Vision Based GNC Systems for Planetary Exploration

Transcription

1 Abstract Vision Based GNC Systems for Planetary Exploration S. Mancuso ESA ESTEC, Noordwijk, The Netherlands In the frame of planetary exploration, among the Guidance Navigation and Control activities in ESA, Vision Based Navigation is one which has received much attention, because of the flexibility and cost saving aspects associated with camera type sensors. Two applications currently under investigation in this field are: Autonomous Interplanetary Navigation [1] and Navigation for Planetary Approach and Landing [2]. The present paper illustrates the problematic involved, the progresses achieved and the possible future development for these activities in the frame of future missions. Introduction Regarding interplanetary navigation the basic concept consists in estimating the state of an interplanetary spacecraft by using information provided by a camera imaging asteroids and/or planets, and information on the inertial attitude provided by the camera itself or by an additional star sensor. The study has concentrated also on the autonomous and real-time aspects of a possible future flight implementation. Validation of the results will be performed via simulation in three parallel ways: by using images generated synthetically by an image simulator; by using images acquired by a ground telescope, by using images acquired by the Smart-1 AMIE camera during its mission to the Moon. Regarding planetary landing, a very challenging concept has been elaborated which allows estimating the position of the spacecraft with respect to the landing site during its final descent on a planet. The application has been developed in the frame of the Bepi-Colombo mission to Mercury, which represents the baseline mission considered. The estimation of the spacecraft state is based on information provided by the camera, in particular the coordinates of feature points, which are extracted and tracked automatically. In order to be compliant with the strict real-time constraints of the fast dynamic environment typical of a final descent phase, hardware implementation of the image processing algorithms has been performed. In addition, an ad-hoc APS (Active Pixel Sensor) camera will be developed up to the level of an elegant breadboard. Validation will be performed via simulation with hardware in the loop. Autonomous vision-based navigation for interplanetary missions The Autonomous On-Board Navigation for Interplanetary Missions (AutoNav) study has been performed by a consortium of European industries under an ESA contract: EADS-Astrium, France (prime contractor); GMV, Spain; INETI, Portugal; IMCCE, France.

2 Objective The main objective of the study is to develop, for the cruise and encounter phases of an interplanetary mission, an autonomous navigation system based on camera measurement of celestial bodies, in order to restitute the spacecraft position and velocity autonomously (without intervention by ground station). Classical navigation for interplanetary probes is based on radar tracking from very large ground stations, able to provide range and range rate measurements very accurately. Then, by employing complex filtering techniques it is possible to estimate the full state of the spacecraft. These techniques have proven very accurate and robust, however, they require almost constant operator supervision, implying frequent ground-spacecraft contacts, especially during critical phases. Also, classical navigation is not a real-time process; due to the enormous distance involved during interplanetary missions, the signal delay can reach up to 6 hours in our solar system, without considering the time required for the processing of the data itself. For particular applications where relative navigation is required, such as fly-bys, swingbys, planetary entry, rendezvous with celestial objects, the accuracy obtainable is very dependent on the knowledge of the ephemeris of the target celestial body, since we need to provide position and velocity of the spacecraft relative to such bodies; typical accuracy reachable for minor bodies, for which ephemeris are not very well know, is of the order of km. On the contrary, an autonomous system such as the vision-based system under consideration, can overcome some of these drawbacks. In particular, being an onboard and autonomous system it can fulfil the real-time requirements and does not require frequent man in the loop presence. Also, since during relative navigation phases direct measurements relative to the target celestial body are processed, errors on the ephemeris of the target celestial body do not affect the relative navigation accuracy. The absolute navigation accuracy with respect to the inertial space depends largely on the complexity of the filter implemented on-board. Due to limitations in hardware capabilities for space application, it is not conceivable to implement on-board complex filters using very detailed environment models (such as those used on ground), therefore a lower accuracy shall be expected for absolute navigation in comparison with classical ground navigation. However, accuracy requirements for the absolute navigation phase are less stringent, and the accuracy provided by the vision-based on-board system is more than sufficient to navigate through this phase efficiently. Navigation concept The basic idea of the vision-based navigation concept is to use celestial bodies (asteroids and planets) as beacons for the on-board state estimation process. Celestial bodies with well known ephemeris are needed in order to provide sufficient accuracy. Despite the fact that ephemeris are better known for planets (with an error of about 1km) than for asteroids (error can be up to 1000km for the faintest ones), asteroid will be mostly used simply because they are much more numerous than planets and more evenly spread out.

