Scholarlink

<p>Global Navigation Satellite System (GNSS) and Inertial Navigation System (INS) integrated navigation suffers from severe accuracy degradation in urban canyon environments caused by signal blockage and multipath interference, thereby limiting its reliability in autonomous driving applications. To address these challenges, this study proposes a novel cascaded Transformer-LSTM framework combined with Factor Graph Optimization (FGO) for vehicle GNSS/INS integration. This framework employs a Transformer encoder with self-attention to extract global temporal dependencies from IMU measurements while suppressing high-frequency noise through parallel processing. Subsequently, a Long Short-Term Memory (LSTM) network models’ residual sequences to deeply explore nonlinear error patterns and generate adaptive covariance matrices for FGO, replacing fixed prior settings with data-driven uncertainty quantification. Experimental validation on the Wuhan University public dataset demonstrates that the proposed method reduces the root mean square error (RMSE) of position from a consumer-grade IMU (ICM20602) by 76.7% to 2.58&#xa0;cm, improves velocity RMSE by 69.1%, and achieves a confidence level of 99.8%. Notably, the framework maintains centimeter-level positioning accuracy during GNSS signal interruptions, achieving these results across 199 interruptions averaging 1.71&#xa0;s. However, limitations remain, including reliance on single-scenario validation and lack of multi-source sensor integration (wheel odometer, barometer). Future work will focus on deploying lightweight Transformers on edge computing platforms and integrating them with visual inertial odometers to enhance robustness in prolonged GNSS-denied environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Cascaded transformer-LSTM architecture for urban canyon navigation with factor graph optimization

  • 期刊论文

  • 发表日期：

    2026年1月8日

查看全文
  • Huan Zhang,
  • Wenyi Luo,
  • Mengzhu Peng,
  • Jian Kong,
  • Lei Liu,
  • Yibin Yao,
  • Chaoqian Xu,
  • Feifei Tang,
  • Jianping Pan

摘要

Global Navigation Satellite System (GNSS) and Inertial Navigation System (INS) integrated navigation suffers from severe accuracy degradation in urban canyon environments caused by signal blockage and multipath interference, thereby limiting its reliability in autonomous driving applications. To address these challenges, this study proposes a novel cascaded Transformer-LSTM framework combined with Factor Graph Optimization (FGO) for vehicle GNSS/INS integration. This framework employs a Transformer encoder with self-attention to extract global temporal dependencies from IMU measurements while suppressing high-frequency noise through parallel processing. Subsequently, a Long Short-Term Memory (LSTM) network models’ residual sequences to deeply explore nonlinear error patterns and generate adaptive covariance matrices for FGO, replacing fixed prior settings with data-driven uncertainty quantification. Experimental validation on the Wuhan University public dataset demonstrates that the proposed method reduces the root mean square error (RMSE) of position from a consumer-grade IMU (ICM20602) by 76.7% to 2.58 cm, improves velocity RMSE by 69.1%, and achieves a confidence level of 99.8%. Notably, the framework maintains centimeter-level positioning accuracy during GNSS signal interruptions, achieving these results across 199 interruptions averaging 1.71 s. However, limitations remain, including reliance on single-scenario validation and lack of multi-source sensor integration (wheel odometer, barometer). Future work will focus on deploying lightweight Transformers on edge computing platforms and integrating them with visual inertial odometers to enhance robustness in prolonged GNSS-denied environments.