The aim of the HAPI project is to enhance a handheld device, such as the Laser Speckle Contrast perfusion imager, that scans a target area by manually pointing the device at the diseased skin area. To provide 3D geometrical information, visual SLAM (vSLAM) algorithms are involved.
In this work, several SLAM algorithms are chosen that are suitable to perform in a feature-based vSLAM framework. First, different types of Bayesian-based filters were presented. Based on these filters, four common SLAM algorithms were selected: EKF SLAM, SEIF SLAM, Graph SLAM and FAST SLAM. The working principle of the algorithms is explained in this report, including a mathematical derivation.
The performance of these algorithms is measured using an experiment in a 2D environment. Furthermore, the algorithms are tested in a feature-based vSLAM framework in a 3D environment. The input data used in these experiments consist of feature points extracted from real images and simulated feature points.
During the experimentation, it is observed that Graph SLAM provides relatively high accuracy and consistency. However, the algorithm is computationally expensive and therefore it could not be run in real-time. The results of the experiments show that EKF and SEIF have almost the same accuracy. The SEIF algorithms are computationally more attractive, but the needed computation power was higher than expected. It is also shown that the sparsification step in SEIF ensures a low constancy of the filter. Furthermore, the FAST-SLAM algorithm is only tested on the 2D data and shows a relatively high error in the pose and the map.
The research shows that there are several SLAM-back ends that can perform in a feature-based framework. Depending on the needed accuracy, the available computation power, the available knowledge to implement and the need to perform in real-time, a SLAM back-end can be chosen. However, to measure the performance in a more realistic environment, the algorithms need to be tested in a more nonlinear situation with more loop closings.
To join the presentation via Microsoft Teams click here