Video-based system for satellite proximity operations

Transcription

1 7th ESA Workshop on Advanced Space Technologies for Robotics and Automation 'ASTRA 2002' Video-based system for satellite proximity operations Piotr Jasiobedzki(1), Michael Greenspan(2), Gerhard Roth(3), HoKong Ng(1), Neil Witcomb(1) (1) MD Robotics, 9445 Airport Rd., Brampton ON, Canada, L6S 4J3 {pjasiobe, hng, (2) Dept. of Electrical and Computer Engineering, Queen s University, Kingston ON, Canada, K7L 3N6 michael.greenspan@ece.queensu.ca (3) National Research Council of Canada, Bldg. M-50, 1200 Montreal Road, Ottawa ON, Canada, K1A 0R6 gerhard.roth@nrc.ca 1 INTRODUCTION The servicing of satellites in space will extend their life and reduce operational costs. Servicing using an unmanned servicer spacecraft offers a reduction of launching cost when compared with that of manned spacecraft. However, the communication delay, intermittence and limited bandwidth between the ground and on-orbit segments renders direct teleoperated ground control infeasible. The servicer will therefore have to operate with a high degree of autonomy. The on-board sensor system should allow the servicer spacecraft to approach the target satellite, identify its unknown position and orientation from an arbitrary direction, and manoeuvre to align with a docking interface [5,7,10]. Autonomy during rendezvous and docking, proximity operations and actual servicing requires the ability to estimate and track the pose (position and orientation) of the serviced spacecraft, as well as that of its mechanical interfaces and serviceable modules. Automatic analysis of images from servicer cameras can provide such information. The vision systems used currently in space detect visual targets, which simplifies their design. Whereas this approach is attractive when viable, the use of targets introduces operational restrictions, due to the limited ranges of distances and angles for which the targets are visible and can be detected. A better solution would be to develop space vision systems that detect natural surface features and use known models of satellites. 2 SATELLITE PROXIMITY OPERATIONS Currently satellites are deployed, retrieved and serviced during Space Shuttle missions. For example, the Hubble Space Telescope has been serviced four times during its ten years on orbit. Spartan is a family of satellites that carry experimental payloads and Spartans have been flown multiple times on board the shuttle [16]. During satellite rendezvous and capture the shuttle crew uses camera images to locate the satellite at a distance of approximately 100m, to approach it, and finally to capture it with the Canadarm, the shuttle manipulator. Selected images used by the crew and the arm operator during one such mission are shown in Figure 1. The first two images have been captured using the shuttle bay cameras during the approach. Figure 1 Images observed during satellite capture. Left and center images have been captured using the shuttle bay cameras, the right one has been captured by the end-effector camera The second image shows the arm in hovering position just before the final capture phase. The last image has been captured from the end-effector camera after capturing of the interface. It is expected that unmanned satellite servicing will follow the same operational scenarios as does manned servicing. This will entail moving the servicer to the vicinity of the target satellite, preparing the arm for capture and moving the servicer into a position from which the arm can reach the target satellite interface, and finally capturing the satellite. Space vision systems will provide three-dimensional information about the position and orientation of observed spacecraft for such operations. The operating range of such a vision system may extend from approximately 100 m to 1

2 contact. It is expected that other systems relying on GPS, orbital mechanics, radar or Lidar measurements will be used to bring the servicer within this range of the target spacecraft. 2.1 Operational phases Different sensors and algorithms may be used for different phases of the satellite proximity operations, depending on the operational requirements and sensor characteristics (field of view, depth of field, accuracy, etc.). The operations are typically divided into phases according to distance into long, medium and short range. At the long range the vision system must determine the bearing and distance to the satellite of interest. At the medium range the distance, pose and motion must be determined accurately and with high confidence. Only when this information is known with certainty may the servicer spacecraft approach the target satellite (short range) and dock or capture it with a manipulator. The specific distances that correspond to the limits of the long, medium and short ranges depend on the spacecrafts involved (e.g., size, mass, geometry, and manoeuvrability) and safety requirements. Selection of the sensors and their location on the spacecraft depend on these factors and they may be changed for various missions depending on the requirements. For satellite proximity operations that may involve inspection, and docking by a microsatellite [10] or capture by a manipulator the ranges correspond approximately to long: 100m - 20m, medium: 20m - 2m and short: 2 m to contact. The short range subsystem will be used for well defined operations of docking or grasping and the expected range of distance and viewing angles will be restricted by the allowed approach trajectories to the target spacecraft. Lights mounted together with cameras may be used to illuminate the scene. The medium range poses a much more difficult challenge as the viewing angles are unrestricted, the distance varies significantly, and the vision system mostly relies on ambient illumination from the Sun or Earth albedo. For the purpose of our work we assume that the vision system is activated within the medium range and provided with an approximate attitude towards a satellite of interest. 