Obstacle detection, especially real-time power line detection plays a vital role in the low-altitude flight safety of aircrafts. Most of previous power line detection methods fail to deal with curved power lines due to the small size and unapparent visual features in the complex scene. In the paper, we propose a novel fast power line detection network (Fast-PLDN), a real-time semantic segmentation model, for pixel-wise straight and curved power line detection. Besides, we construct our network with low-high pass block and edge attention fusion module, which extract spatial and semantic information effectively to improve the power line detection result along the boundary. Furthermore, we also build up a new dataset named AIR Power Line dataset based on pixel-wise annotations for power line detection task because public Power Line dataset based on pixel-wise annotations is so limited. Our model can run at 189.6 frames per second (fps) with 71.3% mean intersection over union (mIoU) on AIR Power Line dataset, which outperforms most of the previous power line detection methods and the existing real-time semantic segmentation models.

障碍物检测，尤其是实时电力线检测，对飞机的低空飞行安全起着至关重要的作用。由于复杂场景中的小尺寸和不明显的视觉特征，以前的大多数电力线检测方法无法处理弯曲的电力线。在本文中，我们提出了一种新颖的快速电力线检测网络（Fast-PLDN），一种实时语义分割模型，用于逐像素直线和曲线电力线检测。此外，我们使用低高通块和边缘注意力融合模块构建我们的网络，有效提取空间和语义信息以改善沿边界的电力线检测结果。此外，我们还建立了一个名为 AIR Power Line 数据集的新数据集，该数据集基于用于电力线检测任务的逐像素注释，因为基于逐像素注释的公共电力线数据集非常有限。我们的模型可以以每秒 189.6 帧 (fps) 的速度运行，在 AIR 电力线数据集上的平均交集比 (mIoU) 为 71.3%，优于大多数以前的电力线检测方法和现有的实时语义分割模型。