Matih Ullah, "Landmark Independent Visual Localization with Slow Feature Analysis", University of Applied Sciences Frankfurt , 2022.

Visual localization is an area of interest to research in mobile robots, self-driving cars, etc. Localization using a camera is one of the fundamental requirements for a vision-based mobile robot. In this context, unsupervised learning with Slow Feature Analysis (SFA) directly applies to the images to extract a spatial representation of the environment. In the past, SFA is used with Convolutional Neural Network (CNN) to perform localization. This method produces good results, but it relies on either hand-label data or pre-trained CNN detectable objects as anchors. To bypass these methods, this thesis includes research work with the OpenCV tracker for data collection that is used to train the SFA network to perform localization. This complete pipeline removes the dependency on learning pre-fixed visual landmarks for localization. The image patches are collected with the help of an OpenCV tracker from the scene to train the SFA network. Robot localization is performed by randomly selecting image patches that are treated as a generic landmarks from the scene. Based on these generic landmark’s robot can localize itself. The proposed method can be performed with limited computational resources and does not require hand-labeled data, which gives it an advantage over expensive deep learning methods. In this thesis, different experiments were performed, and results from the data were collected and tabulated.