Classification, detection and 3-DOF pose estimation by Convolutional Neural Networks.

An algorithm for determining the 3-DoF pose of an object detected by a Convolutional Neural Network is presented. In particular, the first chapter of this thesis presents a summary of what a Convolutional Neural Network is, what is intended by object detection, and describes popular object detection frameworks, such as Faster R-CNN, SSD, and YOLO. It continues by describing the MobileNets network architecture, and concludes by justifying the choice of object detection network-framework couple. In Chapter 2, it describes a procedure for achieving the 3 DoF pose estimation goal, which consists in obtaining the contours of such an object by binarizing the bounding box provided by the detector, and later exploiting the corresponding image moments to calculate its orientation. Finally, possible application scenarios are proposed.