Year: 2023
Type: article
Abstract: Accurate understanding and interpretation of the underlying field of view (FoV) are paramount in a real-time planetary landing. This understanding helps detect hazardous bodies and bypasses unfavourable situations well in advance. Existing planetary landing missions rely on the 3-dimensional digital elevation models (DEM) and achieve terrain-relative navigation. These DEMs are used as a reference for spotting the hazards on the pre-defined landing site. These are computationally intensive and time-consuming pattern-matching tasks. The primary concern is the existence of such DEMs before the missions and high storage requirements. This study aims to tackle the abovementioned drawbacks and build a robust intelligent system for autonomous hazard-free planetary landing. This paper utilizes the advanced deep learning approach for accurately detecting and positioning the hazards in the current FoVs using vision sensors of the spacecraft. Deep convolutional neural networks are utilized for feature extraction purposes. These features are further utilized by the recurrent neural network's region proposal algorithm to spot the distinct regions inside the current FoV. These proposals are the keys to detecting the hazards like craters and boulders. The detection results are interpreted by classifying the hazards into craters and boulders. It also classifies the safe landing region as a plain surface. Further, the classes are positioned accurately using the bounding boxes of the coco model. Transfer learning is used to build and train the network. The work also includes the creation of a valid planetary dataset required for generating a ground truth. The overall results are validated through comparative judgments and exhaustive analysis. Experimental results show that the transfer learning approach for hazard detection and localization achieved excellent results.
Institutions: Savitribai Phule Pune University, Department of Space
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7216955909
