Please use this identifier to cite or link to this item:
https://www.um.edu.mt/library/oar/handle/123456789/35198
Title: | Automatic vehicle detection from aerial imagery |
Authors: | Mallia, Stefan |
Keywords: | Neural networks (Computer science) Computer vision Image processing Vehicles -- Aerial photographs |
Issue Date: | 2018 |
Citation: | Mallia, S. (2018). Automatic vehicle detection from aerial imagery (Bachelor's dissertation). |
Abstract: | The detection of vehicles from aerial imagery is a problem that is well suited to Convolutional Neural Network models designed for object detection. The most recent advances in computer vision have come from research that strongly relies on these models. However, these models were designed to be trained and used on standard datasets and often do not perform well when trained for custom datasets. Apart from training, which is already a computationally expensive task, it is also important to take into consideration whether the architecture designed is also suitable. This task of model selection can be tackled as an optimisation problem, but this problem is difficult due to its computational complexity and non-differentiability. In this work, Bayesian Optimisation is performed on the Yolov2 architecture with the aim of finding an improved configuration. An initial set of 16 Yolov2 architecture configurations is specified and evaluated on a car park dataset, after which Bayesian Optimisation continued to search for better performing architectures. An improvement of 13.6 percentage points in Average Precision was attained over the best performing model in the initial set of architectures and 55.2 percentage points over the original Yolov2 architecture configuration. |
Description: | B.SC.(HONS)COMP.SCI. |
URI: | https://www.um.edu.mt/library/oar//handle/123456789/35198 |
Appears in Collections: | Dissertations - FacICT - 2018 Dissertations - FacICTCS - 2018 |
Files in This Item:
File | Description | Size | Format | |
---|---|---|---|---|
18BCS009.pdf Restricted Access | 2.03 MB | Adobe PDF | View/Open Request a copy |
Items in OAR@UM are protected by copyright, with all rights reserved, unless otherwise indicated.