See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/358835063

High-Precision Measurement of 3D Rock Morphology on Mars Using Stereo

Rover Imagery

Article in IEEE Geoscience and Remote Sensing Letters · February 2022

DOI: 10.1109/LGRS.2022.3154629

CITATIONS READS
3 654

7 authors, including:

Zhouxuan Xiao Linzhou Zeng

Sun Yat-Sen University Sun Yat-Sen University
31 PUBLICATIONS 81 CITATIONS 4 PUBLICATIONS 7 CITATIONS

SEE PROFILE SEE PROFILE

Yuan Li Jie Shao

Sun Yat-Sen University Chinese Academy of Sciences
29 PUBLICATIONS 460 CITATIONS 38 PUBLICATIONS 782 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Zhouxuan Xiao on 25 February 2022.

The user has requested enhancement of the downloaded file.

 1

High-Precision Measurement of 3D Rock

Morphology on Mars Using Stereo Rover
Imagery
Zhouxuan Xiao, Student Member, IEEE, Linzhou Zeng, Yuan Li *, Jie Shao, Chaohua Ma, Wuming
Zhang*, Man Peng

boundaries facing the sun are often poorly detected; Region-

Abstract—Fine-scale three-dimensional (3D) morphological based methods significantly reduce the computation cost of
features of rocks on the Martian surface provide important clues rock extraction [7, 11], however, it is usually difficult to set
to Mars exploration missions and scientific discoveries. To obtain optimal segmentation parameters, leading to poor detection of
such information, an automatic approach for high-precision
measurement of 3D morphological features of Martian rocks is
rocks; Machine learning-based methods can provide accurate
proposed in this letter. The approach directly detects 3D rocks detection results through learning the features of rocks [7, 12],
from dense point cloud that is generated by interpolating but requiring extensive manual labeling for model training and
triangulated irregular network (TIN) model, through a coarse-to- sometimes need reduplicative parameter tuning. Furthermore,
fine process combining cloth simulation filtering (CSF) and 3D range data was introduced as auxiliary data to refine 2D
connected-component labeling algorithms. Multiple 3D image segmentation results [13, 14]. This integrated method
morphological features are then precisely extracted by modelling
rock points and fitting local ground plane. Experimental results
benefits from the supplementary of 3D range data, but the
demonstrate that the proposed approach provides an effective way extracted rocks are restricted on the 2D image plane, rather than
for the extraction of complete 3D rocks and comprehensive in 3D space. In this regard, rock morphological features are
morphological features with over 90% accuracy. mostly obtained from regular shape fitting methods [2, 15],
Index Terms—Mars rover imagery, 3D rock extraction, rock which mainly aim to coarsely classify rock types. For instance,
morphology, feature extraction ellipse fitting is performed to obtain 2D rock morphological
features [2], and the height of rocks is extracted via the Least-
I. INTRODUCTION squares [2] or RANSAC global plane-fitting [1]. However,

R OCK morphology is associated with the formation and

evolution of the Martian landscape, and also an important
consideration for Mars rover hazardous avoidance [1-3]. Rover
these regular fitting methods cannot finely describe the
irregular surface features of the rock. Furthermore, without
considering the topographic relief, the extracted 3D rocks are
imagery can capture in-situ information of the Martian incomplete using global plane fitting.
surface, and thus enable high-precision measurement and In summary, increasing attention has been paid to Martian
analysis of rock morphology [4-6]. rocks, but there is still a lack of methods that enable high-
To obtain morphological features, rocks should be firstly precision measurement of 3D rock morphological features. To
extracted from the Martian surface [7]. Previous approaches are address this problem, we develop a novel approach to perform
mostly developed from two-dimensional (2D) images or three- 3D morphology measurement of rocks detected from stereo
dimensional (3D) range data. Methods using only 2D images rover imagery. The 3D point cloud with even point density is
can be grouped into edge-, shadow-. region-, or machine generated interpolating triangulated irregular network (TIN)
learning-based. Edge-based methods obtain rock boundaries model after dense matching. To ensure the completeness and
from detected rock edges [8], but often suffering from dust- precision of the extracted rocks, a coarse-to-fine rock extraction
cover and illumination; Shadow-based methods reduce the rock strategy based on cloth simulation and connected-component
detection problem to finding shadows and infer rock boundaries labeling is adopted to optimize the 3D rock points. A novel
from the labeled shadow pixels [9, 10]. However, rock shape fitting strategy combining triangulated irregular network