2.2 Expected performance The processing rates are different for different operational phases. At the medium range the vision system will estimate pose rates of several Hz as this data will be used either to maintain the servicer at constant relative position with respect to the target or to bring the arm into a hovering position. The short range will be used for the visual servoing of the endeffector towards the mechanical interface and will have to operate at the rate of at least 10 Hz. This will allow the arm controller to compensate for the target drift or motion of the servicer. The expected accuracy of the vision system is related to the capture envelope of a mechanical interface. Depending on the interface this envelope may extend up to 5-25 cm and 2-10 degrees in angular misalignment. These requirements are mostly relevant for the short range system, which should provide measurements with a much higher accuracy. For the medium range system, an accuracy of 1% of the distance and several degrees of the object orientation will be sufficient. On-orbit illumination and the imaging properties of highly reflective materials used to cover satellites pose significant challenges for any vision system. The illumination changes from day to night as the spacecraft is orbiting the Earth, and for the Lower Earth Orbit the period is 90 minutes. The objects can be illuminated by a combination of direct sunlight (an intense directional source creating hard shadows), Earth albedo (an almost shadow-less source) and on-board lights. The insulation materials used in spacecraft manufacture are either highly reflective metallic foils or featureless white blankets that loosely cover the satellite body. This causes specular reflections and a lack of distinct and visually stable features such as lines or corners. 3 SPACE VISION TECHNOLOGIES Data processed by the vision systems may be obtained from a variety of sensors such as video cameras, and scanning and non-scanning rangefinders. The video cameras commonly used in current space operations are available at a relatively low cost, and have low mass and energy requirements. Their main disadvantage is their dependence on ambient illumination and sensitivity to direct sunlight. Scanning rangefinders [9] are independent of the ambient light and less sensitive to sunlight. This comes at a cost/weight/energy requirement of complex scanning mechanisms. Currently space vision systems rely on the presence of visual targets for their operations. This simplifies the image processing tasks as algorithms tuned to specific targets are used. The targets can be used only in specific tasks (limited by the distance and viewing angles) requiring the definition of all operations of interest at the design phase and precluding any unexpected tasks. Lack or mis-detection of a target may lead to a vision system failure, as there will be not enough data to compute the reliable camera pose solution. There are additional installation and maintenance costs. In general, the targets are suitable only for predefined operations such as vision guided capture at close range range. For a recent review of space vision technologies see [5,7]. 2

3 4 MDR VISION SYSTEM The MDR vision system prototype supports the medium and short range operations, i.e., approximately 20m to 0.2m (the minimum distance corresponds to a stand-off between camera and satellite surface in contact position). This large range of distances implies using different algorithms for different operational phases. The vision system uses passive video cameras that operate under ambient illumination and artificial illumination. The algorithms implemented in this system rely mostly on the presence of natural features and object models. Extraction and processing of redundant data provides robustness to partial data loss. To reduce the mass and power requirements the system uses the same sensing and computing hardware during different phases reconfiguring the system as needed. The current vision system prototype operates at two ranges (medium and short) in four different configurations, see Figure 2. Multiple configurations may run in parallel, use the same cameras and share partially processed data. At the medium range the vision system cameras observe either the complete satellite or a significant portion of it. The system computes sparse 3D data using only natural features observed in stereo images. Three medium range functions are currently supported: 1) Model-free motion estimation, 2) Model-based pose acquisition, 3) Model-based pose tracking. Configuration control User interfaces, data log Vision server id, 3D pose, 3D motion, confidence Spacecraft or Robot Controller monitoring 3D location, 3D motion 3D model (satellite) 3D data sparse 3D computation acquisition id, 3D pose stereo cameras tracking 3D pose, 3D motion target detection target pose estimation & tracking 3D pose, 3D motion 3D model (target) 4.1 3D data computation Figure 2 MDR Vision system architecture A sparse three dimensional representation of the scene is computed from images provided by two calibrated and synchronised stereo cameras. The images are rectified first, then edges extracted in both images independently are matched along the epipolar lines and used produce a sparse representation of the scene. The stereo processing parameters are changed adaptively so as to extract about D points from each stereo pair [8]. 