Example Stereo Single Camera

From BoofCV
Revision as of 19:14, 16 October 2018 by Peter (talk | contribs)
Jump to navigationJump to search
Associated inlier features between two views
Associated inlier features between two views
Stereo disparity image
3D point cloud
Stereo disparity image 3D point cloud

A dense point cloud of an environment can be created from a single camera using two views. If the distance between the two views is not known, then the scale of the found point cloud will be arbitrary. In this example natural features are used to determine the geometric relationship between the two views. The algorithm can be summarized as follows:

  • Load camera calibration and two images
  • Detect, describe, and associate image features
  • Compute camera motion (Essential matrix)
  • Rectify image pair
  • Compute dense stereo disparity
  • Convert into 3D point cloud

Example File:


  • Point feature association
  • Epipolar geometry
  • Rectification
  • Dense stereo processing

Related Tutorials/Example Code:

Example Code

 * Example demonstrating how to use to images taken from a single calibrated camera to create a stereo disparity image,
 * from which a dense 3D point cloud of the scene can be computed.  For this technique to work the camera's motion
 * needs to be approximately tangential to the direction the camera is pointing.  The code below assumes that the first
 * image is to the left of the second image.
 * @author Peter Abeles
public class ExampleStereoTwoViewsOneCamera {

	// Disparity calculation parameters
	private static final int minDisparity = 15;
	private static final int maxDisparity = 100;

	public static void main(String args[]) {
		// specify location of images and calibration
		String calibDir = UtilIO.pathExample("calibration/mono/Sony_DSC-HX5V_Chess/");
		String imageDir = UtilIO.pathExample("stereo/");

		// Camera parameters
		CameraPinholeRadial intrinsic = CalibrationIO.load(new File(calibDir , "intrinsic.yaml"));

		// Input images from the camera moving left to right
		BufferedImage origLeft = UtilImageIO.loadImage(imageDir , "mono_wall_01.jpg");
		BufferedImage origRight = UtilImageIO.loadImage(imageDir, "mono_wall_02.jpg");

		// Input images with lens distortion
		GrayU8 distortedLeft = ConvertBufferedImage.convertFrom(origLeft, (GrayU8) null);
		GrayU8 distortedRight = ConvertBufferedImage.convertFrom(origRight, (GrayU8) null);

		// matched features between the two images
		List<AssociatedPair> matchedFeatures = ExampleFundamentalMatrix.computeMatches(origLeft, origRight);

		// convert from pixel coordinates into normalized image coordinates
		List<AssociatedPair> matchedCalibrated = convertToNormalizedCoordinates(matchedFeatures, intrinsic);

		// Robustly estimate camera motion
		List<AssociatedPair> inliers = new ArrayList<>();
		Se3_F64 leftToRight = estimateCameraMotion(intrinsic, matchedCalibrated, inliers);

		drawInliers(origLeft, origRight, intrinsic, inliers);

		// Rectify and remove lens distortion for stereo processing
		DMatrixRMaj rectifiedK = new DMatrixRMaj(3, 3);
		GrayU8 rectifiedLeft = distortedLeft.createSameShape();
		GrayU8 rectifiedRight = distortedRight.createSameShape();

		rectifyImages(distortedLeft, distortedRight, leftToRight, intrinsic, rectifiedLeft, rectifiedRight, rectifiedK);

		// compute disparity
		StereoDisparity<GrayS16, GrayF32> disparityAlg =
						minDisparity, maxDisparity, 5, 5, 20, 1, 0.1, GrayS16.class);

		// Apply the Laplacian across the image to add extra resistance to changes in lighting or camera gain
		GrayS16 derivLeft = new GrayS16(rectifiedLeft.width,rectifiedLeft.height);
		GrayS16 derivRight = new GrayS16(rectifiedLeft.width,rectifiedLeft.height);
		LaplacianEdge.process(rectifiedLeft, derivLeft);

		// process and return the results
		disparityAlg.process(derivLeft, derivRight);
		GrayF32 disparity = disparityAlg.getDisparity();

		// show results
		BufferedImage visualized = VisualizeImageData.disparity(disparity, null, minDisparity, maxDisparity, 0);

		BufferedImage outLeft = ConvertBufferedImage.convertTo(rectifiedLeft, null);
		BufferedImage outRight = ConvertBufferedImage.convertTo(rectifiedRight, null);

		ShowImages.showWindow(new RectifiedPairPanel(true, outLeft, outRight), "Rectification",true);
		ShowImages.showWindow(visualized, "Disparity",true);

		showPointCloud(disparity, outLeft, leftToRight, rectifiedK, minDisparity, maxDisparity);

		System.out.println("Total found " + matchedCalibrated.size());
		System.out.println("Total Inliers " + inliers.size());

	 * Estimates the camera motion robustly using RANSAC and a set of associated points.
	 * @param intrinsic   Intrinsic camera parameters
	 * @param matchedNorm set of matched point features in normalized image coordinates
	 * @param inliers     OUTPUT: Set of inlier features from RANSAC
	 * @return Found camera motion.  Note translation has an arbitrary scale
	public static Se3_F64 estimateCameraMotion(CameraPinholeRadial intrinsic,
											   List<AssociatedPair> matchedNorm, List<AssociatedPair> inliers)
		RansacMultiView<Se3_F64, AssociatedPair> epipolarMotion =
				FactoryMultiViewRobust.baselineRansac(new ConfigEssential(),new ConfigRansac(200,0.5));

		if (!epipolarMotion.process(matchedNorm))
			throw new RuntimeException("Motion estimation failed");

