monovslam

Visual simultaneous localization and mapping (vSLAM) and visual-inertial sensor fusion with monocular camera

Since R2025a

Description

Use the monovslam object to perform visual simultaneous localization and mapping (vSLAM) and visual inertial SLAM (viSLAM) with a monocular camera. The object extracts Oriented FAST and Rotated BRIEF (ORB) features from incrementally read images, and then tracks those features to estimate camera poses, identify key frames, and reconstruct a 3-D environment. The vSLAM algorithm also searches for loop closures using the bag-of-features algorithm, and then optimizes the camera poses using pose graph optimization. You can enhance the accuracy and robustness of the SLAM by integrating this object with IMU data to perform visual-inertial sensor fusion. To learn more about visual SLAM, see Implement Visual SLAM in MATLAB (Computer Vision Toolbox).

Creation

Syntax

vslam = monovslam(intrinsics)

vslam = monovslam(intrinsics,imuParameters)

vslam = monovslam(intrinsics,PropertyName=Value)

Description

vslam = monovslam(intrinsics) creates a monocular visual SLAM object, vslam, by using the camera intrinsic parameters.

The monovslam object does not account for lens distortion. You can use the specified undistortImage (Computer Vision Toolbox) function to undistort images before adding them to the object.

The object represents 3-D map points and camera poses in world coordinates. The object assumes the camera pose of the first key frame is an identity rigidtform3d (Image Processing Toolbox) transform.

Note

The monovslam object runs on multiple threads internally, which can delay the processing of an image frame added by using the addFrame function. Additionally, the object running on multiple threads means the current frame the object is processing can be different than the recently added frame.

vslam = monovslam(intrinsics,imuParameters) creates a monocular visual SLAM object, vslam, by fusing the camera input intrinsics with inertial measurement unit (IMU) readings.

vslam = monovslam(intrinsics,PropertyName=Value) sets properties using one or more name-value arguments. For example, MaxNumFeatures=850 sets the maximum number of ORB feature points to extract from each image to 850.

Input Arguments

intrinsics — Camera intrinsic parameters

cameraIntrinsics object

Camera intrinsic parameters, specified as a cameraIntrinsics (Computer Vision Toolbox) object.

This argument sets the Intrinsics property.

imuParameters — IMU parameters

factorIMUParameters object

IMU parameters, specified as a factorIMUParameters object. The object contains the noise, bias, and sample rate information about the inertial measurement unit (IMU).

Properties

Camera Parameters

Intrinsics — Camera intrinsic parameters

Read-only: cameraIntrinsics object

This property is read-only.

Camera intrinsic parameters, stored as a cameraIntrinsics (Computer Vision Toolbox) object.

Use the intrinsics argument to set this property.

Feature Extraction

ScaleFactor — Scale factor for image decomposition

1.2 (default) | scalar greater than 1

This property is read-only after object creation.

Scale factor for image decomposition, stored as a scalar greater than 1. The scale factor is also referred to as the pyramid decimation ratio. Increasing the value of ScaleFactor reduces the number of pyramid levels, but reduce computation time. Decreasing this value (down to just over 1), increases the number of pyramid levels, which can improve tracking performance, at the cost of computation speed. The scale value at each level of decomposition is ScaleFactor^(level-1), where level is any value in the range [0, Numlevels-1]. Given the input image of size M-by-N, the image size at each level of composition is M_level-by-N_level, where:

$$\begin{array}{l}{M}_{level}=\frac{M}{ScaleFacto{r}^{(level-1)}}\\ N_{level}=\frac{N}{ScaleFacto{r}^{(level-1)}}\end{array}$$

NumLevels — Number of decomposition levels

8 (default) | integer greater than or equal to 1

This property is read-only after object creation.

Number of decomposition levels, specified as an integer greater than or equal to 1. Increase this value to extract keypoints from the image at more levels of decomposition. Along with the ScaleFactor value, NumLevels controls the number of pyramid levels on which the object evaluates feature points.

The image size at each decomposition level limits the number of levels at which you can extract keypoints. The image size at a level of decomposition must be at least 63-by-63 for keypoint detection. The maximum level of decomposition is calculated as

$$leve{l}_{\max }\text{ = floor}\left(\frac{\log \left(\min \left(M,N\right)\right)-\log \left(63\right)}{\log \left(\text{ScaleFactor}\right)}\right)+1$$

If either the default value or the specified value of NumLevels is greater than level_max, the object modifies NumLevels to level_max and returns a warning.

MaxNumPoints — Maximum number of ORB feature points uniformly extracted from each image

1000 | positive integer

This property is read-only after object creation.