This work is supported by Open Fund of State Key Laboratory of Remote [email protected]; [email protected];
Sensing Science under Grant No. OFSLRSS202108, the National Natural [email protected]; [email protected]).
Science Foundation of China (grant Nos. 41971380) and Guangxi Natural Yuan Li is with the School of Geospatial Engineering and Science, Sun Yat-
Science Fund for Innovation Research Team (grant Nos. sen University, 519082 Zhuhai, China, and with the State Key Laboratory of
2019GXNSFGA245001). (Corresponding authors: Yuan Li and Wuming Remote Sensing Science, Aerospace Information Research Institute, Chinese
Zhang) Academy of Sciences. Beijing 100101, China (e-mail:
Zhouxuan Xiao, Linzhou Zeng, Jie Shao, Chaohua Ma and Wuming Zhang [email protected]).
are with the School of Geospatial Engineering and Science, Sun Yat-sen Man Peng is with the State Key Laboratory of Remote Sensing Science,
University, 519082 Zhuhai, China (e-mail: [email protected]; Aerospace Information Research Institute, Chinese Academy of Sciences.
Beijing 100101, China (email: [email protected]).
 2

Fig. 2 Coarse rock measurement using CSF algorithm

Fig. 3 Connected-component labeling segmentation

the cloth are calculated. If h is greater than a given threshold,

the corresponding point is preliminarily regarded as candidate
Fig. 1 Overall technical flow of the proposed method rock point, otherwise belong to ground points (Fig. 2).
2) Refinement based on connected-component labeling
and the local ground plane is used to replace the regular fitting The coarse rock candidate points are discrete and unordered
strategy, which considers the topographic relief on the Martian and there can be some outliers. To refine the individual rocks
surface and performs high-precision extraction of rock with high accuracy, a connected-component labeling algorithm
morphological features. [19] is then performed based on a 3D grid, which are deduced
Following the introduction section, the overview and key from an octree structure:
steps of the proposed method will be elaborated in Section II. 𝐿𝐿
𝑑𝑑 = 𝑛𝑛 (1)
Experimental results and their performance assessments are 2
introduced in Section III, after which conclusion is presented at where L is the maximum length of the point cloud space, n is
the end. the octree level, d is the size of leaf cells in the 3D.
Each 3D grid cell with rock points is labeled, and neighboring
II. METHODOLOGY labeled cells are merged into a connected component (Fig. 3).
The flowchart of the proposed approach is illustrated in Fig.1. The grid cells without points become the gaps between
The approach includes 3D data acquisition, coarse-to-fine rock connected components. Each extracted connected components
extraction, and morphological feature extraction, which are represent individual rocks for further morphological features
described detailly in the following subsections. extraction. The outliers are also segmented as individual
components and then eliminated by thresholding (the threshold
A. 3D Data Acquisition is optimized as 300).
Mars surface images collected by NASA’s Mars rovers (e.g.,
C. Morphological Features Extraction
Spirit, Opportunity, Curiosity) are generally archived in the
Planetary Data System [16, 17], and some of the derived 3D To comprehensively characterize rock morphology, six
point cloud data can be directly downloaded from the Analyst’s features, namely length (L), width (W), height (H), surface area,
Notebook website (https://an.rsl.wustl.edu). However, the projected area, and volume (V), are selected in this study. Note
point cloud density progressively decreases from near to far- that the Martian terrain is a surface with topographic relief,
field. To address this problem, a triangulated irregular network which may bring bias in estimating rock morphological features.
(TIN) interpolation is performed to generate dense point cloud Considering the extraction of these morphological features is
data with even density in this study. Besides, constructing TIN closely associated with the surrounding terrain of the rocks, the
can also fill the holes caused by occlusions to some extent and following scheme is constructed, which includes local ground
present the Martin surface in 3D with better consistency. plane fitting and morphological features extraction.
1) Local ground plane fitting
B. Coarse-to-Fine Rock Extraction Previous global plane fitting methods can not eliminate the
1) Coarse extraction based on simulated cloth effects of topographic relief on features extraction. To address
A 3D rock has a continuous surface, which is usually higher this problem, the local terrains around the rocks are established
than the topographic relief. Inspired by the CSF algorithm [18], for subsequent rock morphological analysis. Local ground
a simulated cloth is constructed to separate rocks from the points surrounding each rock can firstly be searched based on
ground. The 3D point cloud is first reversed in the vertical the buffers, which are derived from the refined rock boundaries
dimension and the cloth is generated to cover the reversed point (the size of the buffer is 5 times the point cloud density).
cloud. Then, the vertical distances (h) between the points and Principal component analysis (PCA) is then used for extracting
 3

