Scholarlink

<p>Accurate stochastic modeling of GNSS Doppler measurements is fundamental for velocity estimation in smartphones, yet conventional stochastic models have poor robustness in complex environments. The existing robust estimation methods require hyperparameter tuning and struggle with multipath/NLOS interference in complex environments. To address these limitations, we propose a novel stochastic modeling method involving Transformer-based Neural Network (TNN) to estimate Doppler-based velocity. First, the traditional Weight Least Squares and Solution Separation (SS) is conducted to get features for TNN training. The TNN then automatically learns the stochastic model and assigns measurement weights based on multi-source features including estimation residuals and SS-Statistics, enabling robust Doppler-based velocity estimation. The experiment results using smartphone GNSS in suburban environments show that the accuracy of the velocity solutions using the new stochastic model is improved compared to the use of traditional methods such as the Huber equivalent weight function and the elevation-based stochastic model.</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

A transformer-based stochastic model and application to robust Doppler velocity estimation on smartphones in complex environments

  • 期刊论文

  • 发表日期：

    2026年4月24日

查看全文
  • Jingzhou Liu,
  • Yang Jiang,
  • Chun Wang,
  • Wei Ding,
  • Yang Gao

摘要

Accurate stochastic modeling of GNSS Doppler measurements is fundamental for velocity estimation in smartphones, yet conventional stochastic models have poor robustness in complex environments. The existing robust estimation methods require hyperparameter tuning and struggle with multipath/NLOS interference in complex environments. To address these limitations, we propose a novel stochastic modeling method involving Transformer-based Neural Network (TNN) to estimate Doppler-based velocity. First, the traditional Weight Least Squares and Solution Separation (SS) is conducted to get features for TNN training. The TNN then automatically learns the stochastic model and assigns measurement weights based on multi-source features including estimation residuals and SS-Statistics, enabling robust Doppler-based velocity estimation. The experiment results using smartphone GNSS in suburban environments show that the accuracy of the velocity solutions using the new stochastic model is improved compared to the use of traditional methods such as the Huber equivalent weight function and the elevation-based stochastic model.