Maximum number of ORB feature points uniformly extracted from each image, specified as a positive integer. Values are typically in the range of [800, 2000], depending on the resolution of the image. When the number of extracted features is less than the value of MaxNumPoints, then the object uses all feature points.

Tracking

TrackFeatureRange — Range for the number of feature points to identify key frame

[30 100] | two-element vector of positive integers

This property is read-only after object creation.

Key frame feature point range, stored as a two-element vector of positive integers in the form [lowerLimit upperLimit]. This property specifies the minimum and maximum numbers of tracked features points a frame must contain for the object to identify it as a key frame. The TrackFeatureRange and the SkipMaxFrames properties enable you to control the frequency at which frames in the tracking process become key frames.

The success of tracking depends on the number of tracked points in the current frame, with one of these results:

• Tracking is lost — The number of tracked feature points in the current frame is less than the lowerLimit set by the TrackFeatureRange argument. This indicates that the image does not contain enough features, or that the camera is moving too fast. To improve the tracking, you can increase the upperLimit value of the TrackFeatureRange property and decrease the SkipMaxFrames property to add key frames more frequently.

• Tracking is successful — The object identifies the current frame as a key frame. The number of tracked feature points in the current frame is in the range set by TrackFeatureRange.

• Tracking adds key frames too frequently — The number of tracked feature points in the current frame is greater than the upperLimit set by the TrackFeatureRange property. This indicates that the camera is moving very slowly, which produces an unnecessary number of key frames. To improve the tracking, you can reduce the frequency of adding key frames by increasing the value of the SkipMaxFrames property.

For more details, see the addFrame object function.

SkipMaxFrames — Maximum number of skipped frames

20 | positive integer

This property is read-only after object creation.

Maximum number of skipped frames, stored as a positive integer. When the number of tracked features is consistently greater than the upperLimit set by the TrackFeatureRange property, use the SkipMaxFrames property to control the frequency at which the object adds new key frames. The object identifies the current frame as a key frame when the number of skipped frames since the most recently added key frame equals the value of SkipMaxFrames.

LoopClosureThreshold — Minimum number of matched feature points between loop closure key frames

60 | positive integer

This property is read-only after object creation.

Minimum number of matched feature points between loop closure key frames, stored as a positive integer.

CustomBagOfFeatures — Custom bag of features for loop detection

[] (default) | bagOfFeaturesDBoW object

This property is read-only after object creation.

Custom bag of features for loop detection, specified as a bagOfFeaturesDBoW (Computer Vision Toolbox) object. The bagOfFeaturesDBoW (Computer Vision Toolbox) enables you to create a custom bag of words (BoW) from feature descriptors, alongside options to utilize a built-in vocabulary or load a custom one from a specified file.

Verbose — Progress information display

[] (default) | 1 | 2 | 3

This property is read-only after object creation.

Progress information display, specified as [], 1, 2, or 3. Paths to the location of log files, when they are created, is displayed on the command line.

Verbose value	Display description	Display location
[] or false	Display is turned off	—
1 or true	Stages of vSLAM execution	Command window
2	Stages of vSLAM execution, with details on how the frame is processed, such as the artifacts used to initialize the map.	Log file in a temporary folder
3	Stages of vSLAM, artifacts used to initialize the map, poses and map points before and after bundle adjustment, and loop closure optimization data.	Log file in a temporary folder

IMU Fusion

CameraToIMUTransform — IMU sensor transformation

rigidtform3d

This property is read-only after object creation.

IMU sensor transformation, specified as a rigidtform3d (Image Processing Toolbox) object. The transformation describes the rotation and translation of the camera in the coordinate system of the IMU sensor.

NumPosesThreshold — Number of estimated camera poses to trigger camera-IMU alignment

30 (default) | integer equal to or greater than 2

This property is read-only after object creation.

Number of estimated camera poses to trigger camera-IMU alignment, specified as an integer equal to or greater than 2. The process of aligning camera and IMU data is initiated after a specific number of camera poses have been estimated. This alignment serves two primary purposes: first, to estimate a scale factor that translates the up-to-scale results from a monocular camera into actual world units (meters), and second, to synchronize the IMU and camera frames, effectively eliminating the influence of gravity on accelerometer data. The timing for this alignment, determined by a threshold for the number of camera poses, is key to its success. A threshold set too low may not provide enough data for accurate calibration, while one set too high risks incorporating noise from measurement drift into the calibration. For a more in-depth understanding of this calibration technique, see the estimateGravityRotationAndPoseScale function.

AlignmentFraction — Subset of pose estimates

0.8 (default) | scalar in the range (0,1]

This property is read-only after object creation.

