EDGE-YOLO11: A lightweight dual-stream RGB-T pedestrian detector under varying illumination

RGB-T pedestrian detection has attracted increasing attention because visible and infrared modalities provide complementary information under challenging illumination conditions. However, many existing RGB-T detectors rely on relatively heavy dual-stream or fusion structures, which limits their applicability in resource-constrained deployment scenarios. To address this issue, this paper proposes EDGE-YOLO11, a lightweight dual-stream RGB-T pedestrian detector under varying illumination. The proposed framework processes visible and infrared inputs with a compact dual-stream architecture and selectively introduces GhostConv and DualConv into specific layers rather than replacing all standard convolutions. This design reduces redundant computation while preserving the mid-level RGB-T fusion pipeline. In addition, an efficient multi-scale attention module is employed as a lightweight feature recalibration mechanism to partially compensate for the representation loss introduced by lightweight convolutional compression. Experiments are conducted on the LLVIP dataset with full validation and illumination-specific evaluation, including day, transition, and night subsets. Additional experiments are performed on a constructed FLIR-person subset to evaluate the applicability of EDGE-YOLO11 beyond LLVIP. Compared with YOLOv11-RGBT-s on LLVIP, EDGE-YOLO11-s achieves competitive detection performance with 96.31% mAP50 and 64.68% mAP50-95, while reducing parameters by 77.3% and FLOPs by 71.7%. In CPU ONNX Runtime inference, EDGE-YOLO11-s achieves 44.61 ms latency and 22.42 FPS, which is approximately 2.56 × faster than YOLOv11-RGBT-s. These results show that EDGE-YOLO11 provides a favorable efficiency–accuracy trade-off for deployment-oriented RGB-T pedestrian detection under varying illumination.