Semantic Mapping for Disaster Response
Abstract
One of the key challenges in the semantic mapping problem in post-disaster environments is how to analyze a large amount of data efficiently with minimal supervision. To tackle this challenge, we investigate the following research questions: 1) how to train semantic classifier when we have only scarcely labeled training data, 2) how to quickly learn if a new class is introduced to a classifier that has been already trained for other classes, and 3) how to generate navigation paths when a map is no longer consistent with an environment and the only available information is a set of aerial images. In this report, we describe our semantic mapping approach, focusing on three main subtasks. First, we develop learning algorithms that can be quickly trained to generate traversability cost maps using only raw sensor data such as aerial view imagery. Second, we investigate on the problem of learning to detect a new class of object using only a few training examples. Third, we develop a new dataset for rare scenes focusing on post-disaster scenarios, DIsaster SCenarios (DISC) Dataset. This report includes the technical details of our approaches and the evaluation results that show state-of-the-art performance in our experiments. In all subtasks, we collaborate closely with the KAIST team in terms of data sharing as well as technical collaboration.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 07, 2024
- Accession Number
- AD1227765
Entities
People
- Martial Hebert
Organizations
- Carnegie Mellon University