Scholarlink

<p>Accurate long-term trajectory prediction is critical for vehicle safety in complex traffic. Existing models often suffer from error accumulation, poor interpretability, and weak integration of driving behavior. This paper proposes a multi-level personalized trajectory prediction framework (DS-TCTM). A K-Means++ enhanced GMM (K-GMM) classifies driving styles into conservative, neutral, and aggressive. The classification uses acceleration variation rate and average headway time under different traffic densities. A comprehensive style label is generated through density weighting and scoring. The prediction module applies a cascaded GRU, BGRU, and Bi-LSTM structure for longitudinal modelling. Three parallel style-specific sub-networks are used to reduce cross-style interference. Experiments are conducted on the SQM dataset from the UTE project. DS-TCTM achieves lower RMSE and NLL than LSTM, Social-LSTM, CS-LSTM, SV-LSTM, TransTM, GRUTrans and TCTM without style modelling. At 1&#xa0;s and 5&#xa0;s, RMSE is 0.11 and 4.46, while NLL is 0.51 and 3.89. The average inference time is 13.4&#xa0;ms. Results show that DS-TCTM improves accuracy and efficiency. It can be applied to active safety tasks such as collision warning and behavior understanding.</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

Vehicle trajectory prediction with driving style and multiple models

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月28日

查看全文
  • Heng Wang

摘要

Accurate long-term trajectory prediction is critical for vehicle safety in complex traffic. Existing models often suffer from error accumulation, poor interpretability, and weak integration of driving behavior. This paper proposes a multi-level personalized trajectory prediction framework (DS-TCTM). A K-Means++ enhanced GMM (K-GMM) classifies driving styles into conservative, neutral, and aggressive. The classification uses acceleration variation rate and average headway time under different traffic densities. A comprehensive style label is generated through density weighting and scoring. The prediction module applies a cascaded GRU, BGRU, and Bi-LSTM structure for longitudinal modelling. Three parallel style-specific sub-networks are used to reduce cross-style interference. Experiments are conducted on the SQM dataset from the UTE project. DS-TCTM achieves lower RMSE and NLL than LSTM, Social-LSTM, CS-LSTM, SV-LSTM, TransTM, GRUTrans and TCTM without style modelling. At 1 s and 5 s, RMSE is 0.11 and 4.46, while NLL is 0.51 and 3.89. The average inference time is 13.4 ms. Results show that DS-TCTM improves accuracy and efficiency. It can be applied to active safety tasks such as collision warning and behavior understanding.