An Easy-portable Displacement-offset-based Trajectory Prediction Method for Autonomous Vehicles

Abstract

Trajectory prediction algorithm is an important component of autonomous driving system (ADS) or advanced driver assistance system (ADAS), which enable autonomous vehicles to evaluate critical tasks in advance thus reduce vehicle collisions and improve traffic safety. Most existing trajectory prediction methods suffer from difficulties in portability and application across coordinate systems. To address these difficulties, this study proposes an easy-portable, vehicle displacement-offset-based trajectory prediction model, which can be rapid deployed and reused in different highway road sections and does not require second training. Specifically, first a novel trajectory sampling method to homo-dimension the vehicle data of different lengths is proposed. Second, the traditional neural network used for sequence prediction is modified to develop an enhanced trajectory prediction model, and which is trained and tested using the input reduction method. Finally, the trained model is saved locally and embedded in a co-simulation environment composed of CARLA, SUMO and Keras, afterwards the proposed method is tested in real-time simulation and across coordinate systems. The experimental results show that the proposed method can be quickly ported and deployed for reuse without second training, and has a fairly high prediction accuracy and long prospective time.