The OBB’s half-lengths ℎ2 and ℎ3 represent the width and

length of the rock, respectively. Other morphological features
(i.e., surface area, projected area, and volume) can be extracted
using differential computation based on the triangular patches
in the TIN, which have six different intersection cases with the
local ground plane. For the most common case shown in Fig.
5a, the differential volume can be calculated as 𝑉𝑉 (𝑎𝑎) in (3).
Fig. 4 (a)Local ground plane fitting;(b) Surface modeling of rock based on TIN
Triangular patches lying below the plane are not included in the
calculation (Fig. 5b). When the triangular patch intersects the
local ground plane, as the four cases shown in Fig. 5c, Fig. 5d,
Fig. 5e, and Fig. 5f, the differential volume can be calculated as
𝑉𝑉 (𝑐𝑐) , 𝑉𝑉 (𝑑𝑑) , 𝑉𝑉 (𝑒𝑒) and 𝑉𝑉 (𝑓𝑓) in (3), respectively. The final rock
volume is considered as the volume of the triangular prism
formed by TIN. The final surface area and projected area of a
rock are defined as the cumulative surface areas and projected
areas of all triangular patches above the local ground plane
(excluding duplicate projections), respectively, and the rock
height is defined as the maximum distance of all the vertices in
the TIN.
1
⎧ 𝑉𝑉 (𝑎𝑎) = 𝑆𝑆𝑝𝑝 ∗ (𝐻𝐻1 + 𝐻𝐻2 + 𝐻𝐻3 )
⎪ 3
⎪ (𝑐𝑐)
1
𝑉𝑉 = 𝐷𝐷1,2 𝐷𝐷3 (𝐻𝐻1 + 𝐻𝐻2 )
⎪ 6
⎪ 1
(𝑑𝑑)
𝑉𝑉 = 𝑆𝑆𝑝𝑝 ∗ 𝐻𝐻2 (3)
⎨ 3
3
⎪𝑉𝑉 (𝑒𝑒) = 1 𝑆𝑆𝑝𝑝 ∗ � 𝐻𝐻3
− 𝐻𝐻3 + 𝐻𝐻1 + 𝐻𝐻2 �
⎪ 3 (𝐻𝐻1 + 𝐻𝐻3 )(𝐻𝐻2 + 𝐻𝐻3 )
⎪
⎪ 1 𝐻𝐻23
𝑉𝑉 (𝑓𝑓) = 𝑆𝑆𝑝𝑝 ∗
⎩ 3 (𝐻𝐻1 + 𝐻𝐻2 )(𝐻𝐻2 + 𝐻𝐻3 )
where H1, H2, H3 is the vertical distance from the three points
on the triangular patch to the local ground plane, respectively.
Fig. 5 The relationships between triangular patch and the local plane 𝑆𝑆𝑝𝑝 is the projected area of triangular patch and D1,2 is the
projected distance of point-1 and point-2, while D3 is the
the first and second principal components to fit the local ground distance from point-3 to the Line1,2 (Fig. 5).
plane (Fig. 4a).
2) Morphological features III. EXPERIMENTAL ANALYSIS
The rocks extracted from the 3D point cloud are a set of points, A. Study Area and Data
which is difficult to be directly used for calculating shape
properties. Therefore, the 3D rock surface should be modeled The "Spirit" rover traversed the ejecta of Bonneville crater
for extracting the morphological features. Considering the rock and found a volcanic rock dubbed "Humphrey" on Sol.60 (Fig.
surface is usually irregular in shape, we aim to precisely model 6a). In-situ chemical analysis of Humphrey rock indicates the
3D rock surface using triangulated irregular network (TIN) [20], surface water probably shaped the Martian landscape [21], and
which consists of massive small triangular patches including thus the morphological features of this rock are of great interest.
rock points (Fig. 4b). On-board the Spirit, the navigation cameras (Navcams) are
The length and width of the rocks represent the maximum and mounted with a baseline of 0.2m and ~1.5m height above the
minimum scales in two orthogonal horizontal dimensions and ground (Fig. 6b). To verify the performance of our approach,
can be optimally using the oriented bounding box (OBB) [15]. three stereo image pairs taken by the Navcams at Humphrey
Regarding the 3D centroid c as the center point and three sites are selected as experimental data (Fig. 6c). The resolution
mutually orthogonal unit eigenvectors obtained by PCA as axes, of each Navcam image is 1024*1024 pixels, and the focal
the OBB can be denoted as: length of the cameras is 14.67mm with a field of view of 45°.
3 Table I lists the details of stereo images and generated point
OBB = �𝒄𝒄 + � 𝜆𝜆𝑖𝑖 ℎ𝑖𝑖 𝒗𝒗𝒊𝒊 ，𝝀𝝀𝒊𝒊 ∈ [−1,1]� (2) cloud data.
𝑖𝑖=1
where orientation is specified with three mutually orthogonal
unit vectors (𝒗𝒗𝒊𝒊 ) and corresponding eigenvalues (𝜆𝜆𝑖𝑖 ) and half-
lengths (ℎ𝑖𝑖 ) (i = 1, 2, 3).
 4

