TY - JOUR
T1 - Multi-type Terrain Detection and Evaluation on Planet Surface Based on Deep Learning
AU - Ting, Song
AU - Yongjun, Zhou
AU - Gao, Ai
N1 - Publisher Copyright:
Copyright © 2023 by the International Astronautical Federation (IAF). All rights reserved.
PY - 2023
Y1 - 2023
N2 - Aiming at the potential threat of diverse and heterogeneous Terrain on planetary surfaces to the landing safety of probes, this paper proposes a multi-type terrain detection method based on deep learning. Landing detection is a necessary prerequisite for in situ exploration and scientific research on planetary surfaces, and the assessment and selection of mission landing sites is the first key problem to be solved to realize safe landing on planetary surfaces. Currently, the commonly used engineering method is to screen the Terrain and select the target landing area by the ground operator, which requires human involvement and lacks real-time performance. The future planetary landing exploration mission will have higher autonomy requirements, in order to ensure the safety of the landing exploration mission and realize the autonomous selection of landing sites, there is an urgent need to develop an online detection algorithm for different Terrain. Since different kinds of Terrain will present different textures, brightness, shadows and other features in the image, the current traditional methods usually need to design different feature extraction algorithms for different categories of Terrain, and there is a lack of unified detection methods. For deep learning-based methods, the quality and size of the dataset are crucial to the effectiveness of the Terrain detection algorithm, and the unstructured nature of most of the terrains on the planetary surface increases the difficulty of labeling the data. To address the above problems, this paper proposes a deep learning-based Terrain detection method, which utilizes weakly supervised learning to construct a terrain segmentation dataset and adopts a convolutional neural network model to realize terrain detection. First, a planetary surface terrain classification network model is constructed on the basis of the planetary surface terrain classification database, and the class activation map method is used to extract the coarse localization results of Terrain in the image. Second, the inner pixel method model and random wandering method are used to improve the terrain detection results of class activation maps, and the corresponding masks of Terrain are generated. Then, the semantic segmentation network is trained to generate the comprehensive evaluation results of Terrain based on the prediction results of the network. Finally, the effectiveness and feasibility of the present method is verified by testing it on real planetary surface images.
AB - Aiming at the potential threat of diverse and heterogeneous Terrain on planetary surfaces to the landing safety of probes, this paper proposes a multi-type terrain detection method based on deep learning. Landing detection is a necessary prerequisite for in situ exploration and scientific research on planetary surfaces, and the assessment and selection of mission landing sites is the first key problem to be solved to realize safe landing on planetary surfaces. Currently, the commonly used engineering method is to screen the Terrain and select the target landing area by the ground operator, which requires human involvement and lacks real-time performance. The future planetary landing exploration mission will have higher autonomy requirements, in order to ensure the safety of the landing exploration mission and realize the autonomous selection of landing sites, there is an urgent need to develop an online detection algorithm for different Terrain. Since different kinds of Terrain will present different textures, brightness, shadows and other features in the image, the current traditional methods usually need to design different feature extraction algorithms for different categories of Terrain, and there is a lack of unified detection methods. For deep learning-based methods, the quality and size of the dataset are crucial to the effectiveness of the Terrain detection algorithm, and the unstructured nature of most of the terrains on the planetary surface increases the difficulty of labeling the data. To address the above problems, this paper proposes a deep learning-based Terrain detection method, which utilizes weakly supervised learning to construct a terrain segmentation dataset and adopts a convolutional neural network model to realize terrain detection. First, a planetary surface terrain classification network model is constructed on the basis of the planetary surface terrain classification database, and the class activation map method is used to extract the coarse localization results of Terrain in the image. Second, the inner pixel method model and random wandering method are used to improve the terrain detection results of class activation maps, and the corresponding masks of Terrain are generated. Then, the semantic segmentation network is trained to generate the comprehensive evaluation results of Terrain based on the prediction results of the network. Finally, the effectiveness and feasibility of the present method is verified by testing it on real planetary surface images.
KW - Planetary landing
KW - Semantic segmentation
KW - Terrain detection
KW - Weakly supervised learning
UR - http://www.scopus.com/inward/record.url?scp=85187998549&partnerID=8YFLogxK
M3 - Conference article
AN - SCOPUS:85187998549
SN - 0074-1795
VL - 2023-October
JO - Proceedings of the International Astronautical Congress, IAC
JF - Proceedings of the International Astronautical Congress, IAC
T2 - 74th International Astronautical Congress, IAC 2023
Y2 - 2 October 2023 through 6 October 2023
ER -