Basic Idea for the Thesis - web version

Intended topic: Time-independent Binocular Visual Odometry

Category:         Robotics/Vision

Based on Mobile Robotics Software Challenge (the DARPA Grand Challenge in reverse)

http://www.mobilerobot.org/MRSC/

Problem:

Instead of using a supplied set of coordinates to determine the path of an autonomous vehicle, this project supply a system containing a computer and sensors that will be placed on top of a mobile robot.  The robot will transport the system along an unknown trajectory for and from these observations your system will generate a listing of the trajectory of the robot.

Survey:

-         Highly-calibrated binocular pair

-         Monocular Vision using consumer grade

Historically, visual odometry systems have had difficulty overcoming a number of problems, including numerical instabilities common in SFM-like projective geometric

techniques , sensitivity to the low quality of point correspondences available from automatic tracking algorithms , a requirement for omnidirectional views or the inability to disambiguate simultaneous rotation and translation . A more recent result overcomes these difficulties in monocular- and stereo-camera cases but achieves high accuracy only with a calibrated stereo pair.

Landmarks:

A Robust Visual Odometry and Precipice Detection System Using Consumer-grade Monocular Vision, Jason Campbell, Rahul Sukthankar, Illah Nourbakhsh, and Aroon Pahwa

http://www.cs.cmu.edu/~personalrover/PER/ResearchersPapers/CampbellSukthankarNourbakhshPahwa_VisualOdometryCR.pdf

In the landmark paper above achieves good real-world performance (similar to the calibrated-stereo results) at substantially lower implementation complexity. However limited to the followings:

-         time dependent

-         assumption of lab-environment on flat 2D surface

-         acceleration/movement control of the robotic platform needed.

My proposed thesis overcome these limitations.

Novelty:

-         time independent

-         non-calibrated binocular vision (no high precision calibration necessary)

-         commodity equipment (web camera)

-         not-limited to flat surface lab environment (when time for research allow for application of horizontal level compensation)

Information on Visual Odometry:

Visual odometry offers the prospect of substantially reduced sensing costs, allowing more reliable navigation through unstructured areas and safer operation in close

proximity to humans. As the cost of computation falls, an inexpensive camera can replace a typical sensor suite consisting of dozens of range sensors and a set of encoders and provide a broader field of view and the ability to perform

range and appearance-based sensing simultaneously. Passive vision systems also avoid the multi-path interference problems typical with sonar rangefinders and the high sensitivity to lighting common in low-cost infrared rangefinders.

For kinematically indeterminate robots (e.g., where friction is low and actuation powerful), visual odometry offers a low-latency error signal which can be used in a feedback loop to correct motion. For robots that operate in highly unstructured indoor environments (e.g., urban search and rescue) visual odometry can be significantly more practical and functional than other localization systems because no radio coverage is required, no beacons need be carried/ deployed, static drift is low compared to low-cost inertial measurement units, and high degrees of wheel slip pose

no difficulty.

Implementation:

Platform ?laptop + webcam x 2

Samsung Q25 (Pentium M Processor LV 758 1.5GHz, 256MB RAM)

iPrc WebCamera, 24-bit RGB, 800x600 px, auto white balance, manual focus

Basic Idea:

Picture 1: Calculating angles for the pixel proprotion from the difference between the captured image

Picture 2: Calculating Coordinate of a feature point from the angle from the difference between the 2 camera view

Picture 3: Finding Translational and Positional displacement by gathering feature point coordinates