A multilayer perceptron hazard detector for vision-based autonomous planetary landing

A hazard detection and target selection algorithm for autonomous spacecraft planetary landing, based on Artificial Neural Networks, is presented. From a single image of the landing area, acquired by a VIS camera during the descent, the system computes a hazard map, exploited to select the best target, in terms of safety, guidance constraints, and scientific interest. ANNs generalization properties allow the system to correctly operate also in conditions not explicitly considered during calibration. The net architecture design, training, verification and results are critically presented. Performances are assessed in terms of recognition accuracy and selected target safety. Results for a lunar landing scenario are discussed to highlight the effectiveness of the system.