Trajectory Prediction

Transformer Network to predict Trajectories for traffic agents

In realistic traffic scenarios, trajectory prediction is important to guarantee the safety of an autonomous vehicle. To predict future trajectories, interactions between surrounding traffic are needed to be modelled. While modelling such interactions, the use of local semantic maps is helpful as it determines the presence of obstacles and other surrounding traffic agents in the traffic scene.

Objectives

• Learning spatio-temporal relationships between the traffic agents to improve prediction accuracy.
• Fusion of diverse data-types in the form of numerical data, categorical data and local semantic maps.

Results

The experiments show that adding local semantic maps encoded with a CNN to the input of the Transformer-based model, significantly improves prediction accuracy.