Subset of pose estimates, specified as a scalar in the range of (0,1]. This value specifies a fraction of the number of recent pose estimates, calculated as round(NumPosesThreshold*AlignmentFraction), for use in the camera-IMU alignment process. It effectively filters out initial, potentially noisy pose estimates, ensuring only the most relevant data contributes to the alignment for improved accuracy.

Object Functions

addFrame	Add image frame to visual SLAM object
hasNewKeyFrame	Check if new key frame added in visual SLAM object
checkStatus	Check status of visual SLAM object
isDone	End-of-processing status for visual SLAM object
mapPoints	Build 3-D map of world points
poses	Absolute camera poses of key frames
plot	Plot 3-D map points and estimated camera trajectory in visual SLAM
reset	Reset visual SLAM object

Examples

Monocular Visual SLAM Using TUM RGB-D Data Set

Perform monocular visual simultaneous localization and mapping (vSLAM) using the data from the TUM RGB-D Benchmark. You can download the data to a temporary directory using a web browser or by running this code:

baseDownloadURL = "https://cvg.cit.tum.de/rgbd/dataset/freiburg3/rgbd_dataset_freiburg3_long_office_household.tgz"; 
dataFolder = fullfile(tempdir,"tum_rgbd_dataset",filesep); 
options = weboptions(Timeout=Inf);
tgzFileName = dataFolder+"fr3_office.tgz";
folderExists = exist(dataFolder,"dir");

% Create a folder in a temporary directory to save the downloaded file
if ~folderExists  
    mkdir(dataFolder) 
    disp("Downloading fr3_office.tgz (1.38 GB). This download can take a few minutes.") 
    websave(tgzFileName,baseDownloadURL,options); 
    
    % Extract contents of the downloaded file
    disp("Extracting fr3_office.tgz (1.38 GB) ...") 
    untar(tgzFileName,dataFolder); 
end

Create an imageDatastore object to store all the RGB images.

imageFolder = dataFolder+"rgbd_dataset_freiburg3_long_office_household/rgb/";
imds = imageDatastore(imageFolder);

Specify your camera intrinsic parameters, and use them to create a monocular visual SLAM object.

intrinsics = cameraIntrinsics([535.4 539.2],[320.1 247.6],[480 640]);
vslam = monovslam(intrinsics,TrackFeatureRange=[30,120]);

Process each image frame, and visualize the camera poses and 3-D map points. Note that the monovslam object runs several algorithm parts on separate threads, which can introduce a latency in processing of an image frame added by using the addFrame function.

for i = 1:numel(imds.Files)
    addFrame(vslam,readimage(imds,i))

    if hasNewKeyFrame(vslam)
        % Display 3-D map points and camera trajectory
        plot(vslam);
    end

    % Get current status of system
    status = checkStatus(vslam);
end 

Plot intermediate results and wait until all images are processed.

while ~isDone(vslam)
    if hasNewKeyFrame(vslam)
        plot(vslam);
    end
end

After all the images are processed, you can collect the final 3-D map points and camera poses for further analysis.

xyzPoints = mapPoints(vslam);
[camPoses,addedFramesIdx] = poses(vslam);

% Reset the system
reset(vslam)

Compare the estimated camera trajectory with the ground truth to evaluate the accuracy.

% Load ground truth
gTruthData = load("orbslamGroundTruth.mat");
gTruth     = gTruthData.gTruth;

% Evaluate tracking accuracy
mtrics = compareTrajectories(camPoses, gTruth(addedFramesIdx), AlignmentType="similarity");
disp(['Absolute RMSE for key frame location (m): ', num2str(mtrics.AbsoluteRMSE(2))]);

Absolute RMSE for key frame location (m): 0.20043

% Plot the absolute translation error at each key frame
figure
ax = plot(mtrics, "absolute-translation");
view(ax, [2.70 -49.20]); 

Monocular Visual-Inertial SLAM Using the BlackBird Data Set

Perform monocular visual-inertial SLAM using the data from the Blackbird data set (NYC Subway Winter). Download the MAT file which contains the images, camera intrinsics and IMU measurements.

uavDataTarFile =  matlab.internal.examples.downloadSupportFile('shared_nav_vision/data','BlackbirdVIOData.tar');  

% Extract the file.
outputFolder = fileparts(uavDataTarFile);
if (~exist(fullfile(outputFolder,"BlackbirdVIOData"),"dir"))
    untar(uavDataTarFile,outputFolder);
end

uavData = load(fullfile(outputFolder,"BlackbirdVIOData","data.mat"));

images     = uavData.images;
timeStamps = uavData.timeStamps;

