الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 92 |  |  |
| | | from | 92 | | |
Abstract
The basic objective of this thesis is to build a system, which stays in the middle of the lane during autonomous driving without slipping for transporting materials and equipment.
For improving the driving/braking performance under adhesion coefficient road, slip ratio is controlled by pid controller. Half model of vehicle is studied for the controlling of the slipping.
Using udacity as an open source simulator to depict a real-time environment with the help of a model trained by deep neural networks, the driving behavior of a human is mimiced on the simulator (Behavioral Cloning). Keras and TensorFlow are used as the backend for dataflow programming. Keras provides sequential models for building a linear stack of network layers. Convolutional Neural Network (CNN) models are neural network layers are optimized in series combinations of Time-Distributed Convolution Layers, Maxpooling, Flatten, Dense, Dropout and so on, which are experimented for reaching the best model of control. Finally, model 2 is performed as the best in the end of 60 epochs with the least loss = 0.007.
Reinforcement Learning with a prototype is used for a perfect way of working on the middle of the track where the distance between the camera and the real track is measured according to number of pixels at different positions and angles on the track. For calculating the mean of errors according to different numerical methods, Root mean square error (RMSE) & Mean absolute error (MAE) are used. The error values
(RMSE = 7.02 mm & MAE = 6.5 mm) are acceptable with the comparison to other method of control (Deep Learning Models). Reinforcement Learning algorithm is used for reaching the good performance on the track and the results are improved using the update equation through driving on different tracks.