LTNIN: Lightweight Transformer Network for Inertial Navigation

Abstract

Pedestrian dead reckoning is an important task for smartphones and smart wearable devices. Traditional gait model-based methods make it difficult to match multiple motion patterns of different users. The performance of inertial odometry based on deep learning is impressive.At present, most of the data-driven pedestrian dead reckoning backbone networks use Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTMs), Residual Network(ResNet) and Temporal Convolutional Network(TCN) and related variants. In recent years, Transformer architecture has greatly progressed in large models and multi-scenes. However, the Transformer has many parameters and is not friendly to mobile edge devices. Given the demand for edge devices for lightweight models and the time series characteristics of inertial sensor data, this paper designs a lightweight Transformer inertial odometry network based on LightViT for edge devices. It uses group convolution to extract the characteristics of IMU data while reducing network parameters and fuses IMU data time and space information through a multi-head self-attention mechanism and bidirectional attention multi-layer perceptron. It is verified on three mainstream data sets (RONIN, OXIOD, RIDI). Compared with the ResNet network, which has the best effect in RONIN, ATE is reduced by 8% on average. When the network parameters are reduced by 76.6%, FLOPs are reduced by 77%, and the inference time is 12% faster. With comparable accuracy to ResNet, the network reduces parameters by 87%, FLOPs by 86%, and accelerates inference by 76%. It provides a new reference for future backbone network selection in resource-constrained pedestrian dead reckoning systems.