Abstract
Visible–thermal object detection benefits from the complementary properties of RGB and thermal imagery, but repeated cross-modal fusion can increase model complexity, particularly in lightweight detectors. This paper proposes SDR-YOLO, a scale-selective detector designed to make better use of shallow spatial details without adding an extra prediction scale. A P2-guided Cross-modal Detail Enhancement module uses visible and thermal P2 features as auxiliary detail sources and injects the resulting residual into the P3 fused feature. P4 and P5 retain simple concatenation-based fusion, while a Lightweight Shared Convolutional Detection Head reduces redundant prediction parameters across scales. Experiments are conducted on DroneVehicle and M3FD. On DroneVehicle, SDR-YOLO achieves 80.4% mAP@50 and 56.7% mAP@50:95, improving both metrics by 1.0 percentage point over YOLO11s, with 14.06 M parameters and 32.51 GFLOPs; under workstation profiling conditions, it reaches 197.62 ± 1.57 FPS. On M3FD, the model obtains 82.5% mAP@50 and 55.6% mAP@50:95, corresponding to changes of 0.2 and −0.1 percentage points relative to YOLO11s. Since these differences are comparable to the run-to-run variation observed across three training seeds, they are treated as numerical trends rather than statistically established improvements. On the Jetson Orin Nano, TensorRT FP16 deployment achieves 45.78 ± 0.16 FPS, with a latency of 21.84 ± 0.08 ms and peak additional unified-memory usage of 590.4 ± 5.1 MiB, although YOLO11s-RGBT runs faster on this platform. Overall, the proposed design provides a clearer benefit in dense UAV-view scenes while maintaining comparable accuracy and slightly lower model complexity on M3FD, although its runtime efficiency remains hardware-dependent.