LSD-SLAM is a novel approach to real-time monocular SLAM. It is fully direct (i.e. does not use keypoints / features) and creates large-scale, semi-dense maps in real-time on a laptop. For more information see http://vision.in.tum.de/lsdslam where you can also find the corresponding publications and Youtube videos, as well as some example-input datasets, and the generated output as rosbag or .ply point cloud.
-
LSD-SLAM: Large-Scale Direct Monocular SLAM, J. Engel, T. Schöps, D. Cremers, ECCV '14
-
Semi-Dense Visual Odometry for a Monocular Camera, J. Engel, J. Sturm, D. Cremers, ICCV '13
First, install LSD-SLAM following 2.1 or 2.2, depending on your Ubuntu / ROS version. You don't need openFabMap for now.
Download the Room Example Sequence and extract it.
Launch the lsd_slam viewer:
rosrun lsd_slam_viewer viewer
Launch the lsd_slam main ros node:
rosrun lsd_slam_core live_slam image:=/image_raw camera_info:=/camera_info
Play the sequence:
rosbag play ~/LSD_room.bag
You should see one window showing the current keyframe with color-coded depth (from live_slam), and one window showing the 3D map (from viewer). If for some reason the initialization fails (i.e., after ~5s the depth map still looks wrong), focus the depth map and hit 'r' to re-initialize.
We tested LSD-SLAM on two different system configurations, using Ubuntu 12.04 (Precise) and ROS fuerte, or Ubuntu 14.04 (trusty) and ROS indigo. Note that building without ROS is not supported, however ROS is only used for input and output, facilitating easy portability to other platforms.
Install system dependencies:
sudo apt-get install ros-fuerte-libg2o liblapack-dev libblas-dev freeglut3-dev libqglviewer-qt4-dev libsuitesparse-dev libx11-dev
In your ROS package path, clone the repository:
git clone https://github.com/tum-vision/lsd_slam.git lsd_slam
Compile the two package by typing:
rosmake lsd_slam
We do not use catkin, however fortunately old-fashioned CMake-builds are still possible with ROS indigo. For this you need to create a rosbuild workspace (if you don't have one yet), using:
sudo apt-get install python-rosinstall
mkdir ~/rosbuild_ws
cd ~/rosbuild_ws
rosws init . /opt/ros/indigo
mkdir package_dir
rosws set ~/rosbuild_ws/package_dir -t .
echo "source ~/rosbuild_ws/setup.bash" >> ~/.bashrc
bash
cd package_dir
Install system dependencies:
sudo apt-get install ros-indigo-libg2o ros-indigo-cv-bridge liblapack-dev libblas-dev freeglut3-dev libqglviewer-dev libsuitesparse-dev libx11-dev
In your ROS package path, clone the repository:
git clone https://github.com/tum-vision/lsd_slam.git lsd_slam
Compile the two package by typing:
rosmake lsd_slam
If you want to use openFABMAP for large loop closure detection, uncomment the following lines in lsd_slam_core/CMakeLists.txt
:
#add_subdirectory(${PROJECT_SOURCE_DIR}/thirdparty/openFabMap)
#include_directories(${PROJECT_SOURCE_DIR}/thirdparty/openFabMap/include)
#add_definitions("-DHAVE_FABMAP")
#set(FABMAP_LIB openFABMAP )
Note for Ubuntu 14.04: The packaged OpenCV for Ubuntu 14.04 does not include the nonfree module, which is required for openFabMap (which requires SURF features). You need to get a full version of OpenCV with nonfree module, which is easiest by compiling your own version. We suggest to use the 2.4.8 version, to assure compatibility with the current indigo open-cv package.
An optional way to install OpenCV with nonfree modules without rebuilding from source:
sudo add-apt-repository ppa:xqms/opencv-nonfree
sudo apt-get update
sudo apt-get install libopencv-nonfree-dev
LSD-SLAM is split into two ROS packages, lsd_slam_core
and lsd_slam_viewer
. lsd_slam_core
contains the full SLAM system, whereas lsd_slam_viewer
is optionally used for 3D visualization.
