Itransformer Neural Network Based Traffic Aware Motion Prediction for Autonomous Highway Driving

Autonomous vehicles require accurate, traffic-aware motion prediction to offer various advantages over manual driving, including convenience, reduced traffic collisions, fuel savings and mobility efficiency. To gain widespread global adoption, autonomous vehicles must recognise potential dangers in real-time traffic conditions to increase driving safety. Understanding driver intentions and improving methods for predicting the trajectories of surrounding vehicles is crucial for adaptive traffic-aware autonomous driving. In this paper, we explore a unique application of the iTransformer (inverted Transformer) neural network, which uses real-time surrounding traffic data to enhance the ego vehicle's speed prediction by capturing complex interactions. This implementation adopts multivariate forecasting, taking advantage of mapping the high-dimensional features to low-dimensional spaces by using the inverted self-attention mechanism to learn the distinctions and interactions between the time-series data. The training of the iTransformer is done using the popular open-source NGSIM dataset. The proposed iTransformer neural network improves the root mean square error (RMSE) by 71.69 % and 57.31 % when compared against Long Short-Term Memory (LSTM) and Vanilla Transformer (VT), respectively.