Example Stereo Visual Odometry

From BoofCV

Stereo visual odometry estimates the camera's egomotion using a pair of calibrated cameras. Stereo camera systems are inherently more stable than monocular ones because the stereo pair provides good triangulation of image features and resolves the scale ambiguity. The example below shows how to use a high level interface with visual odometry algorithms. The basic high level interface hides internal data structures which are useful in many applications, which is why an optional interface is provided for accessing some of those structures.

Example File: ExampleVisualOdometryStereo.java


  • Structure from Motion
  • Geometric Vision
  • Feature Tracking

Relevant Videos:

Related Tutorials/Example Code:

Example Code

 * Bare bones example showing how to estimate the camera's ego-motion using a stereo camera system. Additional
 * information on the scene can be optionally extracted from the algorithm if it implements AccessPointTracks3D.
 * @author Peter Abeles
public class ExampleVisualOdometryStereo {

	public static void main( String args[] ) {

		MediaManager media = DefaultMediaManager.INSTANCE;

		String directory = UtilIO.pathExample("vo/backyard/");

		// load camera description and the video sequence
		StereoParameters stereoParam = CalibrationIO.load(media.openFile(new File(directory , "stereo.yaml").getPath()));
		SimpleImageSequence<GrayU8> video1 = media.openVideo(
				new File(directory , "left.mjpeg").getPath(), ImageType.single(GrayU8.class));
		SimpleImageSequence<GrayU8> video2 = media.openVideo(
				new File(directory,"right.mjpeg").getPath(), ImageType.single(GrayU8.class));

		// specify how the image features are going to be tracked
		PkltConfig configKlt = new PkltConfig();
		configKlt.pyramidScaling = new int[]{1, 2, 4, 8};
		configKlt.templateRadius = 3;

		PointTrackerTwoPass<GrayU8> tracker =
				FactoryPointTrackerTwoPass.klt(configKlt, new ConfigGeneralDetector(600, 3, 1),
						GrayU8.class, GrayS16.class);

		// computes the depth of each point
		StereoDisparitySparse<GrayU8> disparity =
				FactoryStereoDisparity.regionSparseWta(0, 150, 3, 3, 30, -1, true, GrayU8.class);

		// declares the algorithm
		StereoVisualOdometry<GrayU8> visualOdometry = FactoryVisualOdometry.stereoDepth(1.5,120, 2,200,50,true,
				disparity, tracker, GrayU8.class);

		// Pass in intrinsic/extrinsic calibration.  This can be changed in the future.

		// Process the video sequence and output the location plus number of inliers
		while( video1.hasNext() ) {
			GrayU8 left = video1.next();
			GrayU8 right = video2.next();

			if( !visualOdometry.process(left,right) ) {
				throw new RuntimeException("VO Failed!");

			Se3_F64 leftToWorld = visualOdometry.getCameraToWorld();
			Vector3D_F64 T = leftToWorld.getT();

			System.out.printf("Location %8.2f %8.2f %8.2f      inliers %s\n", T.x, T.y, T.z, inlierPercent(visualOdometry));

	 * If the algorithm implements AccessPointTracks3D, then count the number of inlier features
	 * and return a string.
	public static String inlierPercent(StereoVisualOdometry alg) {
		if( !(alg instanceof AccessPointTracks3D))
			return "";

		AccessPointTracks3D access = (AccessPointTracks3D)alg;

		int count = 0;
		int N = access.getAllTracks().size();
		for( int i = 0; i < N; i++ ) {
			if( access.isInlier(i) )

		return String.format("%%%5.3f", 100.0 * count / N);