Page 44 - ME-03-digital-heb_Neat

The innovation In his research work, Itzik investigated and developed
three methods for perceptual understanding of information
represented by point clouds: Classification, segmentation,
and geometric analysis that focuses on estimating normal
directions. In classification, the goal was to determine the
class of a given point cloud (e.g., a car, a pedestrian, a
traffic sign.) In segmentation, the goal was to divide the
point cloud into sub-clouds, in order to identify the object’s

sub-class (a wheel, a car hood, headlights, etc.) In
normal direction estimation, the goal was to gain
a better understanding of the geometry of the
different objects. Deep learning networks
were used in all three methods, and the
normal direction estimation method
also involved the development
of a unique Mixture of Experts
network, which enables the
indirect understanding of
complex geometric properties.
Because these networks
do not know how to “read”
the point clouds as natural
input, Itzik developed an
innovative representation
that replaces the point cloud
with structured statistical
information describing it—a
representation that enables
using deep learning methods on
the point clouds. These methods
are a significant contribution toward
creating autonomous robotic systems
that operate in the real world.

Faculty of Mechanical Engineering | MEgazine | 29