Please also read General Notes for good results below.
We provide two different usage modes, one meant for live-operation (live_slam
) using ROS input/output, and one dataset_slam
to use on datasets in the form of image files.
If you want to directly use a camera.
rosrun lsd_slam_core live_slam /image:=<yourstreamtopic> /camera_info:=<yourcamera_infotopic>
When using ROS camera_info, only the image dimensions and the K
matrix from the camera info messages will be used - hence the video has to be rectified.
Alternatively, you can specify a calibration file using
rosrun lsd_slam_core live_slam /image:=<yourstreamtopic> _calib:=<calibration_file>
In this case, the camera_info topic is ignored, and images may also be radially distorted. See the Camera Calibration section for details on the calibration file format.
rosrun lsd_slam_core dataset_slam _files:=<files> _hz:=<hz> _calib:=<calibration_file>
Here, <files>
can either be a folder containing image files (which will be sorted alphabetically), or a text file containing one image file per line. <hz>
is the framerate at which the images are processed, and <calibration_file>
the camera calibration file.
Specify _hz:=0
to enable sequential tracking and mapping, i.e. make sure that every frame is mapped properly. Note that while this typically will give best results, it can be much slower than real-time operation.
LSD-SLAM operates on a pinhole camera model, however we give the option to undistort images before they are being used. You can find some sample calib files in lsd_slam_core/calib
.
fx/width fy/height cx/width cy/height d
in_width in_height
"crop" / "full" / "none" / "e1 e2 e3 e4 0"
out_width out_height
Here, the values in the first line are the camera intrinsics and radial distortion parameter as given by the PTAM cameracalibrator, in_width and in_height is the input image size, and out_width out_height is the desired undistorted image size. The latter can be chosen freely, however 640x480 is recommended as explained in section 3.1.6. The third line specifies how the image is distorted, either by specifying a desired camera matrix in the same format as the first four intrinsic parameters, or by specifying "crop", which crops the image to maximal size while including only valid image pixels.
This one is without radial distortion correction, as a special case of ATAN camera model but without the computational cost:
fx/width fy/height cx/width cy/height 0
width height
none
width height
fx fy cx cy k1 k2 p1 p2
inputWidth inputHeight
"crop" / "full" / "none" / "e1 e2 e3 e4 0"
outputWidth outputHeight
-
r
: Do a full reset -
d / e
: Cycle through debug displays (in particular color-coded variance and color-coded inverse depth). -
o
: Toggle on screen info display -
m
: Save current state of the map (depth & variance) as images tolsd_slam_core/save/
-
p
: Brute-Force-Try to find new constraints. May improve the map by finding more constraints, but will block mapping for a while. -
l
: Manually indicate that tracking is lost: will stop tracking and mapping, and start the re-localizer.
A number of things can be changed dynamically, using (for ROS fuerte)
rosrun dynamic_reconfigure reconfigure_gui
or (for ROS indigo)
rosrun rqt_reconfigure rqt_reconfigure
Parameters are split into two parts, ones that enable / disable various sorts of debug output in /LSD_SLAM/Debug
, and ones that affect the actual algorithm, in /LSD_SLAM
.
Note that debug output options from /LSD_SLAM/Debug
only work if lsd_slam_core is built with debug info, e.g. with set(ROS_BUILD_TYPE RelWithDebInfo)
.
