UDRT-DETR: a small UAV detection method based on infrared imaging and RT-DETR

The rapid development and wide application of unmanned aerial vehicles (UAVs) have made illegal and unauthorized flights a serious threat to public safety, making timely detection essential. However, existing UAV detection methods still struggle to accurately detect small UAVs because of low feature resolution, background clutter, and dense spatial distribution. Transformer-based detectors, such as Detection Transformer (DETR), have shown promising improvements in detection accuracy; however, they demand significant computation and exhibit high inference latency, limiting deployment on resource-constrained platforms. To address these challenges, we propose UDRT-DETR, a small UAV detection method based on infrared imaging and the Real-Time Detection Transformer (RT-DETR). To reduce the computational cost of conventional transformer backbones, we design a cascaded inverted residual backbone (CIRB) that combines cascaded inverted residual mobile blocks (CI-RMBs) with depthwise separable convolutions and structured reparameterization, thereby enhancing feature representation and reducing computation. To reduce the cost of multi-head self-attention, we propose a super token attention-based intra-scale feature interaction (STA-IFI) module that projects tokens into a compact super-token space, eliminating redundant interactions while preserving global context for detecting densely distributed small UAVs. For effective cross-scale integration, we design a slim-neck-ASF, which combines lightweight convolutional units with adaptive upsampling for a precise multi-scale fusion. We design the inner-MPDIoU loss to refine bounding-box regression using auxiliary constraints, thereby improving localization accuracy. Experiments on our self-built infrared UAV dataset demonstrate that UDRT-DETR achieves a precision of 90.71%, which is 4.66% higher than RT-DETR, while reducing GFLOPs by 17.84%, confirming state-of-the-art accuracy and enabling real-time UAVs surveillance.