Abstract: Autonomous navigation for unmanned aerial vehicles (UAVs) has achieved promising performances in a known or structured environment. However, the autonomous flight technology is still a challenge in the unstructured environment such as forests. Detecting a forest trail and following it would be the most efficient and safest way for UAV in a forested environment. Considering the difficulty of trail recognition in forests, a model of two-column deep neural networks (2CDNN) is proposed. The proposed method addresses this challenge in a different way by getting feature maps using histogram equalization and edge detection. Then the feature maps and RGB images are fed to a two-parallel deep residual network to extract texture and color features from forest images. Finally, UAV controlling command is generated according to classified results. The proposed method has been evaluated on IDSIA dataset and achieved accuracy of 91.31%, which delivers a 4.41% improvement than previous methods. The experiment results show that the proposed method can improve trail perception performance of UAVs in forests effectively and has some generalization and practical significance in the field of autonomous navigation.

Key words: deep residual network, deep learning, unmanned aerial vehicle (UAV), trail perception

摘要： 无人机自主导航在已知或结构化环境中已取得大量研究成果，但在森林等非结构化环境中的技术仍不够成熟。无人机在复杂的森林环境中通过识别森林步道飞行是一种安全有效的行进方式。提出了一种针对森林环境下路径识别的双列深度神经网络模型（two-column deep neural networks，2CDNN），该网络模型通过直方图均衡化结合边缘提取的方法获取特征图，再将特征图与RGB图馈入两路并列的深度残差网络，从而提取出森林场景中的色彩与纹理特征，最终根据网络分类结果来确定飞行方向指令。该模型在森林数据集IDSIA上进行评估，准确率高达91.31%，比现有的方法提高了4.41%。实验结果表明该模型可以有效地提高无人机在森林环境中的路径感知性能，在自主导航领域具有一定的泛化性和实用意义。

关键词: 深度残差网络, 深度学习, 无人机（UAV）, 路径感知