3 The position of the spacecraft can be easily estimated via trigonometric relations as the planar example of Figure 1 shows. LOS Measurement error of beacon 1 Vehicle position uncertainty Spacecraft LOS Measurement error of beacon 2 Figure 1: Navigation concept This concept is not new and was shown to work during the US mission Deep Space 1 which flew in 1998 [3]. The mission objective for showing the feasibility of vision-based navigation was fully accomplished. Two main phases are identified for the interplanetary mission: Cruise and Encounter phase. The navigation processing is stopped at about 3 hours before encounter, since afterwards it would become too expensive to perform correction manoeuvres. During the cruise phase, navigation is absolute; the inertial position of the spacecraft is determined by knowing the Line of Sight (LOS) to the beacons, the beacons position and the inertial spacecraft attitude. During this phase no particular accuracy requirement is needed, the only constraint is to allow switching to the following encounter phase; typical navigation accuracy is of the order km. An image plan shall be established a priori, typically in the average two images a day will be needed for an arc of about one month. However, taking images very frequently would impose constraints at system level, a more efficient strategy is to concentrate all the measurements for one arc at the beginning or at the end, and then process the batches of time-tagged data in the navigation filter. During Encounter Phase, navigation is relative to the target body. The relative state of the Spacecraft can be determined by knowing the LOS to the target body and the previous state history (see Figure 2). Knowledge of the shape of the target can improve the state estimation process, especially with respect to the along track direction. A summary of the accuracy requirements for the various phases is reported in Table 1.

4 LOS t1 r1 LOS t2 r2 LOS Figure 2: Navigation principle for encounter phase Phase Requirement Justification Asteroid Encounter 1 km cross-track 3% of range along-track (up to encounter) Planet Encounter (Gravity Assist) Cruise (Asteroid Encounter) Cruise (Planet Encounter) Low thrust deep space maneuver 10 km cross-track 3% of range along-track (up t0 3 hours) Camera specifications and image processing To allow close encounter. To enable instrument operation scheduling. To minimize error correction consumption 500 km To enable detection of faint objects. To meet along track error requirements km To meet along track error requirements km To allow maneuver realization error at minimum cost. Table 1: Accuracy requirements As the Deep Space 1 mission showed, performances achievable are very dependent on the quality and capability of the camera; the LOS error accuracy is a good quality index with this respect: LOS error accuracy = IFOV * subpixel accuracy (1) where IFOV is the instantaneous field of view, namely the field of view of a single pixel. It has been found that for minor bodies encounter a LOS error accuracy of 5-15 µrad is needed in order to meet the performance requirements specified in Table 1. For approach to major bodies (planets) this requirement can be relaxed due to the brightness and size of the object which prevents achieving high accuracy with image processing algorithms; LOS accuracy of 50 µrad is sufficient in this case. The FOV of the camera is another important parameter; for given detector, the smaller the FOV the higher the accuracy, however with a small FOV we will have

