Scholarlink

In this paper, we present a data-driven approach for predicting lane changes of vehicles in a highway scenario based on observing position, speed, and movements of surrounding vehicles. To train a prediction model, we employ federated learning (FL) with various locations acting as clients. The study employs Long Short-Term Memory (LSTM) networks that utilize 1 s of historical data to forecast lane changes over a 1, 3 and 5-s prediction horizon. We show that personalized FL performs well for a distributed setup without data sharing. The findings demonstrate FL’s potential in automotive safety applications, nearly matching centralized performance while significantly improving data security and privacy across distributed locations. This study supports using federated learning as a viable and robust solution for privacy-preserving predictive tasks in dynamic environments.

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

Federated Learning for Lane-Change Prediction

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Lilit Yenokyan,
  • William Lindskog-Muenzing,
  • Christian Prehofer,
  • Matthias Schubert

摘要

In this paper, we present a data-driven approach for predicting lane changes of vehicles in a highway scenario based on observing position, speed, and movements of surrounding vehicles. To train a prediction model, we employ federated learning (FL) with various locations acting as clients. The study employs Long Short-Term Memory (LSTM) networks that utilize 1 s of historical data to forecast lane changes over a 1, 3 and 5-s prediction horizon. We show that personalized FL performs well for a distributed setup without data sharing. The findings demonstrate FL’s potential in automotive safety applications, nearly matching centralized performance while significantly improving data security and privacy across distributed locations. This study supports using federated learning as a viable and robust solution for privacy-preserving predictive tasks in dynamic environments.