Procedura di identificazione dei parametri cinematici di un veicolo e
Transcription
Procedura di identificazione dei parametri cinematici di un veicolo e
IMEKO 2010 London, UK, 1-3 September 2010 Calibration of a vision-based system for displacement measurement in planetary exploration space missions Marco Pertile1, Marina Magnabosco2, Stefano Debei1 1 Dept. of Mechanical Engineering, University of Padova, Via Venezia 1, 35131 Padova (PD), Italy. 2 Cranfield University, UK. PURPOSE To perform a detailed uncertainty analysis and calibration of a measurement system of displacement based on a stereo camera and applied to a simulated planetary scene. METHOD Experimental tests using a reference displacement instrument, to obtain a calibration curve Probabilistic propagation to evaluate the uncertainty of an indirect measurement performed by a mathematical model IMEKO 2010 Experimental tests to evaluate the uncertainty of input quantities used to calculate the output quantity (displacement) London, UK, 1-3 September 2010 INTRODUCTION In planetary exploration space missions, the position of a vehicle on the planet surface can not be measured in a easy way: 1. An odometry system (e.g. optical encoders), that measures the rotation of the wheels, has wide uncertainty due to slippage of wheels on a natural, often sandy or slippery, surface. 2. GPS-like positioning systems are accurate but are not yet available on extraterrestrial planets. 3. Inertial navigation sensors exhibit unacceptable drifts. Need of a reliable and accurate displacement instrument is particularly relevant. In this presentation a measurement system based on a stereo camera is described IMEKO 2010 London, UK, 1-3 September 2010 MEASUREMENT ALGORITHM Camera model A pin-hole camera model is assumed: -X, Y, Z defines the 3D position of landmarks (3D points); -x, y are the coordinates of the projection of a landmark using an ideal camera, aligned like the real camera but with a focal length equal to 1 (in length units). • Intrinsic parameters: define the functional relationship between projection x, y, expressed in length units, and projection x’, y’, expressed in pixels; they comprise: the pixel densities along the two axes, the focal length, the position in the image of the principal point (intersection of the optical axis with the sensor), distortion parameters. • Extrinsic parameters: the relative position and rotation between the two cameras. IMEKO 2010 London, UK, 1-3 September 2010 MEASUREMENT ALGORITHM The measurand is the displacement of a calibrated stereo system and the measurement is performed using the images acquired in an initial position and in a second one. Indirect measurement Output: a numerical evaluation of the displacement Input : 1. Positions in pixels of the projections on the image plane of 3D points (landmarks); 2. Intrinsic and extrinsic parameters of each camera IMEKO 2010 London, UK, 1-3 September 2010 MEASUREMENT ALGORITHM 1. Positions of 2D keypoints on the image plane a) A detector locally analyses the images and finds out regions that are projections of landmarks and can be used as features. The Hessian-Affine detector is selected. b) A descriptor provides representations of the detected regions. Thus, the descriptor allows to search corresponding features (regions that are projections of the same landmark) in the acquired images and to perform their matching: • between the two cameras; • between images acquired by the same camera but in two different positions. This matching phase is required to measure the 3D position of features and then the position of the stereo system. The SIFT (Scale Invariant Feature Transform) descriptor is selected. Image 1 IMEKO 2010 London, UK, 1-3 September 2010 Image 2 MEASUREMENT ALGORITHM 2. Computation of 3D position of landmarks Once the features are identified in images and matched between the two cameras, the 3D positions of landmarks are computed by the middle point triangulation method. 3. Stereo system displacement Once the 3D landmarks are evaluated for two positions of the stereo system, the displacement is computed as the rigid translation that makes the corresponding 3D points overlap. IMEKO 2010 London, UK, 1-3 September 2010 CALIBRATION Determination of intrinsic and extrinsic parameters of the stereo system The first step to calibrate the whole measurement system is the determination of intrinsic and extrinsic parameters of the stereo system. The known Zhang method is selected: 1. It uses a plurality of images of a chessboard acquired in different positions and orientations; 2. It comprises: a) a first analytical evaluation of parameters for both cameras; b) a nonlinear optimization technique based on the maximum likelihood criterion (Levenberg-Marquardt algorithm); lens distortions, especially radial distortion, are taken into account the for each camera. IMEKO 2010 London, UK, 1-3 September 2010 CALIBRATION Calibration of the whole measurement system The stereo system is mounted on a mechanical slide having 2 degrees of freedom and is aimed at a simulated planetary scene obtained with crumpled brown paper. The whole system is calibrated separately along two orthogonal directions: a first substantially aligned one with the optical axes of the cameras (axial displacement) and the second one substantially orthogonal to the optical axes (transverse displacement). For each direction the stereo system is moved from an initial position to a final one with 11 steps, while the displacement is measured by the laser interferometer and by the stereo system in order to build a calibration curve for each direction of translation. IMEKO 2010 London, UK, 1-3 September 2010 UNCERTAINTY ANALYSIS The uncertainty associated with the following quantities are analyzed and evaluated: 1. intrinsic and extrinsic parameters of the stereo system, whose uncertainties are evaluated during the stereo calibration by applying a Monte Carlo simulation to the Zhang method; in this sub-problem the input quantities are the chessboard dimensions, and the positions of the chessboard points on the image plane; 2. FEATURE POSITIONS in the image plane, which have two main contributions: 1. the sensor and reading electronics uncertainty; 2. the lighting angular variation of the scene. IMEKO 2010 London, UK, 1-3 September 2010 UNCERTAINTY ANALYSIS 2D position of features on the image plane A. uncertainty (commonly referred to as image noise in computer vision) associated with