Vision-Based Hazard Detection with Artificial Neural Networks for Autonomous Planetary Landing

In this paper a hazard detection and landing site selection algorithm, based on a single, visible light, camera acquisition, processed by Artificial Neural Networks (ANNs), is presented. The system is sufficiently light to run onboard a spacecraft during the landing phase of a planetary exploration mission. Unsafe terrain items are detected and arranged in a hazard map, exploited to select the best place to land, in terms of safety, guidance constraints and scientific interest. A set of statistical indexes is extracted from the raw frame, progressively at different scales in order to characterize features of different size and depth. Then, a set of feed-forward ANNs interprets these parameters to produce a hazard map, exploited to select a new target landing site. Validation is carried out by the application of the algorithm to images not considered during the training phase. Landing sites maps are compared to ground-truth solution, and performances are assessed in terms of false positives ratio, false negatives ratio and final selected target safety. Results for different scenarios are shown and discussed, in order to highlight the effectiveness of the proposed system.