Template reference : 100181708K-EN Crater detection with segmentation-based image processing algorithm M. Spigai, S. Clerc (Thales Alenia Space-France) V. Simard-Bilodeau (U. Sherbrooke and NGC Aerospace, Canada)
Outline Page 2 Context Reference Scenarios, Needs Quick state-of-the-art Algorithm principle Performances, sensitivity Conclusion/Perspective
Context : Precision landing for interplanetary mission Page 3 Mission to mars : reach points of scientific interest on Mars Mission to the Moon : future human base on the Moon Typical need of the required precision of landing is : 100 m 200 m. Asteroid exploration
Standard lander navigation A standard lander navigation Takes an initial position knowledge initialized from ground tracking Propagates the position by the measurement of non-gravitational acceleration with an IMU. The initial position error is typically around 1 km and the IMU integration error is around 600 m 10 km. Page 4 Regarding the need of precision of landing on the Moon, there is clearly a need of terrain-relative sensors to reduce navigation error (altimeter, lidar and/or camera) : Joint PhD thesis NGC Aerospace (Canadian SME) / ESA / TAS-F, 2008-2011, V. Simard- Bilodeau : absolute position and surface relative velocity using surface features during the proximity operation of a planetary mission (orbiting phase and landing phase). Internal TAS-F study on vision-based navigation to improve landing precision : crater detection and identification Moreover terrain-relative sensors are also needed for other purposes such as fine control of terminal velocity and Hazard detection and avoidance. TAS-I studies (Turin, Italy)
Lunar Lander Scenario Reference for the study At a given time, the goal is to detect in an image as much craters as possible and to match them with a reference database in order to improve the current localisation of the lander. Page 5 Craters position errors as input of an Extended Kalman Filter as state estimator
Needs for the present image processing study Reference scenario = Lunar lander braking phase Principle, on the current grey level image acquired : To detect as much craters as possible (not necessarily all the craters) To model each detected crater by an ellipse : localisation, axis, orientation Study focus on the Pd/Pfa and the precision of localisation Expected crater detection algorithm properties: A low probability of false alarm with a reasonable probability of detection A mean localization error that is lower than 3-4 pixels Relatively simple algorithm with real time capabilities Robust to changes of pose, resolution and illumination No a priori information concerning potential crater positions at the current time is taken into account in the study : We are in the worst case! Page 6
Quick State of the art Mainly coming from : TAS-I and TAS-F Cannes Phd Thesis V. Bilodeau. Hough-transform based : Pros : Robust against edge discontinuity. Cons : Requires high computation power, Sensitive to noise. Edge-based : Pros : The algorithm false detection rate is low and accuracy of the crater localisation in pixel is good. Cons : Not robust to noisy edges of old craters (but more robust than the majority of Hough-based algorithm). Requires high computation power. Local low-level features (Harris, SIFT, SURF, ) These features lead from very slow to very fast algorithms and there is a limited robustness to pose/scale/lighting conditions Ellipses Detected Autonomous ly Page 7 Image Filtered Thin Edges 50 50 50 100 pixels 100 150 200 250 pixels 100 150 200 250 pixels 150 200 250 300 300 300 350 400 100 200 300 400 500 600 700 pixels 350 400 100 200 300 400 500 600 pixels 350 400 100 200 300 400 500 600 pixels
Parallel field of research : Target recognition in SAR imagery Exemple of database (MSTAR, non confidential) Page 8 Computation of a feature vector in order to distinguish vehicle classes
Algorithm principle : High-level surface features Main hypothesis : a typical crater is composed of a shadow followed by a bright object representing the illuminated part of the crater. Page 9 Crater object = set of pixels (x1 = columns, x2 = lines), the ellipse estimation is obtained by eigenvalues/eigenvectors of the covariance matrix of (x1,x2). = σ σ 2 1 21 σ σ 12 2 2
Algorithm Steps Initial Image Unsupervised segmentation Example : K-means with N classes Page 10 Selection of potential Dark/Bright objects Prior information : Sunlight «rough» angle and elevation, Min/Max size of objects of interest Ellipse characterization Geometrical check
Data available for the study Page 11 TAS-F Cannes synthetic images With ESA software : PANGU With fine ground-truth Quantitative results Nadir/slant view, High, low and very low sun elevation TAS-F TAS-I Database Real images coming from ESA,NASA Synthetic images, ESA software : PANGU Without fine ground-truth Qualitative visual results Moon, Mars, Asteroid. NASA
Qualitative results (real images) Page 12 Tests on real images : Same algorithm performs correctly on very different images Very low level of false alarm
Qualitative results (real images) Page 13 But the algorithm is not always adapted : For instance, detection is more difficult on highly eroded martian craters Mars, Hourglass Crater, ESA Mars, Crater ice, NASE HiRISE
Quantitative results (synthetic images) Page 14 Test images generated with PANGU Images 512x512, gray level, and associated «ground truth» of crater locations and sizes Three Sun elevations: 77.5, 22.5 and 2.5 (~ Moon pole case) 2 different views: nadir and slant «Raw» images Nadir view, Sun high Nadir view, Sun low Nadir view, Sun very low
Quantitative results (synthetic images) Page 15
Quantitative results (synthetic images) Page 16
Quantitative results (synthetic images) Page 17
Quantitative results (synthetic images) Page 18
Sensitivity to algorithm parameters The general sensitivity to parameters is acceptable. Page 19 The two most critical parameters are : The number of clusters of the k-means segmentation algorithm. An interval of n=[5-6], set empirically, has given quite good robustness during exploitations. A link with the physical description of the different typical terrain of the area of landing should bring benefit to that. The maximum/minimum size of the crater (number of pixels). This parameter is critical because depends a lot on the scene, phase of descent and the resolution of the image. Nevertheless, with the knowledge of the scene, it should be easy to set this parameters. Concerning other parameters : The algorithm is quite robust. It should be noticed that some enhancements should be possible by using a priori knowledge on the scene, phase of descent, etc.
Elements for real times capabilities Page 20 To give a feeling of the RT capabilities, we give here some values of CPU time needed, and which part is the most demanding in CPU. It should be notice that : Algorithm has been coded in MATLAB without optimization. Tests have been performed on a TAS Laptop designed for non-scientific matters => So absolute performances can be improved!
Synthesis on the algorithm Working with shape rather than edges has some advantages Bright/dark pairing is easier Ellipse fitting is direct But also some drawbacks Craters are not detected when dark/bright object is connected to background CPU time / complexity seems compatible with our hypothesis Probably faster than all other algorithms Robustness Seems very good. Behavior is similar on real and synthetic images. Few parameters to adjust, knowledge of sun direction essentially Precision Seems compatible with our hypothesis Crater size error seems only weakly correlated with crater size Detection performance Detection probability is lower than expected, but still manageable Page 21
Conclusion/Perspective Conclusion The algorithm based on segmentation of the scene and ellipse estimation with pixels covariance has been defined and its performances and robustness studied in the case of the lunar lander braking phase. This algorithm seems to be a good candidate for on board navigation. Page 22 Perspective Take into accout prior information Coming from database, including the identification Coming from simple geometric considerations on craters, or in a priori information on phase of descent/position of sun, physical consitutuon of the ground, etc. Test possible enhancements Pre-processing : filtering, enhancement, etc. Other segmentation algorithms