Year: 2018
Type: dissertation
Abstract: Autonomous vehicles are currently being heavily researched with the purpose of finding a manner in which a vehicle can drive safely.An approach for doing this is imitation learning by which human behavior is imitated, and which recently has found success in steering vehicles using images.For this thesis, the purpose is to investigate whether imitation learning can be applied to path planning for autonomous vehicles using several sources of input data.We also want to investigate if merging different types of information will improve the safety and robustness of the vehicle.The aim is therefore to design and implement a neural network which merges input data from several sources and predicts the path of the vehicle 5 seconds into the future.This thesis compares a variety of neural network models used to predict the path.Some models use data from only one source, while other merge the data.The results show that it is possible to use deep learning to predict the path for autonomous vehicles in some situations.They also show that including images as input to the model does not significantly improve the performance.Instead, using only object data make the networks performs as good as merging the inputs.However, more research are needed in order to find a more accurate and stable model.
Authors: Maria Bergqvist, Oskar Rödholm
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2884115443