2026-05-22T17:05:26Zhttps://keep.lib.asu.edu/oai/request

oai:keep.lib.asu.edu:node-1992152024-12-23T18:01:48Zoai_pmh:alloai_pmh:repo_items

199215 https://hdl.handle.net/2286/R.2.N.199215 http://rightsstatements.org/vocab/InC/1.0/ All Rights Reserved 2024 32 pages Masters Thesis Academic theses Text eng Bhatt, Zeel Shaileshkumar Yang, Yezhou Jayasurya, Suren Zhang, Yu Arizona State University Partial requirement for: M.S., Arizona State University, 2024 Field of study: Engineering Visual Odometry is an essential component in a Visual SLAM system. The aim of Visual SLAM (Simultaneous Localization and Mapping) is to create a 3D map of the world while simultaneously estimating the camera’s position in the created map. Visual odometry attempts to estimate the camera’s motion by analyzing the changes in images due to camera movement. This element is pivotal in various fields, including SLAM, 3D reconstruction, augmented reality, and more. A classic pipeline for visual odometry consists of camera calibration, feature detection, feature matching, triangulation, and local optimization (Bundle Adjustment). This geometry-based method has been broadly implemented in various SLAM algorithms. On the other hand, deep learning-based methods have dominated many computer vision tasks, but learning-based visual odometry is not on par with strong geometric methods. This is attributed to the insufficient diversity of data and scale ambiguity in triangulation. These issues can explicitly be addressed using Contrastive Learning. Contrastive learning utilizes data augmentation to improve learning performance, allowing it to derive benefits from sparsely diverse data available for visual odometry tasks through valid augmentations. Contrastive loss pulls feature vectors of negative pairs far apart while keeping positive pairs close together in latent space. This control over stretching in latent space can be highly useful for visual odometry problems. Robotics Artificial Intelligence Autonomous Vehicle Contrastive Learning Visual Localization Visual Odometry Learning to Localize: A Contrastive Approach