Scholarlink

<p>The prediction of vehicle motion trajectories using artificial neural networks is considered. Existing approaches and methods for predicting vehicle motion trajectories along with their limitations are considered. The traffic prediction problem is presented in a scene-centric setting, using a simplified bird’s eye view (BEV) input representation and an occupancy map output representation. A semi-supervised physical maneuver model is proposed to reprocess the unimodal dataset and generate a multimodal dataset, solving the problem of limited dataset for training. A modified version of the Transformer artificial neural network architecture is used for trajectory prediction. The Transformer architecture is also further modified to handle large spatio-temporal relationships. The efficiency of the proposed method is compared with the two best agent-centric and scene-centric algorithms. The proposed method improves the efficiency of known methods by up to 40% in certain metrics and achieves results comparable to the best agent-centric approaches.</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

Semi-Supervised Prediction of Multimodal Vehicle Trajectory Based on Transformers

  • 期刊论文

  • 发表日期：

    2025年9月30日

查看全文
  • V. M. Sineglazov,
  • K. S. Lesohorskyi

摘要

The prediction of vehicle motion trajectories using artificial neural networks is considered. Existing approaches and methods for predicting vehicle motion trajectories along with their limitations are considered. The traffic prediction problem is presented in a scene-centric setting, using a simplified bird’s eye view (BEV) input representation and an occupancy map output representation. A semi-supervised physical maneuver model is proposed to reprocess the unimodal dataset and generate a multimodal dataset, solving the problem of limited dataset for training. A modified version of the Transformer artificial neural network architecture is used for trajectory prediction. The Transformer architecture is also further modified to handle large spatio-temporal relationships. The efficiency of the proposed method is compared with the two best agent-centric and scene-centric algorithms. The proposed method improves the efficiency of known methods by up to 40% in certain metrics and achieves results comparable to the best agent-centric approaches.