Augmented Reality For Indoor Navigation In A Warehouse Setting: Examining Supervised Learning-Based Pretrained CNNs for Feature Extraction in Visual Odometry
Author
Petersson, Bastian Starup
Term
4. term
Publication year
2023
Submitted on
2023-05-30
Pages
53
Abstract
This report examines how Augmented Reality (AR) can improve warehouse management tasks such as picking, receiving, and relocating items. AR overlays digital information on the real-world view and can guide workers through the warehouse. The work focuses on the technical needs for warehouse navigation, especially Simultaneous Localization and Mapping (SLAM), which builds a map while estimating position for accurate tracking. It also looks at visual odometry, that is, estimating motion from camera images. We propose a deep learning-based method to extract image features for visual odometry in large industrial environments. The method combines the FAST keypoint detector with pre-trained convolutional neural networks (CNNs) to produce feature descriptors, and we evaluate it against a classical technique, ORB. Tests showed that ORB was more efficient and more accurate than the deep learning-based methods. Future work includes studying which layers in the pre-trained networks yield the best features and using GPUs (graphics processors) to speed up computation.
Denne rapport undersøger, hvordan udvidet virkelighed (Augmented Reality, AR) kan forbedre lagerstyring i opgaver som plukning, modtagelse og omplacering af varer. AR lægger digital information oven på det, man ser, og kan guide medarbejdere gennem lageret. Arbejdet fokuserer på de tekniske krav for lager-navigation, især samtidig lokalisering og kortlægning (SLAM), som beregner position og bygger et kort på samme tid for mere præcis sporing. Derudover undersøges visuel odometri, dvs. at beregne bevægelse ud fra kamerabilleder. Vi foreslår en deep learning-baseret metode til at udtrække billedfeatures til visuel odometri i store industrielle miljøer. Metoden kombinerer FAST-nøglenpunktsdetektoren med forhåndstrænede konvolutionsneuronale netværk (CNN'er) for at danne feature-beskrivelser, og den sammenlignes med en klassisk teknik, ORB. Testene viste, at ORB var både mere effektiv og mere nøjagtig end de deep learning-baserede metoder. Fremtidigt arbejde omfatter at undersøge, hvilke lag i de forhåndstrænede netværk der giver de bedste features, samt at udnytte GPU'er (grafikprocessorer) til hurtigere beregning.
[This apstract has been rewritten with the help of AI based on the project's original abstract]