关键词: 无人机目标检测, 红外与可见光融合, 非配准双光图像, 几何先验, 轻量化算法

Abstract: For infrared and visible images captured by UAV visible-infrared dual-modal payloads in complex low-altitude environments, spatial misalignment often exists between the two modalities. Existing online registration methods usually suffer from high computational cost and significant latency fluctuation, making them difficult to satisfy the constraints of onboard edge computing. To address these issues, this paper proposes a lightweight geometric-prior-constrained method for UAV-based object detection with misaligned visible-infrared images. First, based on the geometric prior of rigidly coupled dual cameras, cross-modal corresponding feature points are annotated offline, and a homography matrix from the visible image coordinate system to the infrared coordinate system is robustly estimated using RANSAC. In this way, the estimation of the dominant global mapping relationship is moved to the offline stage, avoiding frame-by-frame explicit homography estimation and providing a macroscopic coarse-alignment prior for online fusion. Second, a four-channel early-fusion input is constructed based on the lightweight YOLO26 detector. Only the first convolutional layer is extended to adapt to the RGB-IR fusion input, and the newly added infrared channel is initialized randomly to avoid directly transferring visible-light texture priors to the infrared channel. Finally, a lightweight residual alignment module, namely RAM, is designed. With the infrared channel used as a fixed reference, RAM applies a detection-task-driven constrained residual transformation to the visible channels to correct systematic residual misalignment after homography-based coarse alignment. Experimental results on the public DVTOD dataset show that the proposed method achieves an mAP@0.5 of 0.851 and an mAP@0.5:0.95 of 0.539. Meanwhile, the model has 9.975M parameters and a per-image processing time of 2.7 ms, demonstrating that the proposed method achieves a favorable trade-off between accuracy and efficiency with low model complexity.

Key words: UAV object detection, Infrared and visible image fusion, Misaligned visible-infrared images, Geometric prior, Lightweight algorithm