一种Anchor Free的无人机检测方法

目前几乎所有最先进的目标检测方法都依赖于预先定义的anchor,但是由于anchor的存在,网络会增加与anchor相关的计算开销,而且在现实世界中,目标尺度多种多样,比如地对空拍摄的无人机目标,预先定义的anchor不可能穷举所有的无人机尺度。因此,该文抛弃预先定义的anchor,采用anchor free的方法来检测无人机。该anchor free方法以目标中心区域的像素为训练样本来逐像素预测边界框的偏移量。针对地对空拍摄的无人机目标大部分都是小尺度的情况,采用高分辨率网络HRNet作为主干网络来提取小目标细粒度的特征,从而提高小目标检测的精度。相对于基于anchor的目标检测方法,该网络框架简单、灵活,并且可以自适应地预测无人机目标的边界框。在自己设计的无人机数据集上,该方法获得了更高的召回率和精确率。

An Anchor Free Method for Detecting Drones

At present,almost all the most advanced object detection methods rely on a pre-defined anchor.However,due to the existence of anchor,the network will increase computation and memory footprint,and object scales are varied in the real world,anchor cannot exhaustive all drone scales.Therefore,we abandon a pre-defined anchor and adopt anchor free method to detect drones,which uses the pixels in the center area of the object as training samples to predict the offset of the bounding box.At the same time,in order to solve the problem that most drones are small,a high-resolution network HRNet is used as the backbone network to extract fine-grained features of small objects,thereby improving the accuracy of small object detection.Compared with the anchor-based object detectors,this network framework is simple,flexible,and can adaptively predict the bounding box of the drone.The proposed anchor free method achieves higher precision and recall on the drone dataset designed by ourselves.