		// save inlier set for debugging purposes

		return epipolarMotion.getModelParameters();

	 * Convert a set of associated point features from pixel coordinates into normalized image coordinates.
	public static List<AssociatedPair> convertToNormalizedCoordinates(List<AssociatedPair> matchedFeatures, CameraPinholeRadial intrinsic) {

		Point2Transform2_F64 p_to_n = LensDistortionOps.narrow(intrinsic).undistort_F64(true, false);

		List<AssociatedPair> calibratedFeatures = new ArrayList<>();

		for (AssociatedPair p : matchedFeatures) {
			AssociatedPair c = new AssociatedPair();

			p_to_n.compute(p.p1.x, p.p1.y, c.p1);
			p_to_n.compute(p.p2.x, p.p2.y, c.p2);


		return calibratedFeatures;

	 * Remove lens distortion and rectify stereo images
	 * @param distortedLeft  Input distorted image from left camera.
	 * @param distortedRight Input distorted image from right camera.
	 * @param leftToRight    Camera motion from left to right
	 * @param intrinsic      Intrinsic camera parameters
	 * @param rectifiedLeft  Output rectified image for left camera.
	 * @param rectifiedRight Output rectified image for right camera.
	 * @param rectifiedK     Output camera calibration matrix for rectified camera
	public static void rectifyImages(GrayU8 distortedLeft,
									 GrayU8 distortedRight,
									 Se3_F64 leftToRight,
									 CameraPinholeRadial intrinsic,
									 GrayU8 rectifiedLeft,
									 GrayU8 rectifiedRight,
									 DMatrixRMaj rectifiedK) {
		RectifyCalibrated rectifyAlg = RectifyImageOps.createCalibrated();

		// original camera calibration matrices
		DMatrixRMaj K = PerspectiveOps.pinholeToMatrix(intrinsic, (DMatrixRMaj)null);

		rectifyAlg.process(K, new Se3_F64(), K, leftToRight);

		// rectification matrix for each image
		DMatrixRMaj rect1 = rectifyAlg.getRect1();
		DMatrixRMaj rect2 = rectifyAlg.getRect2();

		// New calibration matrix,

		// Adjust the rectification to make the view area more useful
		RectifyImageOps.allInsideLeft(intrinsic, rect1, rect2, rectifiedK);

		// undistorted and rectify images
		FMatrixRMaj rect1_F32 = new FMatrixRMaj(3,3);
		FMatrixRMaj rect2_F32 = new FMatrixRMaj(3,3);
		ConvertMatrixData.convert(rect1, rect1_F32);
		ConvertMatrixData.convert(rect2, rect2_F32);

		ImageDistort<GrayU8,GrayU8> distortLeft =
				RectifyImageOps.rectifyImage(intrinsic, rect1_F32, BorderType.SKIP, distortedLeft.getImageType());
		ImageDistort<GrayU8,GrayU8> distortRight =
				RectifyImageOps.rectifyImage(intrinsic, rect2_F32, BorderType.SKIP, distortedRight.getImageType());

		distortLeft.apply(distortedLeft, rectifiedLeft);
		distortRight.apply(distortedRight, rectifiedRight);

	 * Draw inliers for debugging purposes.  Need to convert from normalized to pixel coordinates.
	public static void drawInliers(BufferedImage left, BufferedImage right, CameraPinholeRadial intrinsic,
								   List<AssociatedPair> normalized) {
		Point2Transform2_F64 n_to_p = LensDistortionOps.narrow(intrinsic).distort_F64(false,true);

		List<AssociatedPair> pixels = new ArrayList<>();

		for (AssociatedPair n : normalized) {
			AssociatedPair p = new AssociatedPair();

			n_to_p.compute(n.p1.x, n.p1.y, p.p1);
			n_to_p.compute(n.p2.x, n.p2.y, p.p2);


		// display the results
		AssociationPanel panel = new AssociationPanel(20);
		panel.setImages(left, right);

		ShowImages.showWindow(panel, "Inlier Features", true);

	 * Show results as a point cloud
	public static void showPointCloud(ImageGray disparity, BufferedImage left,
									  Se3_F64 motion, DMatrixRMaj rectifiedK ,
									  int minDisparity, int maxDisparity)
		DisparityToColorPointCloud d2c = new DisparityToColorPointCloud();
		double baseline = motion.getT().norm();
		d2c.configure(baseline, rectifiedK, new DoNothing2Transform2_F64(), minDisparity, maxDisparity);

		CameraPinhole rectifiedPinhole = PerspectiveOps.matrixToPinhole(rectifiedK,disparity.width,disparity.height,null);

		// skew the view to make the structure easier to see
		Se3_F64 cameraToWorld = SpecialEuclideanOps_F64.eulerXyz(-20,0,0,0,0.2,0,null);

		PointCloudViewer pcv = VisualizeData.createPointCloudViewer();

		pcv.getComponent().setPreferredSize(new Dimension(left.getWidth(), left.getHeight()));
		ShowImages.showWindow(pcv.getComponent(), "Point Cloud", true);