the image sensor and its electronics B. Variation of lighting conditions A. Sensor and reading electronics uncertainty depends on the acquired A simulated rocky scene is employed scene Sensor and reading electronics uncertainty is assumed of random type IMEKO 2010 200 images of the same scene were acquired with both cameras in fixed positions and in the same shooting conditions London, UK, 1-3 September 2010 UNCERTAINTY ANALYSIS 2D position of features on the image plane B. Variation of lighting conditions On a planet surface, the sun angular position may change between two acquisition positions. Time tm required for the stereo system to move from an initial position to a final one is evaluated assuming a moving velocity of a rover on a planetary surface Time tm allows to estimate a maximum angular variation (between two consecutive image acquisitions) of the lighting if the planet (e.g. Mars) rotation and orbit is known IMEKO 2010 London, UK, 1-3 September 2010 UNCERTAINTY ANALYSIS Dedicated experimental tests lamp camera slide Simulated scene PC IMEKO 2010 London, UK, 1-3 September 2010 UNCERTAINTY ANALYSIS 2D position of features on the image plane B. Variation of lighting conditions The angular variation of lighting between two images of the same scene make the shadows move and yields a translation of features on the image plane IMEKO 2010 20 images of the same scene are acquired varying the lamp position to simulate the maximum angular variation of sun on the planet surface. London, UK, 1-3 September 2010 RESULTS Measured AXIAL displacement: stereo system vs. laser interferometer. IMEKO 2010 London, UK, 1-3 September 2010 RESULTS Measured TRANSVERSE displacement: stereo system vs. laser interferometer. IMEKO 2010 London, UK, 1-3 September 2010 RESULTS Difference between measurements of the stereo system and the interferometer Axial direction Transverse direction IMEKO 2010 London, UK, 1-3 September 2010 RESULTS Observations: 1. The uncertainty evaluated for axial displacements is larger than that obtained for transverse displacements. Explanations: a) the uncertainty of 3D points acquired by the stereo system is much wider along the axial direction than along a transverse direction, due to the small distance between the two cameras. b) This disadvantage along axial direction is partially compensated by the fact that the number of image features correctly matched in case of axial displacement is generally greater than the number of matched features in case of transverse displacement. 2. Along both directions, the evaluated uncertainty increases with the measured displacement. Explanation: the larger the displacement and the fewer the features correctly matched among images (the averaging effect associated with a large number of matched features decreases along the displacement direction). IMEKO 2010 London, UK, 1-3 September 2010 CONCLUSIONS 1. A vision-based displacement instrument was described and calibrated using a simulated planetary rocky scene. 2. Dedicated experimental tests were performed to evaluate the most significant uncertainty sources using the simulated scene. Particular attention was dedicated to the uncertainty contributions of the feature detector and of lighting conditions. 3. Two different motion directions were analyzed and the evaluated uncertainty were compared. IMEKO 2010 London, UK, 1-3 September 2010 APPENDIXES IMEKO 2010 London, UK, 1-3 September 2010 MEASUREMENT ALGORITHM Computation of 3D position of landmarks The 3D position of landmarks is computed by the middle point triangulation method: For the two projections of the same 3D landmark, the algorithm finds the 3D points with the minimum distance, belonging respectively to the preimage lines of cameras 1 and 2. This points define a segment orthogonal to the two skew preimage lines. The middle point of this segment is selected as the measured 3D point of the landmark (feature). Stereo system displacement To compute the displacement, the following steps are performed: 1. The 3D position is calculated for all features (landmarks) detected by both cameras when the vision system is in an initial position P1. 2. The vision system is moved (cameras are rigidly connected) from the initial position P1 to a second position P2, the same procedure is used to compute the 3D positions of the features detected by both cameras in the second position P2. 3. Common features detected in both positions P1 and P2 are identified. 4. The displacement is computed as the rigid translation that makes the corresponding features overlap. IMEKO 2010 London, UK, 1-3 September 2010 UNCERTAINTY ANALYSIS 2D position of features on the image plane A. Contribution of the sensor and reading electronics uncertainty The scene acquired for uncertainty evaluation is not uniform For each pixel: the mean value calculated using all 200 images is subtracted from the same pixel of all images (normalization) The probability density function (PDF) of the uncertainty contribution is evaluated by the frequency histogram of the normalized grey levels of all pixels of all images. PDF of the normalized grey levels grey levels IMEKO 2010 London, UK, 1-3 September 2010 UNCERTAINTY ANALYSIS 2D position of features on the image plane A. detector and descriptor Corresponding features The sensor and reading electronics uncertainty contribution is added to all pixels Distance between the features found in both images Monte Carlo simulation IMEKO 2010 London, UK, 1-3 September 2010 UNCERTAINTY ANALYSIS 2D position of features on the image plane A. Contribution of the sensor and reading electronics uncertainty All the distances obtained from all iterations are used to build a frequency histogram along the x and y axes This histogram is used to evaluate the PDF of the 1D displacements of image features due to the sensor and reading electronics uncertainty PDF of feature displacements along x axis [pixels] The evaluated PDF are substantially equal along the two x and y axes IMEKO 2010 London, UK, 1-3 September 2010 UNCERTAINTY ANALYSIS 2D position of features on the image plane B. Variation of lighting conditions Comparing two images at a time, displacements of all common features are calculated. PDF of feature displacements along x axis [pixels] Along x IMEKO 2010 The PDFs of feature displacements are evaluated by the frequency histograms along x and y axes. PDF of feature displacements along y axis [pixels] Along y London, UK, 1-3 September 2010