CCS-TR), which integrates with a Scale-Isolate Fusion Attention (SIFA) mechanism and an Instructive Contrastive Learning (ICL) strategy. Evaluation results demonstrate that CCS-TR achieves a 5.1 %–8.8 % improvement in detection accuracy while maintaining lower computational costs, making it suitable for real-time on-site detection. Future work will address detection in interacting complex scenes and develop multi-modal collaboration strategies for extreme lighting.