基于改进ERFNet的无人直升机着舰环境语义分割

Semantic Segmentation of Landing Environment for Unmanned Helicopter Based on Improved ERFNet

为增强无人直升机对着舰环境的感知理解,促进其安全高效地实现自主着舰,将ERFNet网络模型应用于无人直升机着舰场景语义分割任务中。首先,结合非对称残差模块和弱瓶颈模块对ERFNet网络模型进行改进,提高运行速度、减少精度损失;其次,利用MultiGenCreator和VegaPrime等技术开发无人机自主着舰仿真系统,并建立无人机自主着舰场景数据集;最后,采用PyTorch深度学习框架实现网络模型,采取模型再训练方法对网络进行学习和训练。实验结果表明,所提网络综合优势明显,平均交并比(Mean Intersection over Union,MIOU)达到76.35%,前向传播时间为22.37 ms。

In order to enhance the perception of landing environment for the unmanned helicopter and promote its safe and efficient realization of the autonomous landing,the ERFNet network model is applied to the semantic segmentation task of unmanned helicopter landing scenes.Firstly,the ERFNet network model is improved by combining the asymmetric convolution module and the non-bottleneck-1D module to improve the operation speed and reduce precision loss.Secondly,the MultiGenCreator and VegaPrime technology are used to develop the autonomous landing simulation system,and the autonomous landing scene dataset is also established.Finally,the PyTorch deep learning framework is used to implement the network model,and the model retraining method is adopted for network learning and training.The experimental results show that the network has better comprehensive network advantages,with the mean intersection over union(MIOU)reaching 76.35%and the forward propagation time 22.37 ms.