Example Stereo Visual Odometry

• 

  Screenshot from visual odometry demonstration. Example just outputs estimated location.

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

Concepts:

• Structure from Motion
• Geometric Vision
• Feature Tracking

Relevant Videos:

• Youtube

Related Tutorials/Example Code:

• Tutorial Geometric Vision
• Visual Odometry Monocular Plane
• Visual Odometry Depth

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 which tracker and how it will behave
		var configKlt = new ConfigPKlt();
		configKlt.pyramidLevels = ConfigDiscreteLevels.levels(4);
		configKlt.templateRadius = 4;
		configKlt.toleranceFB = 3;
		configKlt.pruneClose = true;

		var configDet = new ConfigPointDetector();
		configDet.type = PointDetectorTypes.SHI_TOMASI;
		configDet.shiTomasi.radius = 4;
		configDet.general.maxFeatures = 300;
		configDet.general.radius = 5;

		// We will estimate the location of features using block matching stereo
		var configBM = new ConfigDisparityBM();
		configBM.errorType = DisparityError.CENSUS;
		configBM.disparityMin = 0;
		configBM.disparityRange = 50;
		configBM.regionRadiusX = 3;
		configBM.regionRadiusY = 3;
		configBM.maxPerPixelError = 30;
		configBM.texture = 0.05;
		configBM.validateRtoL = 1;
		configBM.subpixel = true;

		// Configurations related to how the structure is chained together frame to frame
		var configVisOdom = new ConfigVisOdomTrackPnP();
		configVisOdom.keyframes.geoMinCoverage = 0.4;
		configVisOdom.ransac.iterations = 200;
		configVisOdom.ransac.inlierThreshold = 1.0;

		// Declare each component then visual odometry
		PointTracker<GrayU8> tracker = FactoryPointTracker.klt(configKlt, configDet,GrayU8.class, GrayS16.class);
		StereoDisparitySparse<GrayU8> disparity = FactoryStereoDisparity.sparseRectifiedBM(configBM, GrayU8.class);
		StereoVisualOdometry<GrayU8> visodom = FactoryVisualOdometry.stereoMonoPnP(configVisOdom,disparity,tracker, GrayU8.class);

		// Optionally dump verbose debugging information to stdout
//		Set<String> configuration = new HashSet<>();
//		configuration.add(VisualOdometry.VERBOSE_RUNTIME);
//		configuration.add(VisualOdometry.VERBOSE_TRACKING);
//		visodom.setVerbose(System.out,configuration);

		// Pass in intrinsic/extrinsic calibration.  This can be changed in the future.
		visodom.setCalibration(stereoParam);

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

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

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

			System.out.printf("Location %8.2f %8.2f %8.2f    %s\n", T.x, T.y, T.z, trackStats(visodom));
		}
		System.out.printf("FPS %4.2f\n", video1.getFrameNumber()/((System.nanoTime()-startTime)*1e-9));
	}

	/**
	 * If the algorithm implements AccessPointTracks3D create a string which summarizing different tracking information
	 */
	public static String trackStats(StereoVisualOdometry alg) {
		if( !(alg instanceof AccessPointTracks3D))
			return "";

		AccessPointTracks3D access = (AccessPointTracks3D)alg;

		int count = 0;
		int N = access.getTotalTracks();
		int totalNew = 0;
		for( int i = 0; i < N; i++ ) {
			if( access.isTrackInlier(i) )
				count++;
			else if( access.isTrackNew(i) ) {
				totalNew++;
			}
		}

		return String.format("inlier %5.1f%% new %4d total %d", 100.0 * count / (N-totalNew),totalNew,N);
	}
}