Lane change detection and prediction using real-world connected vehicle data

Abstract

Prediction of lane changes (LCs) provides critical information to enhance traffic safety and efficiency in a connected and automated driving environment. It is essential to precisely detect LCs from driving data to lay the groundwork for LC prediction. This study aims to develop LC detection and prediction models using large-scale real-world data collected by connected vehicles (CVs). At first, an autoencoder was used to detect LCs, and proved to be more precise and robust than conventional methods. Next, a transformer-based LC prediction model was developed, which concentrated computation power on key information via an attention mechanism. It outperformed the baseline models in terms of accuracy and computational efficiency. The prediction horizon was also analyzed and LC could be accurately predicted up to two seconds in advance. At last, the transformer model was implemented for real-time prediction and demonstrated a great potential for practical applications.