Certifiable and Self-Supervised Category-Level Tracking
Abstract
The goal of this project is to lay the mathematical foundations of certifiable category-level target tracking and shape estimation. We develop theory and certifiable algorithms to estimate the trajectory and shape of a target (e.g., a vehicle, aircraft, or spacecraft) from heterogeneous data (e.g., images and depth). We bridge traditional target tracking with modern techniques for pose and shape estimation, which aim at recovering an accurate 3D model of the target, which can be used to classify the type of observed vehicle and obtain more accurate situational awareness and trajectory prediction. The resulting problem is challenging since (i) modern keypoints detectors —based on deep learning— tend to produce several misdetections and outliers and perform poorly outside the training domain, and (ii) the resulting optimal estimation problem is intractable, due to both the presence of outliers and the non-convexity of the state space. Building on previous work by the PI on certifiable geometric perception and estimation contracts.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 06, 2024
- Source ID
- FA95502310382
Entities
People
- Luca Carlone
Organizations
- Air Force Office of Scientific Research
- Massachusetts Institute of Technology
- United States Air Force