minUseGrad
: [double] Minimal absolute image gradient for a pixel to be used at all. Increase if your camera has large image noise, decrease if you have low image-noise and want to also exploit small gradients.cameraPixelNoise
: [double] Image intensity noise used for e.g. tracking weight calculation. Should be set larger than the actual sensor-noise, to also account for noise originating from discretization / linear interpolation.KFUsageWeight
: [double] Determines how often keyframes are taken, depending on the overlap to the current keyframe. Larger -> more keyframes.KFDistWeight
: [double] Determines how often keyframes are taken, depending on the distance to the current Keyframe. Larger -> more keyframes.doSLAM
: [bool] Toggle global mapping component on/off. Only takes effect after a reset.doKFReActivation
: [bool] Toggle keyframe re-activation on/off: If close to an existing keyframe, re-activate it instead of creating a new one. If false, the map will continually grow even if the camera moves in a relatively constrained area; If false, the number of keyframes will not grow arbitrarily.doMapping
: [bool] Toggle entire keyframe creating / update module on/off: If false, only tracking stays active, which will prevent rapid motion or moving objects from corrupting the map.useFabMap
: [bool] Use openFABMAP to find large loop-closures. Only takes effect after a reset, and requires LSD-SLAM to be compiled with FabMap.allowNegativeIdepths
: [bool] Allow idepth to be (slightly) negative to avoid introducing a bias for far-away points.useSubpixelStereo
: [bool] Compute subpixel-accurate stereo disparity.useAffineLightningEstimation
: [bool] EXPERIMENTAL: Correct for global affine intensity changes during tracking. Might help if you have problems with auto-exposure.multiThreading
: [bool] Toggle multi-threading of depth map estimation. Disable for less CPU usage, but possibly slightly less quality.maxLoopClosureCandidates
: [int] Maximal number of loop-closures that are tracked initially for each new keyframe.loopclosureStrictness
: [double] Threshold on reciprocal loop-closure consistency check, to be added to the map. Larger -> more (possibly wrong) loop-closures.relocalizationTH
: [double] How good a relocalization-attempt has to be to be accepted. Larger -> more strict.depthSmoothingFactor
: [double] How much to smooth the depth map. Larger -> less smoothing.
Useful for debug output are:
plotStereoImages
: [bool] Plot searched stereo lines, and color-coded stereo-results. Nice visualization of what's going on, however drastically decreases mapping speed.plotTracking
: [bool] Plot final tracking residual. Nice visualization of what's going on, however drastically decreases tracking speed.continuousPCOutput
: [bool] Publish current keyframe's point cloud after each update, to be seen in the viewer. Nice visualization, however bad for performance and bandwidth.
- Use a global shutter camera. Using a rolling shutter will lead to inferior results.
- Use a lens with a wide field-of-view (we use a 130° fisheye lens).
- Use a high framerate, at least 30fps (depending on the movements speed of course). For our experiments, we used between 30 and 60 fps.
- We recommend an image resolution of 640x480, significantly higher or lower resolutions may require some hard-coded parameters to be adapted.
- LSD-SLAM is a monocular SLAM system, and as such cannot estimate the absolute scale of the map. Further it requires sufficient camera translation: Rotating the camera without translating it at the same time will not work. Generally sideways motion is best - depending on the field of view of your camera, forwards / backwards motion is equally good. Rotation around the optical axis does not cause any problems.
- During initialization, it is best to move the camera in a circle parallel to the image without rotating it. The scene should contain sufficient structure (intensity gradient at different depths).
- Adjust
minUseGrad
andcameraPixelNoise
to fit the sensor-noise and intensity contrast of your camera. - If tracking / mapping quality is poor, try decreasing the keyframe thresholds
KFUsageWeight
andKFDistWeight
slightly to generate more keyframes. - Note that LSD-SLAM is very much non-deterministic, i.e. results will be different each time you run it on the same dataset. This is due to parallelism, and the fact that small changes regarding when keyframes are taken will have a huge impact on everything that follows afterwards.
The viewer is only for visualization. It can also be used to output a generated point cloud as .ply. For live operation, start it using
rosrun lsd_slam_viewer viewer
You can use rosbag to record and re-play the output generated by certain trajectories. Record & playback using
rosbag record /lsd_slam/graph /lsd_slam/keyframes /lsd_slam/liveframes -o file_pc.bag
rosbag play file_pc.bag
You should never have to restart the viewer node, it resets the graph automatically.
