Shuhul Mujoo

How do we improve loop detection for SLAM?

Started Aug. 17, 2021

Abstract or project description

The goal of the project is to improve upon loop detection for SLAM using novel methods created for solving the same problems used for bag of words and feature-based image clustering. Loop detection is a method in simultaneous localization and mapping (SLAM) that is used to detect when a robot's pose has reached a previously known location and close the path. The current state of the art methods are bag of words and feature image clustering which both use a symbolic representation of landmarks to determine when a location has been reached for a second time. The paper will compare both prior methods as well as improvements to the methods, and alternatives to those methods developed by Shuhul. The comparison will be done on publicly available datasets using a common baseline SLAM algorithm. In this paper, we present a method of loop closure detection that relies on both geometric and human salient features. Our algorithm is built on SuperPoint, which identifies a set of interest points based on geometric features, and SalientDSO that creates a saliency map from which features are extracted. The SuperPoint model is inherently prone to rotations, so we retrained it on rotated images to make it more resistant. Both the point and semantic features are then fed into ORB SLAM to return a set of features. Finally a bag of words model is used to detect loop closure. A more detailed version is as follows. Keyframes are extracted from the video from the monocular camera. Keyframes are fed separately into SuperPoint feature extractor and SalGan feature extractor. Salgan returns a heatmap which is then processed such that the maximum gradient points are selected. These are then combined into a custom feature with a custom feature descriptor based off of ORB. These then are used in only the loop closure part of ORB SLAM by being added to ORB SLAMs built in bag of words model. Along with a custom vocabulary this then is compared against a threshold, which determines whether there was a loop closure. The algorithm is then tested with the New College and Tum Mono datasets. Overall this is a much more reliable solution to loop closures as the number of false positives is drastically lower than ORB SLAMs built in loop closure. The number of false negatives is also low, comparable to the built- in loop closure.