4.2 Model-free motion estimation Model-free motion estimation processes sequences of stereo images and estimates motion of the observed satellite using structure from motion algorithms and sparse 3D data. This motion estimation is used to determine that it is safe to approach the observed satellite and to produce reliable 3D data that is used by pose acquisition module. When dealing with an isolated object, it is the case that the change in camera position is simply the opposite of the change in the object position. Using this principle we compute the satellite motion by simply finding the relative camera motion. Once the satellite reaches the medium range we achieve reasonably accurate depth from the stereo processing. In this situation we combine the stereo and structure from motion (SFM) to compute the camera position. At each position of the stereo cameras we get a set of sparse 2D points. Now using only the right images of the stereo sequence we match the 2D pixel locations of these 2D feature points. More precisely, assume we have two stereo camera positions: left1, right1 and left2, right2. We take the 2D pixel locations of the matching stereo features in right1 and attempt to match them to the stereo features in right2. We are basically performing SFM across the image sequence [14], but using only the 2D stereo features as SFM features. This type of processing is not the same as traditional tracking because we use 3

4 rigidity to prune false matches, which is a very strong constraint. We then use the 2D co-ordinates of these matched features to compute the transformation between the right images of the stereo cameras in the sequence. This method has a number of advantages. First, it avoids the problem of motion degeneracy that is inherent in SFM. Degeneracy occurs when the motion is pure rotation, or when the translational component of the motion is very small. Second, since the stereo rig is calibrated we know the true scale of the computed camera motion. This is not the case when we are using SFM, since without extra information about the scene geometry we cannot know the true scale of the reconstruction of camera motion. Third, the results for small camera motions are no less accurate than the accuracy of the stereo data. When using SFM with small camera motions the reconstruction of the camera path is rarely accurate. However, with our approach the accuracy does not depend on the SFM accuracy, but on the stereo accuracy. We have implemented the second approach and show the results for a sequence in which the camera is tracking a grapple fixture. The final result of this process is the camera path (or satellite motion in the camera reference frame), along with the accumulated set of 3D feature points of the observed satellite, see Figure 3. Figure 3 First and last image of a sequence (left), and computed camera path and Points (right) 4.3 Model based pose acquisition The pose acquisition module uses 3D data produced by the model-free estimation process along with the model of the observed satellite to estimate its pose without using any prior information or constraints. The main idea is to apply 3D binary template matching to the data [17]. The grid elements of a 3D template are called voxels and each such element in a voxel grid represents a small distinct region of space and can be assigned a binary value to indicate whether that region is empty or occupied. The resolution of the voxel grid is quite coarse, a benefit of which is to reduce the size of the pose search space by masking high frequency details. As the satellite can be positioned arbitrarily, a pure template matching approach would require a separate template for each possible pose. The approach described here is to exploit the structure of the object to resolve a subset of the pose parameters, and to apply 3D binary template matching over the reduced space of the remaining free pose parameters. The objective is to determine a rigid transformation X, generally comprising 3 rotations and 3 translations, which aligns a surface model M with the sensed data P of the satellite. The satellite can be positioned arbitrarily within the sensorcentric coordinate system U, and the sensed data is a sparse cloud of 3D points that sample its surface at arbitrary locations. The solution has two distinct phases, each of which solves a distinct subset of the six free degrees-of-freedom (dofs) of the pose. The first phase is motivated by the observation that many satellites, such as the Radarsat, have an elongated structure. We assume that the sensed points are distributed evenly enough along the surface of the satellite to capture this elongation. The vector v that describes the major axis can be efficiently and fairly robustly determined by Principle Component Analysis (PCA). All points in P are first translated by T so that the centroid of the transformed point set Q falls at the origin. Next the covariance matrix C is generated for Q as where the summations are taken over the entire point set. The vector v that points in the direction of the major axis of elongation of Q (and thus P) is determined as the eigenvector corresponding to the maximum eigenvalue of C. The final step of phase one is to identify the two rotations that align v with U x. These two rotations are combined into the transform R yz and applied to the translated point set Q. Examples of the results are illustrated in Figure 4(b). The PCA method has been demonstrated to have a degree of robustness, so that a small number of spurious outliers that may result from background scene elements in the 4

