Multi-Task Prediction of Collision and Trajectories Based on Transformer Network for Safety-Critical Scenarios of Automated Vehicles 위험 시나리오에 대한 트랜스포머 네트워크 기반 자율주행차량의 충돌 및 궤적 예측 알고리즘 개발

Abstract

This research proposed a method for predicting collisions and trajectories using a transformer network with parallel computing capabilities for multiple vehicles. The accurate prediction of the driving trajectories of surrounding vehicles is essential for the decision-making processes of autonomous vehicles(AVs). Furthermore, the ability to predict imminent collisions can significantly enhance the safety of AVs. However, although several studies have addressed these issues individually, it is rare to find research that tackles both problems simultaneously. Hence, to facilitate the multitasks of collision prediction and trajectory prediction, we modified the prediction head associated with the final layer of the network to enable multitasking capabilities. This study explored two model architectures: one with an encoder-decoder structure and another with only a decoder. The performance of the proposed algorithm is compared with existing algorithms in the literature. The results demonstrate that our suggested algorithm outperforms previous methods in terms of parallelization.