Object classification and localization using machine learning techniques

Abstract

When working with object classification and localization in image data, the development of traditional rule-based solutions has stagnated in recent years. In its place, machine learning has become a major field of research in order to handle more and more complex image recognition problems. With machine learning, new state-of-the-art models can be developed by training a model instead of implementing an explicitly programmed feature detector. In this thesis, a literature study covering the field of machine learning has been carried out on behalf of Volvo Advanced Technology and Research. Furthermore, with an autonomous garbage handling project initiated by Volvo in mind, two machine learning models meant for limited hardware-deployment have been designed and trained. The classification model is based on knowledge distillation, where a compact model learns to generalize from a more complex state-of-the-art model, and a localization model, where a typical machine learning implementation is combined with computer vision solutions from the OpenCV framework. Both models, that were trained on images from the ImageNet database, produced poor results in their respective tasks. The process of knowledge distillation, used to train the classifier, was not achievable due to unfortunate choice of cumbersome model combined with hardware limitations during training. The hardware was also an issue for the localization model, which due to this and unwanted performance from the OpenCV corner detector converged early during training and ended up producing unchanged results for different input. However, the thesis as a whole came to important conclusions regarding a proper next step in order to stay competitive within the field of machine learning.