Tools
Tools
Goswami, Rahul (2024) Advancements in Steering Angle Prediction: Deep Learning Approaches for Self-Driving Car. Masters thesis, Dublin, National College of Ireland.
Preview |
PDF (Master of Science)
Download (5MB) | Preview |
Preview |
PDF (Configuration Manual)
Download (653kB) | Preview |
Abstract
Deep learning and computer vision have been driven forward, with time, as the demand in the self-driving automobile sector increases. Under this effort, three unique models will be built and tested for use in self-driving car applications. Training shall make use of the Udacity Self-Driving Car Simulator. Each of these shall be oriented toward making predictions of the steering angle, hence enabling a self-driving car to make its way through a course completely independently of any human intervention. Similarly, three models were built using this dataset of images from the center, left, and right camera with values for steering angle, throttle, brake and speed NVIDIA Model, PilotNet Model, and Custom CNN Model.
Each model was developed under diverse architectural decisions with an eye on improving the accuracy of the steering predictions, taking into consideration various environmental conditions available in the simulator. With that extensive data preprocessing and augmentation were performed: cropping, resizing, changing brightness, and adding Gaussian blur; all this to make sure the dataset covers everything and is representative for most driving conditions. Later on, the performance of models trained based on a mean squared error metric and their generalization capability on previously unseen driving environments was measured. The primary aim of the study was to determine which among the three models-the NVIDIA Model, PilotNet Model, or Custom CNN Model-provides the most consistent and accurate steering prediction to enable safe and reliable autonomous navigation
Actions (login required)
 |
View Item |