UNF-SLAM(ULNC-SLAM) is an abbreviation of Unsupervised Network for Fusion SLAM. In this work, we present a visual-laser network to extract features for SLAM. Building on a previous work which can complete and predict dense depth and confidence, we feed RGB frames and confidence maps into our network to generate feature points that have value and can utilize reliable information from the laser. We partially modified [ORB-SLAM2], (https://github.com/raulmur/ORB_SLAM2) to obtain a more accurate and robust SLAM system. Also, our system can be easily extended to monocular or stereo SLAM.
Performance UNF features:
We use the new thread and chrono functionalities of C++14.
We use Pytorch C++ api(libtorch) for deloying the UNF.
We use Pangolin for visualization and user interface. Dowload and install instructions can be found at: https://github.com/stevenlovegrove/Pangolin.
We use OpenCV to manipulate images and features. Dowload and install instructions can be found at: http://opencv.org.
Required at least OpenCV4.5.1.
Required by g2o (see below). Download and install instructions can be found at: http://eigen.tuxfamily.org.
Required at least 3.1.0.
We use modified versions of the DBoW2 library to perform place recognition and g2o library to perform non-linear optimizations. Both modified libraries (which are BSD) are included in the Thirdparty folder.
Clone the code
git clone https://github.com/salt0107fish/UNF_SLAM-master.git
Then build the project
cd UNF_SLAM-master
./build.sh
Make sure to edit build.sh pointing to your local libtorch installation. Edit run_ulnc.sh to check out how to run with UNF_SLAM.
The ORB-SLAM2 localization result in KITTI05 is
The UNF-SLAM localization result in KITTI05 is
RPE in Meters for KITTI 05 is:
| Configuration | max | mean | median | min | rmse | sse | std |
|---|---|---|---|---|---|---|---|
| ORB-SLAM2 | 10.885378 | 0.107923 | 0.081827 | 0.003509 | 0.352939 | 134.779983 | 0.336033 |
| UNF-SLAM | 5.864444 | 0.059092 | 0.035320 | 0.002328 | 0.202411 | 74.852742 | 0.193593 |
The data used for training and testing are from KITTI data set. The depth predictions and confidence maps are generated by Sparse-Depth-Completion Network .(https://github.com/wvangansbeke/Sparse-Depth-Completion)
The basic framework of UNF-SLAM comes from [ORB-SLAM2]. (https://github.com/raulmur/ORB_SLAM2)