HEFNet-DETR: hybrid efficient feature network for UAV real time object detection

Abstract In recent years, real-time object detection technologies have shown considerable potential for deployment on resource-constrained platforms such as unmanned aerial vehicles. However, maintaining detection accuracy while satisfying real-time and lightweight constraints remains a substantial challenge. To address the limitations of conventional detectors in multi-scale feature extraction, cross-scale feature fusion, and decoder-level information transmission, we propose a real-time object detection method termed HEFNet-DEtection TRansformer (DETR). First, we introduce a novel parallel intra-scale feature interaction encoder, which replaces the conventional stacked architecture with a dual-branch, single-layer encoding framework to efficiently learn global information within each hierarchical level. Second, we develop a lightweight cross-scale repeated selective connection module that integrates multi-scale features with extremely low computational overhead, substantially enhancing the model’s ability to represent objects of varying scales. Finally, we design a selective hybrid query decoder that employs a dynamic cross-layer query selection mechanism, effectively avoiding computational redundancy and error accumulation caused by traditional layer-by-layer query transmission, thereby improving overall decoder performance. Experiments on the VisDrone2019 and Pascal VOC 2012 datasets demonstrate that, compared with the real-time DETR baseline, HEFNet-DETR reduces giga floating-point operations by 22.58%. On VisDrone2019, it achieves a 27.88% improvement in inference speed and a 4.30% improvement in mAP50; on Pascal VOC 2012, it achieves a 26.51% improvement in inference speed and a 2.46% improvement in mAP50, thereby confirming the efficiency and robustness of the proposed method for real-time object detection.