Deep Learning Pedestrian Navigation Method Based on Multi-task Loss Function*

Abstract

In recent years, indoor inertial navigation technology based on pedestrian dead reckoning (PDR) has been widely promoted. Traditional methods often use auxiliary facilities or environmental constraints to suppress PDR heading cumulative errors, but these auxiliary means restrict the application scope of PDR. PDR based on deep learning fills the need for external information dependence, but the heading estimation accuracy is low and the adaptability is poor. To address this problem, an optimized adaptive multitask loss layer based on uncertain weighting is proposed, which constrains the weight of position and attitude estimation in the overall prediction task and dynamically adjusts it adaptively in different stages to enhance attitude estimation capability. A PDR algorithm based on an end-to-end joint residual neural network and bidirectional long short-term memory network is designed to improve the algorithm’s generalization ability. The original inertial navigation data is processed by segmentation and coordinate normalization and is used as input to the deep learning model to detect features and predict trajectories, achieving accurate indoor pedestrian inertial navigation. Finally, the navigation performance of the proposed algorithm is validated in experiments of walking, running, and mixed gait patterns. The results show that the positioning accuracy of the proposed algorithm is better than that of traditional PDR methods and the RONIN algorithm based on deep learning. The positioning errors in walking, running, and mixed gait patterns are reduced by 21.07%, 10.34%, and 32.15%, respectively, compared to the RONIN algorithm.