DynSLAM

Why we included this project

DynSLAM is a stereo SLAM system that takes on a problem most mapping pipelines quietly skip: streets full of moving cars. It takes stereo input, estimates depth with either ELAS or a learned dispnet model, segments the image to find vehicles, and then reconstructs the static background and each moving object as separate volumetric models. That split is what makes it worth studying, since it lets you build a clean map of a city scene without the ghosting artifacts moving vehicles usually leave behind. The code is a research prototype built on InfiniTAM and tied to a specific ICRA 2018 paper, so it reads better as a reference implementation and a starting point than as a drop-in library. It also ships the raw data and Jupyter notebooks used to regenerate the paper's figures, which makes checking the claims and tracing the internals fairly painless.