Advanced Interaction-aware Motion Models for Motorcycle Trajectory Prediction: Experiments on pNEUMA Datasets

Forecasting the motion of motorcycles is a critical task for an autonomous system deployed in complex traffic, considering its distinguished characteristics compared to other vehicles. Motion of motorcycles in a scene is governed by the traffic context, i.e., the motion and relative spatial configuration of neighboring vehicles. In this thesis, we propose to use two advanced interactionaware motion models encompassing the dynamic interaction of the vehicles and maneuver-based encoding, CS-LSTM and PiP, along with their variation models developed specifically for motorcycle trajectory prediction. We evaluate our model using the publicly available pNEUMA datasets. Our results demonstrate the feasibility and superiority of the motorcycle-specific models and their improvements during evaluation in terms of RMSE values.