Modelling Object Movement Around an Ego Vehicle
Publicerad
Författare
Typ
Examensarbete för masterexamen
Program
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
One problem for automated vehicles is that the tra c environment surrounding a vehicle is diverse and
populated by a large set of interacting agents in the form of drivers. With a model for vehicle movement in
tra c, developers will be able design autonomous vehicles with better path planning functionalities.
In this thesis models for vehicle movement around an ego vehicle are developed in a data-driven manner
with di erent machine learning techniques. Analysis is done to nd how accuracy is related to the prediction
horizon, and to determine which features are most important. It is clear that more features are not always
better as removing unnecessary features provides better results.
When comparing the models, baselines based on equations of motion with constant velocity or acceleration
have been used. All methods provide better predictions compared to the baselines, and can make predictions
for longer horizons. For longitudinal position prediction, results are promising. In latitudinal direction the
results are less impressive, especially lane changes are di cult to predict, due to the low amount of lane changes
in the training data.
That leads to analyzing in what other situations the prediction accuracy is limited by the data set, rather
than by the model itself. For example how the accuracy is correlated with the speed of the ego vehicle. It
is clear that the models performs best in situations that is well represented in the training data. To make a
model that handles rare situations, a lot of data with those situations is needed.
Beskrivning
Ämne/nyckelord
Autonomous Vehicles, Machine Learning, Data-driven, Feed Forward Neural Network, Linear Genetic Programming, Principal Components Analysis, Fishers Linear Discriminant, Regression, Path Prediction