5 experimental system do not seriously skew the result. (In a space borne system the background can be easily filtered, and so significant outliers are not anticipated.) It is also not necessary to sense over the complete structure of the satellite to correctly identify the major axis, as the elongation of the satellite is evident in local regions. Following this first phase, there still exist 4 unresolved dofs: the rotation around U x and all three translations. These are determined in the second phase using a correlation approach, which is essentially an exhaustive 3D template matching over the remaining 4 dofs. Let the function V(M) denote the quantization of surface model M onto a voxel grid, such that all voxels that intersect with the surface of M will have the occupied value, and all other voxels will have the empty value. We define the voxel occupancy as the set of occupied voxels of V(XM) for a given pose X of model M. Let M be initially positioned in a canonical pose within U such that its major axis is aligned with U x, its minor axis is aligned with U y, and its centroid is located at the origin. In preprocessing, M is rotated around U x to n distinct pose values. The voxel occupancy of each rotation is calculated and stored in association with its rotation value. The voxel occupancies are therefore essentially 3D binary templates. At runtime, phase one is first executed so that the point set is centered at the origin and its major axis is aligned with U x. The point set is next quantized into the otherwise empty voxel grid, creating a binary 3-D image of the sensed data. Each of the voxel templates generated during preprocessing are then correlated with the voxel grid, at all possible locations. The template location with the highest correlation value indicates the rotation around U x, and the 3 translation values of the satellite pose, within voxel accuracy. This method has been demonstrated to be efficient, executing in a few seconds, and reliable, succeeding in over 90% of the trials in the full-view images. One limitation is that it applies only to elongated objects. For those satellites that are not elongated, it may be possible to apply a strategy that considers other global properties, such as symmetries, to satisfy the axis alignment of phase one. It may also be efficient to apply the template matching approach of phase two to resolve 5 free dofs, if it were only possible to resolve one dof in phase one. 4.4 Model based pose tracking Figure 4 Pose acquisition example After the initial pose of the satellite is determined, the model based tracking mode is invoked and initialized with this estimate. Tracking operates with high precision and update rate by matching the satellite model with 3D data. Satellite servicing operations that do not require contact, such as fly-around, homing and imaging, are performed in this mode using visual servoing of the arm and/or spacecraft. During tracking the pose is determined by iteratively matching 3D data with the model using a version of the Iterative Closest Point algorithm (ICP) [2]. Our implementation of this algorithm consists of the following steps: 1. Selection of the closest points between the data set and the model in the expected pose 2. Rejection of outliers 3. Computation of geometrical registration between matched data points and the model 4. Application of geometrical registration to the data 5. Termination of the iterations after stopping criterion is reached The algorithm iterates the steps 1-4 until the stopping criterion (convergence and/or reaching the maximum allowed number of iterations) is met. The number of iterations required to reach the minimum depends on the difference between the expected and the actual pose. The expected pose is predicted using a pose tracker that estimates both linear and angular velocities, and applies the correction to the model pose. This predictor significantly reduces the number of iterations and thus reduces the processing time. 5

6 Detection of the closest points between the model and the data sets is the most computationally intensive tasks as it involves computing distances for each combination of data points and model features. The complexity is O(N M), where N is the number of data points, M number of model points, and different model representations and acceleration schemes have been used to reduce it. Grouping model points into simple shapes allows using closed form solutions for computing distances [8]. K-d trees have been used to reduce the number of distance computations for models represented as points or meshes. Our technique eliminates all the distance computations from the on-line phase by computing them in advance off-line. The data is stored in an octree type of a structure that allows efficient storage and fast access. This however comes at a price of having to store the data and/or reduce the resolution of the distances. The model based pose tracking does not require detecting high level features in the scene and matching them with the model features, and re-establishes correspondences between the model and data points at every iteration. This significantly reduces its sensitivity to partial shadows, occlusion and local loss of data caused by reflections and image saturation. An adaptable gating mechanism eliminates outliers that are not part of the model. 