Abstract: To address the problems of rotated objects, multiple scales, and large aspect ratios encountered in the coarse positioning of weld seams during welding automation, a coarse positioning method for weld seams based on an improved YOLOv11 oriented bounding box (YOLOv11-OBB) model was proposed. First, an OBB-RLSCD detection head was designed to capture rotational features using a combination of depthwise convolution and pointwise convolution, accurately capturing rotational features under a lightweight architecture and improving the detection efficiency and accuracy. Second, an adaptive lightweight downsampling module was adopted to improve the convolution, effectively retaining the key features of weld seams while reducing the amount of computation. Additionally, dynamic convolution was integrated into the C3k2 module to enable the convolution operation to adaptively adjust according to the input features. Finally, the PIoU loss function was introduced to optimize the positioning accuracy of rotated objects at the pixel level, enhancing the adaptability to complex scenes. The results indicate that the mean average precision of the model reaches 86.5%, which is 3.9% higher than that of the original YOLOv11-OBB model; the number of parameters is reduced by 10.4%; the computation amount is reduced by 18.2%, and the inference speed reaches 89.7 frames/s, which meets the demand for real-time coarse positioning of welding robots. Meanwhile, the model is validated on the public dataset NEU-DET, and it still maintains high detection accuracy, which proves that the algorithm has good adaptability in different industrial scenarios.