Naturalistic Driver Intention and Path Prediction using Machine Learning

Abstract

Autonomous vehicles are still yet to be available to the public. This is because there are a number of challenges that have not been overcome to ensure that autonomous vehicles can safely and efficiently drive on public roads. Accurate prediction of other vehicles is vital for safe driving, as interacting with other vehicles is unavoidable on public streets. This thesis explores reasons why this problem of scene understanding is still unsolved, and presents methods for driver intention and path prediction. The thesis focuses on intersections, as this is a very complex scenario in which to predict the actions of human drivers. There is very limited data available for intersection studies from the perspective of an autonomous vehicle. This thesis presents a very large dataset of over 23,000 vehicle trajectories, used to validate the algorithms presented in this thesis. This dataset was collected using a lidar based vehicle detection and tracking system onboard a vehicle. Analytics of this data is presented. To determine the intent of vehicle at an intersection, a method for manoeuvre classification through the use of recurrent neural networks is presented. This allows accurate predictions of which destination a vehicle will take at an unsignalised intersection, based on that vehicle's approach. The final contribution of this thesis presents a method for driver path prediction, based on recurrent neural networks. It produces a multi-modal prediction for the vehicle’s path with uncertainty assigned to each mode. The output modes are not hand labelled, but instead learned from the data. This results in there not being a fixed number of output modes. Whilst the application of this method is vehicle prediction, this method shows significant promise to be used in other areas of robotics.