4.5 Visual target based pose acquisition and tracking Computed pose of a docking or robotic interface on the target spacecraft is used to servo the end-effector to the capture position. This computation must be performed with high accuracy, reliability and update rate to ensure success and safety. The end-effector must approach the interface along predefined trajectories and may not enter stay-out zones. These constraints on viewing directions and distances allow us to use a visual target mounted next to the interface for this operation. At the close range the observed target satellite surface cannot be simultaneously viewed by both cameras and the vision system processes monocular images. The short range module may use various target designs consisting of linear and circular features arranged in planar or 3D formations. Example targets are shown in Figure 1 and Figure 6. Three dimensional arrangement of features provides higher accuracy with a slightly more complex design. In experiments described in this paper we have used a planar target. The short range processing involves detection of features, pose estimation, and pose tracking and prediction. Each feature is detected by processing an image window centered around its predicted location. The processing involves adaptive segmentation, circle detection, and centroid estimation. Pose estimation maps the centroids to the target model and uses an appropriate closed-form pose estimation algorithm [1,3,6] to handle different number of detected target points and their formations. Each algorithm requires at least four target points to compute the pose. In the case where more than four visual targets are detected, sampling techniques based on Ransac [4] and LMedS [15] are used to reject outliers and select the group of points that gives the best pose information. The system automatically handles occasional loss of features due to adverse illumination or occlusion. Pose tracking estimates linear and angular velocities of the target satellite. This motion model is used to predict pose of the satellite in subsequent observations and reduces the size of search windows used in feature detection. 4.6 Vision system reconfiguration The vision system is re-configured during its operation depending on the requirements of the specific mission, the progress of the operation, and the relative distance and orientation between the satellites. The commands may be sent by a vision system or robotic controller, or by a Guidance, Navigation and Control unit. A diagram illustrating the vision system reconfiguration during the satellite capture is shown in Figure 5. The Vision System controller initializes the medium range configuration upon reaching a distance MR max (the maximum distance where stereo produces 3D data for pose estimation). A short sequence of stereo images is processed using the model-free motion estimation module in order to estimate the satellite motion and compute reliable 3D data. If the detected velocities are below the allowed values the vision system computes the initial pose. This process is repeated until a pose with a sufficiently high confidence value is obtained. Once this is achieved the pose is passed to the medium range tracking module, which starts tracking using this estimate. As the stereo cameras approach the satellite the distance and relative orientation is continuously sent to the controller, and when they reach the values (SR max ) when the short range can be used, this module is then invoked. The short range module acquires the pose of the observed target first and if the confidence is high enough it transitions to a target tracking mode. Both medium and short range modules are executed in parallel until their estimates agree and the distance to the satellite reaches the minimum operational distance for the medium range MR min. The short range operates until the final contact and rigidisation of the end-effector at the distance of SR min. 6

7 Start Stop VS Controller Start Pose Start Pose Stop Medium Range Acq. Tracking Start Stop Short Range Acq. Tracking Far distance M-R max S-R max M-R min S-R min Close distance Figure 5 Vision system reconfiguration during the satellite approach and capture The existing overlap between the operational ranges of different configurations ensures a smooth and correct handover. This provides a safety check, allows for multiple attempts at pose acquisition and does not interrupt the vision based measurements at any time. There is an additional computational cost of running multiple configurations in parallel but reusing partial results produced by common modules reduces this effect. A similar reconfiguration sequence is used during departure from the satellite. A built in hysteresis eliminates unnecessary system reconfigurations if, for example, there is a drift between the satellite and the servicer. 5 EXPERIMENTS AND PERFORMANCE CHARACTERISATION A prototype of the vision system has been tested in a representative laboratory environment under simulated capture of a free-floating satellite. The laboratory setup includes two industrial robots with 6 degrees of freedom each. One of the robots holds a scaled replica of a satellite and the other one an end-effector with cameras. The robots are fully calibrated and they can be programmed to follow predefined trajectories allowing testing of the vision system performance. They can also operate under closed loop control using the vision system tracking modes. The workspaces of both robots overlap and the maximum separation between the replica and the cameras is 6 m. Tests at longer distances are possible, however, without the ability to perform a continuous motion to contact. Testing is performed under illumination simulating direct Sun light and Earth albedo. The Satellite models used in the tests have been manufactured using actual space surface materials to create realistic imaging effects. In the conducted experiments we used two stereo cameras mounted on the end-effector. Cameras were rigidly coupled together and equipped with