Transformer based composite network for autonomous driving trajectory prediction on multi-lane highways

In order to navigate through complex traffic scenarios safely and efficiently, the autonomous vehicle (AV) predicts its own behavior and future trajectory based on the predicted trajectories of surrounding vehicles to avoid potential collisions. Further, the predicted trajectories of surrounding vehicles (target vehicles) are greatly influenced by their driving behavior and prior trajectory. In this article, we propose a novel Transformer-based composite network to predict both driver behavior and future trajectory of a target vehicle in a highway driving scenario. The powerful multi-head attention mechanism of the transformer is exploited to extract social-temporal interaction between target vehicle and its surrounding vehicles. The prediction of both lateral and longitudinal behavior is carried out within the behavior prediction module, and this additional information is further utilized by the trajectory predictor module to ensure precise trajectory prediction. Furthermore, mixture density network is augmented in the model to handle uncertainties in the predicted trajectories. The proposed model's performance is compared with several state-of-the-art models on real-world Next Generation Simulation (NGSIM) dataset. The results indicate the superiority of the proposed model over all contemporary state-of-the-art models, as evaluated using Root Mean Square Error (RMSE) metric. The proposed model predicts a 5s long trajectory with an 11% lower RMSE than the state-of-the-art model.