Department of Aerospace Engineering, KAIST, 291 Daehak-ro, Yuseong-gu, Daejeon 34141, Korea.
Sensors (Basel). 2018 Aug 31;18(9):2886. doi: 10.3390/s18092886.
Terrain-aided navigation (TAN) is a technology that estimates the position of the vehicle by comparing the altitude measured by an altimeter and height from the digital elevation model (DEM). The particle filter (PF)-based TAN has been commonly used to obtain stable real-time navigation solutions in cases where the unmanned aerial vehicle (UAV) operates at a high altitude. Even though TAN performs well on rough and unique terrains, its performance degrades in flat and repetitive terrains. In particular, in the case of PF-based TAN, there has been no verified technique for deciding its terrain validity. Therefore, this study designed a Rao-Blackwellized PF (RBPF)-based TAN, used long short-term memory (LSTM) networks to endure flat and repetitive terrains, and trained the noise covariances and measurement model of RBPF. LSTM is a modified recurrent neural network (RNN), which is an artificial neural network that recognizes patterns from time series data. Using this, this study tuned the noise covariances and measurement model of RBPF to minimize the navigation errors in various flight trajectories. This paper designed a TAN algorithm based on combining RBPF and LSTM and confirmed that it can enable a more precise navigation performance than conventional RBPF based TAN through simulations.
地形辅助导航 (TAN) 是一种通过比较高度计测量的高度和数字高程模型 (DEM) 的高度来估计车辆位置的技术。基于粒子滤波器 (PF) 的 TAN 已被广泛用于在无人机 (UAV) 在高空运行的情况下获得稳定的实时导航解决方案。尽管 TAN 在崎岖和独特的地形上表现良好，但在平坦和重复的地形上其性能会下降。特别是在基于 PF 的 TAN 的情况下，还没有经过验证的技术来确定其地形有效性。因此，本研究设计了一种基于 Rao-Blackwellized PF (RBPF) 的 TAN，使用长短期记忆 (LSTM) 网络来忍受平坦和重复的地形，并训练 RBPF 的噪声协方差和测量模型。LSTM 是一种经过修改的递归神经网络 (RNN)，是一种从时间序列数据中识别模式的人工神经网络。通过使用这种方法，本研究调整了 RBPF 的噪声协方差和测量模型，以最小化各种飞行轨迹中的导航误差。本文设计了一种基于 RBPF 和 LSTM 结合的 TAN 算法，并通过仿真证实了它可以比传统的基于 RBPF 的 TAN 实现更精确的导航性能。