5 less probability to have enough bright stars in the camera background to reconstruct attitude information. It has been found that 0.8 deg is an optimal value. The Image Processing (IP) tasks are responsible for delivering the LOS measurement information to the navigation process. Starting from an image, the IP tasks will have to estimate the correct position of the beacon mass center in detector frame, and at the same time, if possible, provide information about the position of the stars in the background. When the distance spacecraft-beacon is large, the beacon will appear as an unresolved object in the image. In order to obtain subpixel accuracy in this case Multiple Cross-Correlation (MCC) techniques will be employed. These techniques consist of correlating each object in an image with a mask template extracted from the same image. Several non-coincident correlation maxima are generated, which will be interpolated to produce the final reference point with subpixel accuracy. As the spacecraft-beacon distance decreases, the beacon becomes a resolved object in the camera field of view (FOV), different image processing techniques will have to be applied in this case such as center finding techniques, when the object is completely within the camera FOV, or by processing two images of part of the limb to estimate the center of mass of the body from a known a-priori shape. Notice that these processes work quite well for regular shapes (such as sphere, ellipsoid, ), but in the case of an irregular asteroid the estimation error is higher, in this case the Center of Brightness will be estimated instead. Validation The validation of the navigation concept illustrated in the previous sections, will be carried out in three different ways: via simulation using synthesized images, via simulation using real images acquired by ground telescope, via simulation using flight images acquired by the Smart-1 spacecraft with the AMIE camera. The objective of the validation process is twofold: first to validate the developed navigation concept and algorithms, then to validate the real-time aspects. For the validation process two simulators will be employed: the so called Functional Engineering Simulator (FES), and the Prototype Validation Simulator (PVS) which basically represent the AutoNav breadboard and will be used to validate real-time aspects. Matlab/Simulink has been used for the simulators, along with encapsulated C and Fortran code. The first set of simulations will be performed in closed loop using the FES and employing a synthetic image generator; these tests will validate the complete GNC chain. Then tests using images acquired by ground telescopes will be performed using the PVS simulator, the objective is to validate the IP and navigation algorithms on real images, by estimating the position of the observatory in the inertial space, and also to validate the image simulator by comparing synthetic images with real images. Finally, tests using real in-flight images acquired by the Smart-1 satellite with the AMIE camera and the PVS simulator will be performed, the objective is to validate the IP and navigation algorithm on real flight images conducting a realistic orbit determination process and also to validate real time aspects. Notice that while the FES tests are closed loop, the PVS tests (ground and

6 flight images) are necessarily open loop tests, since the guidance and control processes cannot be employed to command the spacecraft. Results The AutoNav study is to be completed by the end of the summer. At the time of writing some preliminary simulations have been performed by using both covariance and Montecarlo methods. Preliminary results show that an accuracy of about 300km can be obtained for the case of the ground tests representing the cruise phase, and an accuracy of 40 km shall be achievable for the Smart-1 case with electrical propulsion, while for encounter the accuracy expected is of the order of few km for encounter with asteroids and of about 10 km for encounter with planet. These are results of preliminary simulation without Image Processing and show that the requirement outlined under Navigation concept, above, can all be fulfilled. The image processing code has just been delivered and is being implemented in the FAS simulator. The above performances will be confirmed with detailed simulations including image processing of the acquired images. Vision-based navigation for planetary approach and landing The Navigation for Planetary Approach and Landing (NPAL) study has been performed by a consortium of European industries under an ESA contract: EADS- Astrium, France (prime contractor); INETI, Portugal; University of Dundee, UK; Science Systems, UK; Galileo Avionica, Italy. Objectives The main objective of the study is to develop a vision-based navigation system for the final descent phase of the landing, to allow soft and precision landing on a planet, based on a single camera. The classical approach for navigation for landing is based on a very accurate orbit determination from ground, before initiation of the descent, and initialization of inertial measurement unit (IMU). After the de-orbit burn, the navigation during the descent is completely inertial. Navigation can be augmented in the last part of the descent with radar or laser altimeters, in this way one ensures that vertical parameters (position and velocity) are within the soft landing domain, however there is no improvement in the horizontal velocity and position, and therefore it is not possible to achieve precision landing with classical approach. Typically, the horizontal overall GNC dispersion at landing will be of the order of km. Also with classical type of navigation it is not possible to perform landing site recognition, hazard avoidance and retargeting. On the contrary, in the vision-based navigation concept relative navigation is performed, better accuracy is achieved and both soft and precision landing are possible. With the vision system it would also be possible to perform retargeting, hazard avoidance and landing site recognition, if an accurate map of the terrain is known a-priori. There are also of course some disadvantages related to a visionbased system like the daylight constraint for landing, which can impose stringent thermal requirements at system level, or adverse environment conditions (heavy dust storms or fog) which could affect the vision system.

