Multi-scale graph-transformer network for trajectory prediction of the autonomous vehicles

The accurate trajectory prediction is a crucial task for the autonomous vehicles that help to plan and fast decision making capability of the system to reach their destination in the complex road scenario with abiding by the traffic rules. For this, autonomous vehicles should have more attention to their goal without affecting the other's task and maintain their safety from road accidents. With this motivation, we proposed a multi-scale graph-transformer-based attention mechanism that provides the interaction between the road agents with different time instances, because from time to time, few new agents may enter the frame scene, and few may leave the frame scene. Each dynamic obstacles trajectory can be defined as state sequences within an interval of time, where spatial coordinates of dynamic obstacles represented by the each state under the world coordinate frame. We have presented graph-based Multi-scale spatial features with transformer network that achieves significant prediction results compared to other existing methods, and we provide an in-depth analysis of the trained weights for different highways scenarios with transformer and the Long-Short Term Memory. We evaluate our model with three publicly available datasets and achieve state-of-the-art performances as presented in the manuscript. The performance balance is more in favour of our model for sparser datasets compared to the dense datasets.