Authors: Stefano Aldegheri, based upon Raul Mur-Artal (ORB-SLAM2)

This implementation aims to target better efficiency in the feature extraction part using a dataflow description of the algorithm and using pipelining. These enhancement ensure a real-time implementation in the embedded Jetson TX2 board, previously unachievable.

⚠️ At the moment, only the monocular version of the KITTI dataset is supported.

Note: The library is intended to be built & run on NVIDIA Jetson TX2 with JetPack 4.2.2, but it should works fine with the newer versions as well.

Stefano Aldegheri, Nicola Bombieri, Daniele D. Bloisi and Alessandro Farinelli. Data Flow ORB-SLAM for Real-time Performance on Embedded GPU Boards. IEEE/RSJ International Conference on Intelligent Robots and Systems. PDF.

NVIDIA JetPack is a comprehensive SDK for Jetson for both developing and deploying AI and computer vision applications. JetPack simplifies installation of the OS and drivers and contains the following components:

• L4T Kernel / BSP
• CUDA Toolkit
• cuDNN
• TensorRT
• OpenCV
• VisionWorks
• Multimedia API's

Jetson TX2 should be flashed by downloading the NVIDIA SDK Manager to a host PC running Ubuntu 16.04 x86_64 or Ubuntu 18.04 x86_64. For more details, please refer to the NVIDIA SDK Manager Documentation.

Pangolin is used for visualization and user interface. Download and install instructions can be found at: https://github.com/stevenlovegrove/Pangolin.

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.

To download the code, navigate to a folder of your choosing on the Jetson (we take as reference the $HOME folder). First, make sure git and cmake are installed:

$ sudo apt-get update
$ sudo apt-get install git cmake

Then clone the dataflow-orbslam project:

$ git clone https://github.com/xaldyz/dataflow-orbslam.git

Build DBoW and g2o modified libraries in Thirdparty folder

$ cd ~/dataflow-orbslam/Thirdparty/DBoW2	# Build DBoW
$ mkdir build && cd build
$ cmake ..
$ make
$ cd ~/dataflow-orbslam/Thirdparty/g2o		# Build g2o
$ mkdir build && cd build
$ cmake ..
$ make

And finally you can build dataflow-orbslam

$ cd ~/dataflow-orbslam
$ mkdir build && cd build
$ cmake ..
$ make

⚠️ The project is set to build only the Monocular mono_kitti example by default.

Note: In the CMakeLists.txt file of the project folder you can set up CUSTOM_VX and PIPELINE variables to switch ON or OFF these optimisations.

First of all you need to download some example sequences from http://www.cvlibs.net/datasets/kitti/eval_odometry.php. We tested the application on the sequences 03, 04, 05 and 06 from grayscale odometry dataset.

Then uncompress the Vocabulary:

$ cd ~/dataflow-orbslam/Vocabulary
$ tar -zxvf ORBvoc.txt.tar.gz

From the project folder go into the build subfolder and run the follow command:

./mono_kitti ../Vocabulary/ORBvoc.txt ../Examples/Monocular/KITTI03.yaml PATH_TO_SEQUENCE_FOLDER

If all works fine you should see the execution below:

⚠️ Edit parameters on the .yaml file

Note: Only the KITTI04-12.yaml in the Examples subfolder is expected to run. This is because there are two parameters that must be set to

• Camera.width: the width of the image (different KITTI streams has different widths)
• Camera.height: the height of the image (different KITTI streams has different heights)