Project
Brief Visual-Inertial SLAM with GTSAM
Details
- Built a SLAM system from scratch using ROS, including an OpenCV frontend, and a GTSAM backend with loop-closing.
- Improved feature matching to $\sim{100}$ correct matches per frame in frontend with optical flow method and correspondence search.
- Evaluated with ground truth data in TartanAir dataset, 2% of matches are outliers.
- Used factor graphs and GTSAM to optimize and fuse factors like SIFT/ORB features, IMU, and objects.
- The backend achieves 3 times better ATE than the input odometry.