I've provided descriptions of a handful of course projects I've worked on, as well as which courses they were for.
Stereo Camera Calibration using an Artificial Neural Network
Course: Intelligent Systems Design (masters course)
Term: Fall 2009
Stereo vision continues to be an important part of the fields of robotics and machine vision, and there is a growing interest in using it as a method for rapid construction of 3-D models for applications such as virtual reality. Furthermore, we are beginning to see offerings of consumer-grade 3D cameras. As such, it is important to be able to quickly, accurately, and cheaply calibrate stereo cameras systems.
Traditional stereo calibration consists of two steps: the correspondence problem, and the callibration problem. The calibration problem is typically solved by extensively characterizing both the intrinsic camera parameters (e.g. focal length, CCD inter-pixel spacing, lens distortion), and the extrinsic parameters (e.g. distance between the two cameras).
This paper presents a method for performing the calibration component of the stereo problem using an artificial neural network; essentially, letting a neural network learn the relationship between the coordinate disparity and the distance from the camera based on a series of test measurements. It also presents an experiment demonstrating the effectiveness of the proposed method, and proposes future work to strengthen this approach.
Change Detection with a Moving Camera for the Purpose of Moving-Object Removal
Course: Machine Vision (masters course)
Term: Winter 2010
The problem of motion detection with a moving camera, or similarly, the tracking of moving objects with a moving camera, has many applications in the computer vision and machine vision fields. Examples include military target tracking, autonomous driver systems, and maneuvering industrial robotic manipulators in a dynamic environment. This problem remains largely open at present, though much work has been done in this area, usually falling under one of two broad categories: sparse feature set tracking, or motion compensation combined with image differentiation.
I propose a system base on compensating for the motion of the camera, then performing image differencing. Motion compensation is approached as a registration problem and leverages established feature detection and optical flow techniques. Traditional image differencing is enhanced to compensate for minor mis-registrations. The system is shown to perform well in testing.
Blending Techniques for Panoramic Image Stitching
Course: Image Processing and Image Communication (masters course)
Term: Winter 2010
This report provides an introduction to the panorama problem and discusses several existing blending techniques: basic averaging, weighted averaging, and p-norm averaging. A new hybrid method is proposed in which the blending width of the p-norm method is varied based on frequency.
In the proposed system, each image in the panorama is low-pass and high-pass filtered. Then, the low-pass images are blended over a wide area to account for exposure differences and vignetting, while the high-pass images are blended over a narrow region to minimize blurring and doubling of features to which the human eye is more sensitive. An implementation of both the existing techniques and the proposed system is described and experimental results are provided.
The proposed method is shown to perform well on test data, combining the strengths of weighted averaging in blending low frequencies with those of the p-norm in blending high frequencies.