Autonomous Rock Segmentation from Lidar Point Clouds Using Machine Learning Approaches
Abstract
Rover navigation on planetary surfaces currently uses a method called blind drive which requires a navigation goal as input from operators on Earth and uses camera images to autonomously detect obstacles. Images can be affected by lighting conditions, are not highly accurate from far distances, and will not work in the dark; these factors negatively impact the autonomous capabilities of rovers. By improving a rover's ability to autonomously detect obstacles, the capabilities of rovers in future missions would improve; for example, enabling exploration of permanently shadowed regions, and allowing faster driving speeds and farther travel distances. This thesis demonstrates how Lidar point clouds can be used to autonomously and efficiently segment planetary terrain to identify obstacles for safe rover navigation. Two Lidar datasets which represent planetary environments containing rock obstacles and sandy terrain were used to train a neural network to perform semantic segmentation. The neural network was based on the RandLA-Net architecture that was designed to efficiently perform semantic segmentation on point clouds using a random sampling algorithm without modifying the point cloud structure. Methods to handle the class imbalance of the datasets were explored to enable the model to learn the minority class and to optimize the model’s performance. The model achieved a recall score of 94.46% and precision score of 84.93% at a frame rate of 0.6238 seconds/point cloud on an Intel Xeon E5-2665 CPU, indicating that it is possible to use Lidar point clouds to perform semantic segmentation on-board planetary rovers with similar compute capabilities.