Set up the IMU noise parameters, camera intrinsics and the camera-IMU extrinsics transform.

imuParams = factorIMUParameters(SampleRate=100,GyroscopeNoise=0.1, ...
            GyroscopeBiasNoise=3e-3,AccelerometerNoise=0.3, ...
            AccelerometerBiasNoise=1e-3,ReferenceFrame="ENU");
intrinsics = uavData.intrinsics;
camera2IMU = rigidtform3d(tform(uavData.camToIMUTransform));

Set up the monocular visual-inertial SLAM object.

vslam = monovslam(intrinsics,imuParams,MaxNumPoints=2000,SkipMaxFrames=2,TrackFeatureRange=[20,100], ...
        CameraToIMUTransform=camera2IMU,NumPosesThreshold=40,AlignmentFraction=0.9, ...
        CustomBagOfFeatures=bagOfFeaturesDBoW('BlackBirdBoW.bin.gz'),LoopClosureThreshold=100);

Process all the images frames and IMU measurements available, and visualize the camera poses and 3-D map points. The sequence does not contain any movement before the 150th frame.

startFrameIdx = 150;

for i = startFrameIdx:length(images)

    if(i>startFrameIdx)
        imuMesurements = helperExtractIMUMeasurements(uavData, i-1, i);
    else
        imuMesurements.gyro=[];
        imuMesurements.accel=[];
    end

    addFrame(vslam, images{i}, imuMesurements.gyro, imuMesurements.accel);

    if hasNewKeyFrame(vslam)
        plot(vslam);
    end
end

Note that the monovslam object runs several algorithm parts on separate threads, which can introduce a latency in processing of an image frame added by using the addFrame function.

% Plot intermediate results and wait until all images are processed
while ~isDone(vslam)
    if hasNewKeyFrame(vslam)
        plot(vslam);
    end
end

The helper function below is used to extract the IMU measurements, based on the camera timestamps.

function imuMesurements = helperExtractIMUMeasurements(data, startFrameIdx, currFrameIdx)

    timeStamps = data.timeStamps;
    
    startTimeStamp = timeStamps.imageTimeStamps(startFrameIdx);
    currTimeStamp  = timeStamps.imageTimeStamps(currFrameIdx);
    
    [~,startIMUIdx] = min(abs(timeStamps.imuTimeStamps - startTimeStamp));
    [~,currIMUIdx]  = min(abs(timeStamps.imuTimeStamps - currTimeStamp));
    
    imuMesurements.accel = data.accelReadings(startIMUIdx:(currIMUIdx-1),:);
    imuMesurements.gyro  = data.gyroReadings(startIMUIdx:(currIMUIdx-1),:);

end

References

[1] Mur-Artal, Raul, J. M. M. Montiel, and Juan D. Tardos. “ORB-SLAM: A Versatile and Accurate Monocular SLAM System.” IEEE Transactions on Robotics 31, no. 5 (October 2015): 1147–63. https://doi.org/10.1109/TRO.2015.2463671.

[2] Galvez-López, D., and J. D. Tardos. “Bags of Binary Words for Fast Place Recognition in Image Sequences.” IEEE Transactions on Robotics, vol. 28, no. 5, Oct. 2012, pp. 1188–97. DOI.org (Crossref), https://doi.org/10.1109/TRO.2012.2197158.

Extended Capabilities

C/C++ Code Generation

Generate C and C++ code using MATLAB® Coder™.

To compile the generated code for non-host target, these dependencies must be installed:

• OpenCV 4.7.0

• Ceres-solver 2.0.0 library — Download the source and build it as a shared library, on your remote device.

• eigen3 Library

Version History

Introduced in R2025a

R2025a: MATLAB Coder support added for IMU sensor support

Added C/C++ support for visual-inertial sensor fusion using the monovslam (Computer Vision Toolbox) object. Starting in R2025a, you can generate C/C++ code using the IMU parameters along with all other object parameters.

R2024b: Fusion support for IMU sensors

Use the imuParameters input argument as a factorIMUParameters object to specify information about IMU parameters. Use the CameraToIMUTransform, NumPosesThreshold, and AlignmentFraction properties to tune IMU parameters. The IMU-based features of the monovslam object do not support C/C++ code generation.

R2024b: Add ability to specify custom bag of words

Use the CustomBagOfFeatures name-value argument to specify a custom bag of words vocabulary as a bagOfFeaturesDBoW (Computer Vision Toolbox) object.

R2024b: Verbose property enhanced to display progress information

The Verbose property added values of 1, 2, and 3 to display levels of progress information.