Fig. 6 (a) Spirit rover traverse map, (b) cameras and (c) experimental images
Fig. 8 Qualitative assessment of extracted rocks in the near and far-field
TABLE I
COLLECTED IMAGES FROM SPIRIT NAVCAM
Original Images Point cloud Density
L-1 2n131246299rad1147p1940l0c1
R-1 2n131246299rad1147p1940r0c1
L-2 2n131344812rad1151p1940l0c1 41,027pts/m2
R-2 2n131344812rad1151p1940r0c1 (2,102,548 points)
L-3 2n131698647rad1159p1940l0c1
R-3 2n131698647rad1159p1940r0c1

Fig. 9 Simulated data and qualitative assessment of extracted rocks

TABLE II
QUANTITATE ASSESSMENT OF ROCK EXTRACTION RESULT
Near Field (<3m) Far Field (>3m)
Items Formula
numbers points numbers points
Fig. 7 (a) Original point cloud, (b) coarse rock candidates, (c) fine extracted
rocks and (d) corresponding bounding box Precision TP/ (TP + FP) 95.31% 98.22% 70.59% 90.64%
Recall TP/ (TP + FN) 88.41% 86.91% 70.59% 48.05%
B. Results and Discussions F-score 2rp/ (r + p) 91.73% 92.22% 70.59% 62.81%
TABLE III
Following the strategy in Section II, the 3D point cloud is MEAN BIAS OF MORPHOLOGICAL FEATURES IN THE NEAR AND FAR FIELD
reconstructed from experimental stereo images (Fig. 7a), and Project Surface
Length Width Height Volume
coarse rock candidates are derived after filtering the ground Field area area
(cm) (cm) (cm) (cm3)
points with a simulated cloth (Fig. 7b). Subsequently, the rock (cm )
2
(cm )2

3D extraction results are refined (Fig. 7c), and the OBB is then Near 1.35 1.32 0.13 19.84 24.45 49.87
Far 3.05 2.79 0.43 85.17 92.03 132.50
constructed for each rock (Fig. 7d).
TABLE IV
1) Qualitative assessment of rock extraction QUANTITATE ASSESSMENT OF ROCK EXTRACTION IN DIFFERENT TERRAINS
The experimental area can be divided into the near field Our Extraction Ground True
(within ~3 m to the rover) and far field (beyond ~3 m to the Terrain type F-score
numbers points numbers points
rover). Fig. 8 shows the comparison between the extracted Flat 25 25974 25 28562 95.26%
rocks and the manually constructed ground truth (GT). The grey Regular 25 27174 25 31567 92.52%
parts refer to extracted rocks that are consistent with the GT Undulating 25 20981 25 27182 87.10%
(true positive, TP). The blue parts refer to rocks that are TABLE V
MEAN BIAS OF MORPHOLOGICAL FEATURES IN DIFFERENT TERRAINS
incorrectly extracted (false positive, FP), and the yellow parts
Project Surface
refer to the missed rocks (false negative, FN). As shown in Fig. Terrain Length Width Height
area area
Volume
7, most rocks at both near and far-field have been successfully type (cm) (cm) (cm) (cm3)
(cm )
2
(cm )2

