CNS/EE 286 Section III
Dynamic Machine Vision for Intelligent Vehicles
- Line following:
The tile patterns on the floor of the Moore building, as well as
the wall baseboard, can be
used to provide visual odometry and lateral guidance.
The robot could use the camera to traverse the hallways of Moore
using any combination of floor tiles, ceiling tiles and wall baseboards
for localization and landmark navigation. The robot would be given goal
positions which it should attempt to reach efficiently.
- Track and follow known object:
The camera can be used to track a known visual pattern. The size and
shape of this pattern in the image gives relative distance and
orientation information. This can be used as control input to have
the robot safely follow the pattern at a fixed distance.
- Neural Network for control:
Most mobile vehicle control systems have explicit representations
of some aspects of the imaged world and robot drive system. Control decisions are
based on these representations. A neural network based control system
would issue drive commands based on visual input based on learned
behavior from training. No explicit representation exists, only
the interconnects between input nodes, hidden layers and control outputs
of the trained system. This project could consists of building, training
and testing a neural network for navigating through hallways.
- VLSI chips: Instead of using conventional CCD imagers
as the vision sensor, specialized VLSI vision chips could be used
to extract visual information that is useful for vehicle control.
Such chips perform low-level vision processing at the pixel
level, reducing the need for large amounts of computational power
at the next level of control.
- Divergent Stereo: It has been demonstrated that motion
control can be modeled after simple biological systems such
as the honeybee. The difference between the optical flow on
the right and left sides of the vehicle can be used as a direct
input into steering control. Keeping the flow difference
zero is equivalent to maintaining a traversal path which is
equidistant between two stationary obstacles. Such systems
do not have explicit representations for obstacles and thus
control vectors can be computed much faster.
- System identification of the Nomad robot and
design of behavior capabilities: By evaluating step responses
in speed and steer rate, reasonable dynamical models shall be derived.
Based on this information feedback control for lateral guidance (wall following)
and stereotypical feedforward control time histories
for curve steering and for obstacle circumnavigation
shall be derived and tested.
- One-D Vision: Image processing is computationally expensive
due to the large amount of data contained in an image array.
Useful information for control can be obtained from single
one dimensional slices of the image plane. This allows for real-time
processing on modest hardware. In the 4-D approach this is exploited
systematically for routine tracking and relative state estimation
tasks by controlling slice selection and feature extraction from
higher level models for given task domains. By use
of several properly oriented 1-D slices, two and three dimensional
properties of given objects can be inferred.
California Institute of Technology,
Pasadena, CA 91125.