Harness GPU acceleration for advanced visual odometry and IMU data fusion with our Unscented Kalman Filter (UKF) implementation. Developed with C++ and powered by CUDA, cuBLAS, and cuSOLVER, the system delivers unmatched real-time performance in state and covariance estimation for robotics applications. Integrated with ROS 2 for seamless sensor data management, it ensures high-efficiency and scalable solutions for complex robotic systems, making it ideal for a wide range of autonomous system applications.
- NVIDIA Driver (Official Download Link)
- CUDA Toolkit (Official Link)
- ROS 2 Humble (Official Link)
- Clone
https://github.com/jagennath-hari/CUDA-Accelerated-Visual-Inertial-Odometry-Fusion.git
- Move
cuUKF
intoROS2_WORKSPACE
- Modify the
CMakeLists.txt
file atset(CMAKE_CUDA_ARCHITECTURES 89)
and change to your GPU architecture. If you don't know which one you can refer to NVIDIA GPU Compute Capability. cd ROS2_WORKSPACE
build workspace usingcolcon build --symlink-install --cmake-args=-DCMAKE_BUILD_TYPE=Release --parallel-workers $(nproc)
Launch the node using ros2 launch cuUKF gpu_filter.launch.py odom_topic:=/odom imu_topic:=/imu
.
The /odom
should be replaced with your nav_msgs/Odometry
and the /imu
should be replaced with your sensor_msgs/Imu
.
The filter odometry gets published as nav_msgs/Odometry
in the /cuUKF/filtered_odom
topic.
Launch RVIZ2 using ros2 run rviz2 rviz2
and subcribe to the /cuUKF/filered_odom
topic.
- The fusion does not consider the IMU's orientation only the visual odometry's orientation for the system dynamics and measurements, as raw IMU don't produce orientation without additional filters such as complementary filter and the Madgwick filter.
- Feel free to change the alpha, beta and kappa values along with the covariance to improve state estimation.
- The dynamics of the system use simple equations, for the best fusion you may need to change the dynamics.
- Consider adding augmented sigma points to further increase the robustness.
- There is also a CPU version of the UKF for quick development and testing.