This is an implementation of a visual SLAM system performed on RGB-D images which consists of keyframe-based camera tracking through direct image alignment with relative entropy keyframe selection strategy, pose graph optimization and automatic loop closure detection.
Download TUM RGB-D dataset fr2/desk and save them under the directory rgbd_dataset_freiburg2_desk/
The evaluation tool scripts evaluate_ate.py and evaluate_rpe.py downloaded from TUM RGB-D benchmark are used to calculate the absolute trajectory error (ATE) and relative pose error (RPE) and plot the resulting trajectory.
Notice: Before running pose graph optimization, you have to run camera tracking first on a specific image scale determined by argument level to get the keyframe map. And before running PGO with automatic loop closure, you have to run automatic loop closure detection to generate loop closures file first.
##Run different algorithm with arguments
camera tracking with direct image alignment with keyframe selection strategy 1
python main.py --tracking 1 --keyframe_selection 0
camera tracking with direct image alignment with keyframe selection strategy 2
python main.py --tracking 1 --keyframe_selection 1
pose graph optimization with manually loop closure
python main.py --pgo 1 --loop_closure 0
automatic loop closure detection
python loop_closure.py
pose graph optimization with automatic loop closure
python main.py --pgo 1 --loop_closure 1
output the optimized trajectory text file your_trajectory.txt
python output_trajectory.py
plot the resulting trajectory
python evaluate_ate.py rgbd_dataset_freiburg2_desk/groundtruth.txt your_trajectory.txt --plot PLOT
