Visual SLAM: Enhancing Direct Visual Odometry Through the Integration of Deep Learning Approaches
Author
Gjerde, Magnus Kaufmann
Term
4. term
Publication year
2023
Submitted on
2023-06-02
Pages
51
Abstract
Visual odometry estimates how a camera moves by analyzing a sequence of images. This thesis focuses on direct visual odometry and explores hybrid designs that combine deep learning with classical, hand-crafted techniques. We propose a system that integrates two components: a deblurring module that sharpens frames and a saliency predictor that identifies the most informative image regions. Together, they guide point sampling so the algorithm tracks the scene using better features, especially when frames are blurred by fast motion or long exposures in low light. In tests on the EuRoC MAV dataset, the method consistently improved trajectory estimates compared with two established systems, DSO and SalientDSO. The average Absolute Trajectory Error (ATE) was 0.26 m, versus 0.335 m for DSO and 0.303 m for SalientDSO. Future work will evaluate other image pre-processing methods—including dehazing, denoising, and general image enhancement—to further increase accuracy.
Visuel odometri beregner en kameras bevægelse ved at analysere en billedsekvens. Denne afhandling fokuserer på direkte visuel odometri og undersøger hybride løsninger, der kombinerer dyb læring med klassiske, håndlavede teknikker. Vi præsenterer et system, der kombinerer to komponenter: et sløringsfjernelsesmodul, som gør billeder skarpere, og en saliencyprediktor, som udpeger de mest informative områder i billedet. Sammen styrer de punktudvælgelsen, så algoritmen sporer scenen med bedre punkter, især når billeder er slørede af hurtige kamerabevægelser eller lange eksponeringstider i svag belysning. I tests på EuRoC MAV-datasættet gav metoden gennemgående bedre bevægelsesbaner end to etablerede systemer, DSO og SalientDSO. Den gennemsnitlige Absolute Trajectory Error (ATE) var 0,26 m mod 0,335 m for DSO og 0,303 m for SalientDSO. Fremtidigt arbejde omfatter at afprøve andre billedforbedringer, herunder dehazing, denoising og generel billedforbedring, for at øge nøjagtigheden yderligere.
[This apstract has been rewritten with the help of AI based on the project's original abstract]