Date of Award

Summer 6-12-2019

Publication Type

Master Thesis

Degree Name

M.Sc.

Department

Computer Science

First Advisor

Boufama, Boubakeur

Keywords

Artificial Intelligence, Change Detection, Computer Vision, Deep Learning, Monocular Vision, Obstacle Detection

Rights

info:eu-repo/semantics/openAccess

Creative Commons License

Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License.

Abstract

We explore change detection using videos of change-free paths to detect any changes that occur while travelling the same paths in the future. This approach benefits from learning the background model of the given path as preprocessing, detecting changes starting from the first frame, and determining the current location in the path. Two approaches are explored: a geometry-based approach and a deep learning approach. In our geometry-based approach, we use feature points to match testing frames to training frames. Matched frames are used to determine the current location within the training video. The frames are then processed by first registering the test frame onto the training frame through a homography of the previously matched feature points. Finally, a comparison is made to determine changes by using a region of interest (ROI) of the direct path of the robot in both frames. This approach performs well in many tests with various floor patterns, textures and complexities in the background of the path. In our deep learning approach, we use an ensemble of unsupervised dimensionality reduction models. We first extract feature points within a ROI and extract small frame samples around the feature points. The frame samples are used as training inputs and labels for our unsupervised models. The approach aims at learning a compressed feature representation of the frame samples in order to have a compact representation of background. We use the distribution of the training samples to directly compare the learned background to test samples with a classification of background or change using a majority vote. This approach performs well using just two models in the ensemble and achieves an overall accuracy of 98.0% with a 4.1% improvement over the geometry-based approach.

Share

COinS