Friday, December 6, 2019 - 10:00am
Location:Traffic21 Classroom 6501 Gates Hillman Centers
Speaker:MENGYUN XU, Masters Student /MENGYUN%20XU
Improving the Accuracy and Runtime of Probabilistic and Deep Learning-Based Pose Estimation
Many applications in robotics require estimating pose, consisting of translation and orientation, between a model frame and a sensor frame. One popular application is robot grasping, where the object pose can be estimated by probabilistic or deep learning-based approaches. This thesis consists of a survey on three methods aimed at improving pose estimation accuracy and runtime. The first method models uncertainty using various distributions and examines their ability to improve the accuracy of the resulting output, when compared to current methods that do not consider uncertainty in their models. The second method searches for the relationship between common system variables and the performance of neural networks with different onfigurations, in order to accelerate computation by developing a guideline to select the optimal parallel model given certain problem characteristics and hardware resources. The third method improves both accuracy and runtime by mapping the original non-linear and non-convex pose estimation problem into an alternative parameter space where the original problem can become truly linear. The lack of linearization or other approximations avoids high sensitivity to initial estimation error and high computation time.
David Held (Chair)