REFINING DEEP LEARNING NEURAL NETWORKS FOR AUTONOMOUS VEHICLE NAVIGATION

Abstract

Machine learning methods have recently increased in appearance in navigation and guidance applications by means of image classification. This thesis sought to advance the ongoing Electrical and Computer Engineering (ECE) Control Systems and Robotics Laboratory project in developing a system that will autonomously navigate across the Naval Postgraduate School (NPS) campus. In pursuit of providing a robust navigation and guidance solution to an autonomous robotic vehicle, a convolutional neural network (CNN) was trained to classify significant objects around NPS. In addition to increasing the number of objects that the neural network could classify, the network was also trained with varying image augmentation techniques applied to the training and validation images. A variety of tests were performed to evaluate the accuracy of the model when exposed to different objects and regions throughout the campus. The tests also included running the image classification model against images that were altered with crop, blur, rotation, and noise. The results demonstrated high classification accuracy and asserted that the output was robust when faced with poor image quality. This work established a strong baseline for more CNN output integration into the navigation and guidance solution of the robotic vehicle.