对机动变换的装甲目标进行快速精确检测是低空无人机的一项重要性能要求, 但目前主流检测方法自身的旋转不变性不能有效应对这一挑战。结合深度卷积神经网络（CNN）提出基于旋转不变Faster R-CNN的低空装甲目标检测方法, 该方法在Faster R-CNN框架的基础上引入旋转不变层, 通过在模型的目标函数上增加正则化约束条件来加强目标CNN特征旋转前后的不变性。实验选取三种典型的装甲目标缩比模型, 在室内外模拟不同场景条件下的低空侦察环境, 利用偏振高光谱相机获取目标的侦察模拟图像作为样本数据用于模型验证。在多模型对比实验中, 改进模型的平均检测准确率提升了2.4%, 取得了最好的检测效果, 初步验证了改进方法的有效性。

Fast and accurate detection of maneuvering armored targets is an important performance requirement for low altitude unmanned aerial vehicles, but the rotation invariance of the current mainstream detection methods is not enough to deal with the challenge effectively. Combined with deep convolution neural network (CNN), we propose a low altitude armored target detection method based on rotation invariant Faster R-CNN. This method introduces the rotation invariant layer on the basis of the original frame of Faster R-CNN to strengthen the invariance of the target′s CNN feature before and after rotation by adding regularization constraints on the objective function of the model. In the experiment, three typical models of armored target are selected to simulate the low altitude reconnaissance environment under different scenes indoors and outdoors, reconnaissance simulated images of the targets are used as sample data for model verification, which are obtained by using a polarizing hyperspectral camera. In the multi model comparison test, the improved model increases the mean average precision by 2.4% on the original basis and achieves the best test result, which preliminary verifies the effectiveness of the improved method.