סמינר מחלקתי 11.5.23

Dr. Dan Feldman
Haifa University Building 505 Room 2
יר צ ה ע ל
Will lecture on
Provably Good Localization for Nano-Drones
Motivated by challenges for autonomous nano-drones (<$100, <100 grams) in our labs, I will formalize some problems in robotics, and suggest solving them using novel algorithms with provable guarantees from computational geometry. Since these drones carry only weak micro-computer (the RPI Zero, <$10, <10 grams), and a monocular RGB camera, most of the problems are in the context of real-time SLAM (Simultaneous Localization and Mapping) via computer vision. For example, in the fundamental registration problem the input is a pair of sets, each containing 3-dimensional n points ("point clouds") that represent two captured images, and the goal is to align the images. Formally, to compute an alignment, which is a rotation (3-by-3 matrix) and a translation (vector in R^3), together with a matching (permutation function) that will minimize the sum of distances from the points in the first set to the matched points in the second set, after its alignment. Heuristics such as the Iterative Closest Point (ICP) algorithm and its variants were suggested for this problem, but none yield a provable non-trivial approximation for the global optimum. We will see the first provable constant factor approximations for this problem in polynomial time, or even O(n) time for a given matching. Extensive experimental results on real and synthetic datasets show that our approximation constants are, in practice, close to 1, and up to x10 times smaller than state-of-the-art algorithms. Based on publications in IROS'22, TPAMI'22, ICCV'21, and joint work with Maalouf, Jubran, Fares, Alfassi, and Ayoub. Short Bio:
Dr. Dan Feldman is a faculty member and the head of the new Robotics & Big Data (RBD) Labs in the CS department of the University of Haifa, after returning from a 3 years post-doc at the robotics lab of MIT.
During his P.hD in the University of Tel-Aviv he developed data reduction techniques known as core-sets, based on computational geometry. Since his post-docs at Caltech and MIT, Dan's coresets are applied for main problems in machine learning, Big Data, computer vision and robotics. His group in Haifa continues to design and implement core-sets with provable guarantees for such problems.