The Effects of Image Augmentation on Efficiency of a Convolutional Neural Network of a Self-Driving Car

The objective of this paper was to study the effectiveness of image augmentation techniques in training a Convolutional Neural Network (CNN) of a self-driving car and identify the most suitable form of image augmentation technique, using the Udacity Car Simulator. Firstly, a dataset of augmented and non-augmented images from a training track, consisting of left-, right-, and front-facing views from the car cameras was created. Various image augmentation techniques were used: zoom, brightness, pan, flip, random (augments the image by arbitrarily choosing a technique from the previous four), and no augmentation. Secondly, training datasets consisting of the aforementioned images and a log of car turning angles, throttle, and brake were built. The final training datasets were then used with NVIDIA training method to train different CNN. The different trained networks generated steering commands from the front-facing camera of the simulation and test track had no effect on the generalization of the CNN. Lastly, different trained networks were used on the test track of Udacity Car Simulator to calculate the following variables: distance travelled, and number of crashes made by the car. After these values were acquired, an efficiency analysis was performed. The results suggested augmentation of training data is a crucial factor when it comes to the process of generalizing a model to perform tasks. Random augmentations performed the best, but a combination of flip and brightness augmentations performed equally efficiently.