fixed zoom and focus lenses. Custom developed algorithms for controlling exposure were used to compensate for changes in brightness of the scene. This optical arrangement has proven to be sufficient for the distances from 0.2 m to 5 m. The first set of experiments has been designed to test the performance of the vision system in different operational scenarios. The robots were commanded to follow predefined trajectories and the vision system processed images. Comparison between the recorded robot telemetry and vision system estimates was used to assess the vision system accuracy and robustness. In the second set of experiments the vision system was tested during visual servoing experiments when the computed pose was used to drive a robotic arm to capture a free-floating satellite [11]. Figure 6 shows three images obtained during one of the visual servoing experiments (start position, intemediate and contact). Figure 6 Images from a sequence recorded during experiments (the first one at 5 m, last at 0.2 m) 7

8 The vision system processing is fully implemented in software and runs on a dual processor Pentium III, 700 MHz computer. Initial estimation of the satellite motion and pose acquisition requires several seconds depending on the length of the processed sequence. The pose acquisition takes several seconds depending on the number of 3D points used (1,000-4,000). The model based pose tracking operates at the rate of 2 Hz including computation of over a 1000 of 3D points from stereo images. Tracking of visual targets operates at rates of over 10 Hz depending on the number of features tracked simultaneously (typically 5-10). Accuracy of pose estimates depends on the distance, object size, camera field of view and stereo baseline. At the medium range the accuracy is in order of cm and a few degrees. Before contact the accuracy is in order of a fraction of a degree and 1 mm, which exceeds by an order of magnitude the capture envelope of the end-effector used in experiments. The vision system is robust to partial data loss caused by shadows, specular reflections, occlusion, and non-uniform illumination. 6 CONCLUSIONS We have described a prototype vision system for satellite proximity operations and the results of laboratory experiments conducted using this system under representative conditions. The system relies on satellite models and natural surface features on the satellite for most of its operation. When the observed object is within the stereo range then the system computes 3D data and uses it for pose estimation. At close ranges the vision system operates in a monocular mode. The vision system has been fully implemented and characterised in laboratory experiments under representative conditions and at the distances from 0.2m to 5m. The minimum distance corresponds to an offset between the camera and target satellite surface in the captured position. The depth of field of the camera optics used limited the maximum distance to approximately 5m. Extending the maximum distance to 20m, the maximum range when accurate pose estimation is required, and beyond might require using an additional pair of stereo cameras mounted on the end-effector or on the servicer spacecraft. The field of view, stereo baseline and setting will have to be optimised for operational range. Our vision system could easily process images from such cameras. Selected components of the described vision system are being adapted for use in the Orbital Express Space Demonstration [13], which will be launched in New research and development focuses on integration of the long range subsystem, integration of the system with robotic and spacecraft controllers, and testing in different operational scenarios. 7 REFERENCES 1. M.A. Abidi, T. Chandra. "Pose estimation for camera calibration and landmark tracking",icra, 1990, pp Besl, P. & N. McKay, "A Method for Registration of 3-D Shapes", IEEE PAMI, vol. 14, no. 2, 1992, pp D. DeMenthon and L. S. Davis. "Model-Based Object Pose in 25 Lines of Code", ECCV, pp , M. A. Fischler and R. C. Bolles. "Random sample consensus: a paradigm for model fitting with application to image analysis and automated cartography", Commun. Assoc. Comp. Mach., 24:381-95, Hollander S.,"Autonomous Space Robotics: Enabling Technologies for Advanced Space Platforms", Space 2000 Conference, AIAA Y. Hung, P.S. Yeh, and D. Harwood. "Passive ranging to known planar points sets", ICRA 1985, pp Jasiobedzki P, Anders C: "Computer Vision for Space Applications: Applications, Role and Performance". Space Technology v.2005, Jasiobedzki P, Greenspan M, Roth G.: "Pose Determination and Tracking for Autonomous Satellite Capture", I- SAIRAS Denis G.; Beraldin, J.-Angelo; Blais, Francois; Rioux, Marc; Cournoyer, Luc. "A Three-Dimensional Tracking and Imaging Laser Scanner for Space Operations", SPIE Vol. 3707, pp Ledebuhr A, Kordas J, Ng L, Jones M, et al. "Autonomous, Agile, Mirco-Satellites and Supporting Technologies for Use in LEO Missions", 12 th AIAA/USU Conf. On Small Satellites. 11. Liu M., Jasiobedzki P.:"Behaviour based visual servo controller for satellite capture", ASTRA Oda M, Kine K, Yamagata F., "ETS-VII a Rendezvous Docking and Space Robot Technology Experiment Satellite",46 Int. Astronautical Congress, Oct 1995, Norway, pp Orbital Express G. Roth, A. Whitehead, Using projective vision to find camera positions in an image sequence, Vision Interface 2000, pp , Montreal Canada, May Rousseeuw, A. Leroy, "Robust regression and outlier detection", John Wiley and Sons, New York, Spartan Greenspan, M.A. and Jasiobedzki, P., Pose Determination of a Free-Flying Satellite, MTOR02: Motion Tracking and Object Recognition, Las Vegas, USA, June 24-27,