Feature based Monocular Visual Odometry for Autonomous Driving and Hyperparameter Tuning to Improve Trajectory Estimation

Abstract

In this paper, we will present an efficient feature-based localization method and use hyperparameter tuning to improve the accuracy of the method. We will first address the advantages of vision-based localization algorithms. These algorithms are more cost-efficient without GPS and are easier to integrate with vision based perception algorithms such as object detection and semantic segmentation. With camera equipped on cars, visual simultaneous localization and mapping, also called visual SLAM, improves many applications such as autonomous driving and advanced driver-assistance systems (ADAS). Currently, there are many different Visual SLAM algorithms and we will analyse the advantages of featured-based sparse methods. In the paper, we will briefly talk about the pipeline of our proposed method including feature extraction, key points matching, 8-point RANSAC algorithm and single value decomposition (SVD). Then we will use the grid search method to tune the parameters in the algorithm and improve the accuracy of the featured based visual localization method.