TRAJECTORY PREDICTION OF DYNAMIC OBSTACLES FOR AUTONOMOUS VEHICLES

Abstract

Self-driving vehicles, also referred to as autonomous vehicles, will bring about many positive impacts. However, for safe and robust driving of AVs, object trajectory prediction is required. This thesis aims to solve the problem of object future trajectory prediction for autonomous vehicles in Singapore with the objectives of developing an annotated dataset of trajectories representative of the traffic in Singapore, developing algorithms and codes for methods of trajectory prediction, and subsequently testing and comparing the developed methods of trajectory prediction on the developed dataset.

Three approaches were developed and tested which are the least-squares fitting method, Kalman filter method, and the deep learning method. The deep learning method using the Autoencoder model gave the lowest error, with the average error across all timesteps of 3.22m.