Prediction of Vehicles' Trajectories based on Driver Behavior Models

Prediction of Vehicles' Trajectories based on Driver Behavior Models

Li, Z.

Hellendoorn, J. (mentor)
Mazo, M. (mentor)
Hoogendoorn, R.G. (mentor)

Mechanical, Maritime and Materials Engineering

Delft Center for Systems and Control

System and Control

2014-08-21

As a component of Dutch Automatic Vehicle Initiative (DAVI) project, this study aims at improving highway driving safety of autonomous vehicle. It is observed that some misbehaved drivers do not use turn indicators forehead a lane change on highway. For a self-driving car in such situations, a lane change is potentially dangerous without an accurate estimation of other vehicles’ movements .If lane changes can be detected or predicted at its initial phase, DAVI vehicle can be advised in time to perform corresponding maneuvers to avoid collision. In this study, a Support Vector Machine (SVM) based method was proposed to fulfil this task. Predictions are generated through analysing vehicles’ motion parameters. At the first place, a number of driving simulator tests were carried out to set up database of typical vehicle maneuvers features, including heading angle, yaw rate, lateral velocity and lateral acceleration. 14 driver databases were used in SVM training after removal of unrealistic info. An off-line trained SVM was obtained to perform lane change predictions. In addition to offline training, an incremental training SVM was introduced. Compared with off-line learning, its incremental feature enables the classifier to update itself with freshly recorded lane change data in a relatively short time. This enables the vehicle to be "learning while predicting" . Based on the databases, SVM based approaches were verified to be feasible of predicting lane changes. With the most optimal parameter combination, this method is able to perform predictions with 100% sensitivity ( predicted all lane changes successfully). Average advance time and average computational time are acceptable for automatic driving. Besides, the performance of sliding window method was evaluated for variation of its size, and a general applicability of overall prediction method was also examined on data from different drivers.

lane change
SVM
incremental training

http://resolver.tudelft.nl/uuid:9f93374b-7f79-4633-a02f-ac81f5e2da33

Student theses

master thesis

(c) 2014 Li, Z.