A deep learning approach is presented to detect safe landing locations using LIDAR scans of the Lunar surface. Semantic Segmentation is used to classify hazardous and safe locations from a LIDAR scan during the landing phase. Digital Elevation Maps from the Lunar Reconnaissance Orbiter mission are used to generate the training, validation, and testing dataset. The ground truth is generated using geometric techniques by evaluating the surface roughness, slope, and other hazard avoidance specifications. In order to train a robust model, artificially generated training data is augmented to the training dataset. A UNet-like neural network structure learns a lower dimensional representation of LIDAR scan to retain essential information regarding safety of the landing locations. A softmax activation layer at the bottom of the network ensures that the network outputs a probability of a safe landing spot. The network is also trained with a cost function that prioritizes the false safes to achieve a sub 1% false safes value. The results presented show the effectiveness of the technique for hazard detection. Future work on electing one landing spot based on proximity to the intended landing spot and the size of safety region around it is motivated.

提出了一种深度学习方法，以使用月球表面的LIDAR扫描来检测安全着陆位置。语义分割用于在着陆阶段通过LIDAR扫描对危险和安全位置进行分类。月球侦察轨道飞行器任务产生的数字高程图用于生成训练，验证和测试数据集。使用几何技术通过评估表面粗糙度，坡度和其他避免危险的规范来生成地面真实情况。为了训练健壮的模型，将人工生成的训练数据扩充到训练数据集。类似于UNet的神经网络结构学习LIDAR扫描的低维表示，以保留有关着陆位置安全性的基本信息。网络底部的softmax激活层可确保网络输出安全着陆点的可能性。还使用成本函数对网络进行了培训，该成本函数将错误保险箱的优先级提高到低于1％的错误保险箱值。提出的结果表明了该技术在危险检测中的有效性。未来的工作是根据与目标着陆点的接近程度和周围安全区域的大小来选择一个着陆点。