extracted. Small rocks are unextracted or partially extracted Flat 0.20 0.26 0.57 8.83 23.66 23.58
because they have very low heights or gentle slopes, making it Regular 0.20 0.23 1.24 27.48 43.30 42.43
hard to be distinguished between such rocks and the terrain Undulating 0.86 0.75 1.34 19.10 59.78 54.60
relief. But generally, the overall performance of the developed F-score), where “numbers” and “points” referred to number of
rock extraction method is satisfactory. rocks and number of 3D points corresponding to rocks,
2) Quantitative assessment of rock extraction respectively. Precision represents the correctness of the
The overall performances are synthesized in Table II and III, detected rocks; recall is used to measure the completeness of
and two evaluations (i.e., numbers- and points-based) are the rock detection; F-score is the overall accuracy by taking the
conducted based on three measures (i.e., precision, recall, and harmonic mean of recall and precision. In the near-filed, the
 5

precision, recall, and F-score can reach 98%, 88% and 92%, Recognition, Jun. 2007: IEEE, pp. 1-7, doi:
10.1109/CVPR.2007.383257.
respectively, indicating that the proposed method had a good
[2] K. Di, Z. Yue, Z. Liu, and S. Wang, "Automated rock detection and
performance on extracting complete 3D rocks with high shape analysis from mars rover imagery and 3D point cloud data,"
precision. On the other hand, the assessment values are only 71% Journal of Earth Science, vol. 24, no. 1, pp. 125-135, Feb. 2013, doi:
in the far-field, probably indicating the poor depth estimation 10.1007/s12583-013-0316-3.
[3] M. P. Golombek et al., "Rock size-frequency distributions on Mars
during dense matching in the far-field, resulting in unintegrated and implications for Mars Exploration Rover landing safety and
rock extraction. operations," (in English), J. Geophys. Res.-Planets, vol. 108, no.
3) Assessment of morphological features extraction E12, Oct. 2003, doi: 10.1029/2002JE002035.
To further verify the effectiveness of the proposed method, [4] J. G. Ward, R. E. Arvidson, and M. Golombek, "The size-frequency
and areal distribution of rock clasts at the Spirit landing site, Gusev
simulated data with three different terrain types (Fig.9a, c and Crater, Mars," (in English), Geophysical Research Letters, vol. 32,
e) and various rock sizes and shapes were used and the no. 11, Jun. 2005, doi: 10.1029/2005GL022705.
corresponding rock extraction results are shown in Fig.9b, d and [5] R. Li, K. Di, A. B. Howard, L. Matthies, J. Wang, and S. Agarwal,
"Rock modeling and matching for autonomous long-range Mars
f. Generally, in the simulated data, rocks were extracted under rover localization," (in English), Journal of Field Robotics, vol. 24,
different terrain conditions with remarkably high accuracies no. 3, pp. 187-203, Mar. 2007, doi: 10.1002/rob.20182.
(Table IV). To further quantitatively evaluate the accuracy of [6] R. V. Morris et al., "Identification of Carbonate-Rich Outcrops on
morphological feature extraction, a deviation formula Mars by the Spirit Rover," (in English), Science, vol. 329, no. 5990,
pp. 421-424, Jul. 2010, doi: 10.1126/science.1189667.
quantifying the bias of extraction results is given as follows: [7] G. Li, Y. Geng, and X. Xiao, "Multi-scale rock detection on Mars,"
∑𝑛𝑛𝑖𝑖=1 |𝑂𝑂𝑂𝑂𝑖𝑖 − 𝐺𝐺𝐺𝐺𝐺𝐺𝑖𝑖 | (in English), Sci. China Inf. Sci. (Germany), vol. 61, no. 10, pp.
bias = (4) 102301 (8 pp.)-102301 (8 pp.), Oct 2018, doi: 10.1007/s11432-017-
𝑛𝑛 9277-x.
where OEi and GTVi represent our extracted values and the [8] R. Castano et al., "OASIS: Onboard autonomous science
corresponding ground true values of the ith simulated rock, investigation system for opportunistic rover science," (in English),
respectively. n represents the number of rocks. Journal of Field Robotics, vol. 24, no. 5, pp. 379-397, May 2007,
doi: 10.1002/rob.20192.
As shown in Table V, compared with the ground-truth of the [9] V. C. Gulick, R. L. Morris, M. A. Ruzon, and T. L. Roush,
simulated data, the mean bias of length, width and height "Autonomous image analyses during the 1999 Marsokhod rover
calculated by the proposed method are mostly less than 1cm, field test," Journal of Geophysical Research: Planets, vol. 106, no.
