LDF-Net: a transformer-based lightweight detail fusion network for UAV engineering vehicle detection
TL;DR
The proposed dataset and detection framework provide an effective solution for engineering vehicle perception in UAV-based intelligent inspection applications and improve multi-range feature modeling and detail-preserving cross-scale fusion.
Abstract
UAV engineering vehicle detection is a fundamental task for intelligent construction monitoring, as engineering vehicles directly reflect construction progress, equipment deployment, and potential safety risks in dynamic construction scenes. However, reliable detection in UAV construction scenes remains difficult due to three major challenges: the lack of a dedicated benchmark dataset, severe scale variation and background interference, and the loss of discriminative details during cross-scale feature fusion. To address these issues, this paper first constructs a dataset named UAV Engineering Vehicles in Challenging Construction Scenes (UEV-CCS), which covers nine representative categories related to engineering vehicles. Then, a Transformer-based detector named LDF-Net is proposed for UAV engineering vehicle detection. Specifically, a Lightweight Grouped Hybrid Attention (LGHA) block is designed to enhance multi-range feature representation in a lightweight manner, while a Detail-Preserving Contextual Fusion (DpcFusion) block is introduced to preserve informative high-resolution details during cross-scale interaction. Together, these two modules improve multi-range feature modeling and detail-preserving cross-scale fusion, enabling LDF-Net to produce more discriminative representations for small and scale-varied engineering vehicles in UAV scenes. Experimental results on the UEV-CCS dataset show that LDF-Net achieves 92.65% mAP@0.5 and 89.23% Recall. Compared with the best competing model, LDF-Net improves mAP@0.5 by 0.62 percentage points while reducing the parameter count by 68.2%. Comprehensive experiments, including ablation studies and external generalization evaluation on VisDrone2019, further verify the effectiveness and robustness of the proposed method. The proposed dataset and detection framework provide an effective solution for engineering vehicle perception in UAV-based intelligent inspection applications.