7 In the frame of the study the reference baseline considered is the Bepi Colombo mission, an ESA scientific mission to Mercury to be launched in In particular the last part of the descent is analyzed here, starting from an altitude of 8 km at a range of about 20 km from the landing site; when the landing site becomes visible in the camera FOV (high gate) approximately 60 sec before touchdown. Navigation concept The basic idea of the vision-based navigation concept, is to utilize information from IMU and from images of the terrain and fuse them together to estimate the state of the spacecraft relative to the landing site. The fusion process is performed by the navigation filter, an extended Kalman filter with UD formulation has been employed in the frame of this study. The main difficulty here is that terrain features are unknown a-priori and therefore landmark recognition cannot be carried out. Instead the navigation processing will have to be based on unknown features. The solution employed in the present study is the following. Instead of using velocity information from the IMU, obtained via integration of the acceleration measurements (therefore affected by errors), the NPAL concept provides a navigation solution via dynamic filtering based on the acceleration estimation. To illustrate this concept refer to the simplified example shown in Figure 3. Spacecraft X Camera θ x Z D Figure 3: NPAL concept The measurement equation can be written as q = 1 / tan(θ) = x / D (2) where q is the raw measurement from the camera (after image processing), x the distance to the ground (to be estimated) and D is the distance between landmarks (unknown but constant). By differentiating twice this equation we obtain & q& = & x& / D (3)

8 where & x& is the total acceleration acting on the spacecraft. The IMU provides information on the non-gravitational acceleration, for the gravitational acceleration we must rely on on-board models, which, although not very accurate, produce an error of one order of magnitude smaller than other estimation errors. Once & x& is known, from Eq. 3 one can find D, and then from Eq. 2, x. With this technique at least 3 points are needed to estimate the state of the spacecraft, however many more (200 in our baseline) will be tracked for robustness. Image processing An essential part of the navigation concept is the Image Processing (IP). The camera will take images at a relatively high frame rate (20 Hz) in order to cope with the fast dynamics of the spacecraft, then two tasks will be performed at image processing level: extraction and tracking. The choice of the algorithms has been made considering several factors such as: robustness with respect to noise, repeatability and simplicity for hardware implementation. For the extraction process, the trade-off has selected the Harris algorithm [4], which works on gradient masks and provides a criterion map. For the tracking process, tracking via correlation is performed by employing a special technique. Each point extracted in frame N is projected in frame N+1 with a prediction search window. A correlation process is performed within this window, using a predefined correlation template, to generate a local maximum: the tracked point in the frame N+1. In this way, the extraction process is only performed in the initial frame and every time a track is lost. This technique has proven very robust also in presence of noise. The output of the IP task is the list of tracked points, to be sent to the Navigation filter. Camera and image processing hardware As mentioned before, due to the fast dynamics of the final descent phase, the requirements in terms of computation speed are quite stringent. Nominally, for robustness, 200 points will be tracked at a frame rate of 20 Hz. This would require a too high data rate flow between camera and On-Board Computer (above 300 Mbit/sec). To solve this problem image processing algorithms will be implemented in an integrated circuit named FEIC (Feature Extraction and Image Correlation). Regarding the camera detector, an Active Pixel Sensor (APS) has been chosen to cope with the compactness and speed requirements, coupled with a large 70 deg FOV to be able to keep the landing site in the FOV while descending. Within the present study only a camera breadboard will be manufactured, which will be as similar as possible to the flight version. The APS Fillfactory Star1000 has been selected as best candidate detector for the breadboard. The final camera will have a mass of about 0.5 kg and power consumption of about 5W. Validation approach The validation of the navigation concept illustrated in the previous sections will be carried out via simulation in successive steps. The simulator employed for the validation has been developed under Matlab/Simulink and it is named VBNAT (Vision Based Navigation Tool). Four main incremental steps are envisioned for the simulator implementation and for the validation process.

