Automated Creation of Labeled Pointcloud Datasets in Support of Machine-Learning Based Perception
Naval Postgraduate School Monterey United States
Pagination or Media Count:
Autonomous vehicles continue to struggle with understanding their environments and robotic perception remains an active area of research. Machine learningbased approaches to computer vision, particularly the increasing application of deep neural networks, have been responsible for many of the breakthroughs in robotic perception over the last decade. We propose a three-phase model for improving pointcloud classification. Progress in applying machine learningbased perception to new problem sets is hampered by the difficulty in creating new training data. As such, our primary contribution is a technique to automate the creation of training data for 3D pointcloud classification problems. Our proposed implementation collects synchronized 2D camera images and 3D LIDAR pointclouds, depth clusters each LIDAR frame to spatially segment a scene, correlates each resultant pointcloud segment to a cropped 2D image, and processes each crop through a 2D image classifier to assign a segment label. Our automated implementation produced labeled 3D pointclouds from raw LIDAR collection and, during testing, yielded a small dataset with 81 percent accuracy of annotations. We also propose a method of scene context discovery to boost pointcloud classification performance. Our approach explores a method to scrape regionally geotagged media for processing through an object-detection neural network. We develop a database mapping of object-type spatial relationships in a specific physical environment and propose applying these relationships as weights to boost pointcloud classifier performance.
- Information Science