E4, pp. 7745-7763, Apr. 2001, doi: 10.1029/1999JE001182.
while the mean bias of projected area, surface area and volume
[10] Y. Li and B. Wu, "Analysis of Rock Abundance on Lunar Surface
are within tens of cm2/cm3. The overall accuracy of feature From Orbital and Descent Images Using Automatic Rock
extraction can reach 95.65%, 91.06% and 88.12% under the flat, Detection," J. Geophys. Res.-Planets, vol. 123, no. 5, pp. 1061-1088,
regular, and undulating terrain conditions, respectively, further Apr. 2018, doi: 10.1029/2017JE005496.
[11] X. Xiao, H. Cui, M. Yao, and Y. Tian, "Autonomous rock detection
validating the high precision of our approach. on mars through region contrast," Advances in Space Research, vol.
60, no. 3, pp. 626-635, Aug. 2017, doi: 10.1016/j.asr.2017.04.028.
IV. CONCLUSION [12] F. Furlan, E. Rubio, H. Sossa, and V. Ponce, "Rock Detection in a
Mars-Like Environment Using a CNN," presented at the 11th
In this letter, we proposed a routine and effective approach to Mexican Conference on Pattern Recognition (MCPR), Queretaro,
perform extraction of 3D rock and morphological features with MEXICO, Jun., 2019.
[13] L. Matthies, A. Huertas, Y. Cheng, and A. Johnson, "Stereo vision
high precision from stereo Mars rover imagery. The main and shadow analysis for landing hazard detection," in 2008 IEEE
contribution of this letter is that we first perform 3D rock International Conference on Robotics and Automation, May. 2008:
extraction from stereo rover imagery, and extract accurate rock IEEE, pp. 2735-2742, doi: 10.1109/ROBOT.2008.4543625.
[14] S. Woicke and E. Mooij, "A stereo-vision hazard-detection
morphology parameters using TIN and local terrain information. algorithm to increase planetary lander autonomy," Acta
Although some small rocks with gentle slopes can be failed to Astronautica, vol. 122, pp. 42-62, May. 2016, doi:
be extracted, but generally the proposed method had a good 10.1016/j.actaastro.2016.01.018.
[15] I. Cruz-Matias et al., "Sphericity and roundness computation for
performance on extracting complete 3D rocks with high particles using the extreme vertices model," (in English), Journal of
precision (~98% in near field and ~91% in far field). However, Computational Science, vol. 30, pp. 28-40, Jan. 2019, doi:
in the 3D processing, inaccuracies in image depth estimation 10.1016/j.jocs.2018.11.005.
and point cloud filtering are likely to result in incomplete rock [16] P. D. System, "Planetary Data System Standards Reference (Version
3)," ed, 2006.
boundaries in the far-field, which consequently lead to biased [17] P. D. System, "Planetary Data System Standards Reference (Version
3D morphological features extraction. In future research, we 4)," ed, 2020.
will investigate the combination of 2D images to get grayscale [18] W. M. Zhang et al., "An Easy-to-Use Airborne LiDAR Data
Filtering Method Based on Cloth Simulation," (in English), Remote
features of rocks for improving the rock extraction accuracy. In Sens., Article vol. 8, no. 6, p. 22, Jun. 2016, Art no. 501, doi:
comparison with the simulated rock data, the proposed method 10.3390/rs8060501.
can perform robust rock morphological features extraction [19] L. He, Y. Chao, K. Suzuki, and K. Wu, "Fast connected-component
labeling," (in English), Pattern Recognit., Article vol. 42, no. 9, pp.
under different terrain conditions (95.65%, 91.06% and 88.12% 1977-1987, Sep. 2009, doi: 10.1016/j.patcog.2008.10.013.
in the flat, regular, and undulating terrain, respectively). [20] Z. C. Marton, R. B. Rusu, and M. Beetz, "On Fast Surface
Reconstruction Methods for Large and Noisy Point Clouds," in 2009
IEEE International Conference on Robotics and Automation, May.
REFERENCES 2009: IEEE, pp. 3218-3223, doi: 10.1109/ROBOT.2009.5152628.
[1] H. Dunlop, D. R. Thompson, and D. Wettergreen, "Multi-scale [21] H. Y. McSween et al., "Basaltic rocks analyzed by the Spirit rover
features for detection and segmentation of rocks in mars images," in in Gusev Crater," (in English), Science, vol. 305, no. 5685, pp. 842-
2007 IEEE Conference on Computer Vision and Pattern 845, Aug. 2004, doi: 10.1126/science.3050842.

View publication stats