RSconv: A Shape-Adaptive Learnable Affine Convolution Mode for Robust UAV Tracking in Infrared Scene
Dakai Sun, Qian Jiang, Wangming Lan, Shiwei Wang, Xin Jin · IEEE Open Journal of Instrumentation and Measurement · 2026
Using visual sensors is a common and low-cost strategy for implementing anti-uncrewed aerial vehicle (UAV) tasks. In UAV object tracking tasks, the frequent changes in flight attitude and object shape pose significant challenges for feature extraction. Traditional convolutional forms possess only translation invariance, making them sensitive to rotation and deformation, which easily affects feature extraction, common factors in UAV tracking scenarios. To address this issue, we propose RSconv (rotation-angle and scale-factor learnable convolution), a novel convolutional mechanism that endows convolution kernels with rotation and scale adaptability. In particular, RSconv performs affine convolution using learnable rotation and scale parameters that are optimized during training, enabling the kernel to adapt to local geometric variations of the target. By introducing learnable parameters s and $\theta $ into the convolution kernel, RSconv learns to adjust the kernel parameters through affine transformation (rotation and scaling) before performing convolution. This enables adaptive feature extraction for UAV targets with varying shapes. At the same time, this does not require adding any additional modules. This design allows the convolution operation to achieve enhanced robustness to rotation and scale variations, significantly improving the robustness of tracking models against UAV pose variations. Experimental results demonstrate that on the anti-UAV dataset, using our designed tracker SiamDfc, which already outperforms other trackers, and considering that trackers struggle to extract effective features in the complex anti-UAV410 dataset, our designed SiamDcn outperforms most mainstream convolutional neural network (CNN)- and Transformer-based trackers. These results validate that RSconv achieves translation, rotation, and scale invariance, thereby providing stronger tracking robustness under diverse UAV flight attitudes and target deformations. The code is publicly available at https://github.com/jinxinhuo/RSConv