9 VBNAT v1 implements the generic blocks; validation at this stage is with respect to the simulator itself, dynamic equations and environment models. Vision-based navigation is not implemented and only inertial or ideal navigation is used. VBNAT v2 is the successive step; it will implement and validate the vision-based navigation block, however no image processing is embedded at this time and the navigation block input is a list of feature points elaborated a-priori. VBNAT v3 will encapsulate the image processing algorithms, which will be fed with images. The output of the image processing block will then be sent to the navigation processing. Images will be provided by the synthetic terrain generator PANGU, developed under a previous ESA study [5], capable of generating high resolution realistic images at a frame rate of a couple of frames per second. VBNAT v4 will be the final step, and will implement and validate the FEIC by integrating the hardware in the loop. The FEIC will be fed with images generated by PANGU residing on a dedicated workstation. Further to the original plan, an extension has been planned to provide a real time demonstration of the Navigation system. The navigation system will be tested by implementing it on a computer with a Real Time operating system (LinuxRT) with the camera breadboard and FEIC in the loop. In order to support the RT tests, images (collected a-priori) will be injected directly into the camera memory. Results The NPAL study is to be completed by the end of At the time of writing the validation campaign with VBNAT v3, the detailed design of the camera and the FEIC prototyping have been performed. What remains to be done is the manufacturing of the camera, the integration of the VBNAT v4 and the validation campaign. The real-time demonstration and an investigation on the applicability of the concept to the case of Mars landing will also be conducted. In Table 2, the main results for navigation accuracy from the VBNAT v3 campaign are summarized. A comparison is made with an analogous scenario based on pure inertial navigation via an IMU sensor (type III, 0.1 deg/h bias stability). Position wrt Landing site Feature point position Velocity Visual based 32 m (cross track) 11 m (along track) 10 m 1 m/sec (cross track) 1.8 m/sec (along track) Table 2: Navigation Estimation Accuracy IMU 760 m (cross track) 520 m (along track) 8 m/sec (cross track) 8 m/sec (along track The results show that the concept developed in the frame of the present study is feasible, and that landing accuracy in the metric range is achieved. The most

10 complex task, i.e. the estimation of the velocity without any further aiding from dissimilar sensors (e.g. altimeters), has also been performed successfully. Conclusions Camera based navigation systems present many advantages with respect to other classical based system, in particular regarding flexibility, compactness (mass and dimension) and cost. The present paper provides an overview of two of the current GNC activity in ESA in the area of vision-based navigation: Autonomous Interplanetary Navigation and Navigation for Planetary Approach and Landing. The associated preliminary results show that both concepts are feasible and provide very good performances. The conclusions provided in the single sections trace the directions for possible future improvements and continuation. Regarding the AutoNav study, next step would be to solve all the open real-time issues with the help of a real-time avionics test-bed. The concept should also be extended for the late encounter phase (after 3hrs before encounter), for the benefits of aerocapture and aerobraking missions. Finally, a flight experiment should be performed by implementing the concept as part of the GNC system of a spacecraft. A possibility could be provided by the ESA Aurora program. Regarding the NPAL study for landing, next steps in the development of the complete visual based landing system are the detailed development of hazard detection/avoidance and piloting techniques (already started in the frame of another ESA study). The natural continuation of the study would be the development of an avionic test bed for real-time simulation, and the development of a real terrestrial demonstrator. This would represent the ultimate and decisive test to prove the robustness of the concept. After that, possible implementation as a flight experiment on lander probes to the Moon or Mars should be investigated. References [1] B. Polle, B. Frapard, O. Saint-Pé, V. Crombez, S. Mancuso, F. Ankersen, J. Fertig, Autonomous On-Board Navigation for Interplanetary Missions, 26 th Annual AAS Guidance and Control Conference, 5-9 Feb. 2003, Breckenridge, Co, USA. [2] B. Frapard, B. Polle, G. Flandin, P. Bernard, C. Vétel, X. Sembely, S. Mancuso, Navigation for Planetary Approach and Landing, 5th International ESA Conference on Guidance Navigation and Control Systems, Oct 2002, Frascati, Italy [3] J. E. Riedel, S. Bhaskara, et al, Autonomous Optical Navigation DS1 Technology Validation Report, Jet Propulsion Laboratory, California, USA. [4] C. G. Harris, M. Stephens, A Combined Corner and Edge Detector, 4th Alvey Vision Conference, 31 Aug - 2 Sept 1998, Manchester, UK. [5] S. M. Parkes, I. Martin, M. Dunstan, S. Mills, Mercury Surface Simulation for Bepi Colombo Lander, DASIA Conference, May 2002, Dublin, Ireland.