If you just want to lead a certain pointcloud from a .bag file into the viewer, you can directly do that using
rosrun lsd_slam_viewer viewer file_pc.bag
-
r
: Reset, will clear all displayed data. -
w
: Print the number of points / currently displayed points / keyframes / constraints to the console. -
p
: Write currently displayed points as point cloud to file lsd_slam_viewer/pc.ply, which can be opened e.g. in meshlab. Use in combination with sparsityFactor to reduce the number of points.
showKFCameras
: Toggle drawing of blue keyframe camera-frustrums. min: False, default: True, max: TrueshowKFPointclouds
: Toggle drawing of point clouds for all keyframes. min: False, default: True, max: TrueshowConstraints
: Toggle drawing of red/green pose-graph constraints. min: False, default: True, max: TrueshowCurrentCamera
: Toggle drawing of red frustrum for the current camera pose. min: False, default: True, max: TrueshowCurrentPointcloud
: Toggle drawing of the latest point cloud added to the map. min: False, default: True, max: TruepointTesselation
: Size of points. min: 0.0, default: 1.0, max: 5.0lineTesselation
: Width of lines. min: 0.0, default: 1.0, max: 5.0scaledDepthVarTH
: log10 of threshold on point's variance, in the respective keyframe's scale. min: -10.0, default: -3.0, max: 1.0absDepthVarTH
: log10 of threshold on point's variance, in absolute scale. min: -10.0, default: -1.0, max: 1.0minNearSupport
: Only plot points that have #minNearSupport similar neighbours (higher values remove outliers). min: 0, default: 7, max: 9cutFirstNKf
: Do not display the first #cutFirstNKf keyframe's point clouds, to remove artifacts left-over from the random initialization. min: 0, default: 5, max: 100sparsifyFactor
: Only plot one out of #sparsifyFactor points, selected at random. Use this to significantly speed up rendering for large maps. min: 1, default: 1, max: 100sceneRadius
: Defines near- and far clipping plane. Decrease to be able to zoom in more. min: 1, default: 80, max: 200saveAllVideo
: Save all rendered images... only use if you know what you are doing. min: False, default: False, max: TruekeepInMemory
: If set to false, the point cloud is only stored in OpenGL buffers, and not kept in RAM. This greatly reduces the required RAM for large maps, however also prohibits saving / dynamically changing sparsifyFactor and variance-thresholds. min: False, default: True, max: True
For convenience we provide a number of datasets, including the video, lsd-slam's output and the generated point cloud as .ply. See http://vision.in.tum.de/lsdslam
LSD-SLAM is licensed under the GNU General Public License Version 3 (GPLv3), see http://www.gnu.org/licenses/gpl.html.
For commercial purposes, we also offer a professional version under different licencing terms.
How can I get the live-pointcloud in ROS to use with RVIZ?
You cannot, at least not on-line and in real-time. The reason is the following:
In the background, LSD-SLAM continuously optimizes the pose-graph, i.e., the poses of all keyframes. Each time a keyframe's pose changes (which happens all the time, if only by a little bit), all points from this keyframe change their 3D position with it. Hence, you would have to continuously re-publish and re-compute the whole pointcloud (at 100k points per keyframe and up to 1000 keyframes for the longer sequences, that's 100 million points, i.e., ~1.6GB), which would crush real-time performance.
Instead, this is solved in LSD-SLAM by publishing keyframes and their poses separately:
- keyframeGraphMsg contains the updated pose of each keyframe, nothing else.
- keyframeMsg contains one frame with it's pose, and - if it is a keyframe - it's points in the form of a depth map.
Points are then always kept in their keyframe's coodinate system: That way, a keyframe's pose can be changed without even touching the points. In fact, in the viewer, the points in the keyframe's coodinate frame are moved to a GLBuffer immediately and never touched again - the only thing that changes is the pushed modelViewMatrix before rendering.
Note that "pose" always refers to a Sim3 pose (7DoF, including scale) - which ROS doesn't even have a message type for.
If you need some other way in which the map is published (e.g. publish the whole pointcloud as ROS standard message as a service), the easiest is to implement your own Output3DWrapper.
Tracking immediately diverges / I keep getting "TRACKING LOST for frame 34 (0.00% good Points, which is -nan% of available points, DIVERGED)!"
- double-check your camera calibration.
- try more translational movement and less roational movement