This is a repository containing my implementation of a SLAM algorithm, utilizing stereo images from the KITTI dataset. Each timestep contains a pair of grayscale stereo images.

The main structure of this repository is as follows:

• utils.py - This file contains the following classes:
  • Keyframe - This class contains information for each keyframe that is extracted, and utilized for detecting loop-closures, optimizing the predicted trajectory, and storing the pointcloud from each keyframe.
  • BoWDatabase - This class implements a Bag of Words database, which is trained on a set of keyframes at the start of any trajectory. This allows for efficient detection of matched keyframes for loop-closure detection.
  • SpatialIndex - This class implements a KD-tree based upon the 3D positions of each keyframe. This allows for efficient querying of existing keyframes near a 3D position.
• g2o_optimizer.py - This file implements a pose-graph optimizer from the g2o library, an open-source graph optimization library.
• loop_closure.py - This file implements a loop closure detector class, which combines a BoWDatabase and a SpatialIndex KD-tree to detect if new keyframes added to the trajectory are loop closures in the trajectory.
• slam.py - This file implements all functions necessary for the visual-odometry trajectory prediction, determining when to create a new keyframe, and building a global pointcloud.
• main.py - This file calls the functions necessary from slam.py to fully implement the SLAM algorithm for a given KITTI sequence, and save the visual-odometry trajectory, optimized trajectory, and global pointcloud created by this algorithm.