Scholarlink

Trajectory prediction plays a pivotal role in intelligent transportation systems, enabling the anticipation of vehicle locations to optimize routes, reduce travel times, and enhance overall transportation efficiency. This is particularly crucial for dynamic carpooling services, where taxis and users must continuously adapt to evolving passenger needs, traffic conditions, and other unpredictable factors. In this study, we present an innovative autoregressive approach framed as a classification task for predicting taxi trajectories, leveraging the BERT (Bidirectional Encoder Representations from Transformers) architecture, originally developed for natural language processing. Our method integrates contextual factors tailored to the operating environment, in order to better capture the complexities of trajectory prediction in challenging urban settings. This approach is especially relevant for the EzyLi platform, currently in beta testing in Africa, where chaotic and unpredictable road conditions demand robust solutions. Preliminary results demonstrate the promise of this context-aware methodology, achieving an accuracy of 83.3% based on this challenging context.

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

Towards Predicting Complex Carpooling Trajectories with Context-Augmented BERT-LLM in Chaotic Environments

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Daril Kengne,
  • Roger Nkambou,
  • Ange Tato,
  • Clara Lacourarie

摘要

Trajectory prediction plays a pivotal role in intelligent transportation systems, enabling the anticipation of vehicle locations to optimize routes, reduce travel times, and enhance overall transportation efficiency. This is particularly crucial for dynamic carpooling services, where taxis and users must continuously adapt to evolving passenger needs, traffic conditions, and other unpredictable factors. In this study, we present an innovative autoregressive approach framed as a classification task for predicting taxi trajectories, leveraging the BERT (Bidirectional Encoder Representations from Transformers) architecture, originally developed for natural language processing. Our method integrates contextual factors tailored to the operating environment, in order to better capture the complexities of trajectory prediction in challenging urban settings. This approach is especially relevant for the EzyLi platform, currently in beta testing in Africa, where chaotic and unpredictable road conditions demand robust solutions. Preliminary results demonstrate the promise of this context-aware methodology, achieving an accuracy of 83.3% based on this challenging context.