STAN: Spatial-Temporal Attention Based Inertial Navigation Transformer

Inertial Navigation Systems (INS) stand as a foundational technology in navigation, finding diverse applications across aerospace engineering, marine exploration, and terrestrial vehicles. Traditional Newtonian-based inertial navigation methods are susceptible to errors, as they involve double integration of IMU measurements. In recent years, data-driven inertial navigation methodologies like IONet and IDOL have demonstrated their efficacy in harnessing well-trained neural networks to achieve precise navigation based on IMU measurements. However, many existing data-driven inertial navigation techniques focus solely on capturing temporal relationships among raw IMU measurements using recurrent neural networks (RNNs). These approaches neglect spatial information inherent in the raw measurements. This oversight may not only result in suboptimal localization performance but also hampers the ability to discern which measurements are most crucial for specific predictions or corrections. This paper introduces STAN, an innovative Spatial-Temporal Attention-based Inertial Navigation Transformer network, devised to address the deficiency in considering spatial information within existing data-driven inertial navigation techniques. STAN represents a significant advancement by extending the conventional Transformer encoder with a novel spatial-temporal attention mechanism. This mechanism enables the simultaneous extraction of spatial and temporal information from raw IMU readings in parallel, obviating the requirement for RNN or CNN modules. The STAN architecture comprises two transformer encoders: one dedicated to capturing temporal attention, incorporating positional encoding, and the other focused on capturing spatial attention, operating without positional encoding. Subsequent to these parallel encoders, a feature fusion layer is employed to merge the extracted features from both spatial and temporal encoders. Finally, to facilitate prediction, several fully connected (FC) layers are integrated into the network. This design ensures STAN’s capability to effectively incorporate spatial information alongside temporal dynamics, leading to enhanced performance in inertial navigation tasks. We showcase the effectiveness of the STAN approach on the widely used KITTI dataset to assess its performance.