FastTrack is an optimized tracking for ORB-SLAM3 that leverages GPU computing power to accelerate the time-consuming components of tracking and enhance the overall performance. These components include ORB extraction, stereo feature matching, and local map tracking. We implemented CUDA kernels to speed up these components. Our results demonstrate a significant reduction in tracking times, achieving up to 2.8× faster performance on desktop and up to 2.7× speedup on Xavier NX. We evaluated FastTrack on a mix of sequences from the EuRoC and TUM-VI datasets using the stereo-inertial configuration.
🚀 FastTrack has been accepted for publication at IROS 2025.
Figure 1: Tracking In FastTrack
Table 1: Machine Specifications
| Machine | Specs |
|---|---|
| Desktop |
20-core Intel Core i7-12700K CPU @ 5.0 GHz NVIDIA RTX 3090 GPU (10496-core) 64 GB RAM |
| Xavier NX |
6-core ARM Carmel CPU @ 1.4 GHz NVIDIA Volta GPU (384-core) 8 GB RAM |
We have tested the library in Ubuntu 22.04 and 20.04, but it should be easy to compile in other platforms. A powerful computer (e.g. i7) will ensure real-time performance and provide more stable and accurate results.
We use the new thread and chrono functionalities of C++14.
We have tested the library with Cuda 12.2. Download and install instructions can be found at: https://docs.nvidia.com/cuda/cuda-installation-guide-linux/.
We use OpenCV to manipulate images and features. Dowload and install instructions can be found at: http://opencv.org. Required at leat 4.4.0.
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 Pangolin for visualization and user interface. Download and install instructions can be found at: https://github.com/stevenlovegrove/Pangolin.
Pangolin avoids to use Eigen in CUDA. To compile this project, guards in line 475 of glsl.hpp and 47 of glsl.h should be commented. Checkout this issue in ORB-SLAM3 github issues.
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.
Required to calculate the alignment of the trajectory with the ground truth. Required Numpy module.
- (win) http://www.python.org/downloads/windows
- (deb)
sudo apt install libpython2.7-dev - (mac) preinstalled with osx
Clone the repository:
git clone git@github.com:sfu-rsl/FastTrack.git
Our system is based on ORB-SLAM3 and ORB-SLAM3 provides a script build.sh to build the Thirdparty libraries and ORB-SLAM3. Please make sure you have installed all required dependencies (see section 2). Execute:
cd FastTrack
chmod +x build.sh
sudo ./build.sh
This will create libORB_SLAM3.so at lib folder and the executables in Examples folder.
EuRoC dataset was recorded with two pinhole cameras and an inertial sensor. We provide an example script to launch EuRoC sequences in all the sensor configurations.
-
Download a sequence (ASL format) from http://projects.asl.ethz.ch/datasets/doku.php?id=kmavvisualinertialdatasets
-
Open the script "euroc_eval_examples.sh" in the Examples directory of the project. Change pathDatasetEuroc variable to point to the directory where the dataset has been uncompressed.
-
Execute the following script to process all the sequences with all sensor configurations:
./run_script $dataset_name 1
TUM-VI dataset was recorded with two fisheye cameras and an inertial sensor.
-
Download a sequence from https://vision.in.tum.de/data/datasets/visual-inertial-dataset and uncompress it.
-
Open the script "tum_vi_examples.sh" in the root of the project. Change pathDatasetTUM_VI variable to point to the directory where the dataset has been uncompressed.
-
Execute the following script to process all the sequences with all sensor configurations:
./run_script $dataset_name 1