Downloads & Free Reading Options - Results

Thanks to Adrien Bartoli for providing DIKU-3DCV2.zip after the original link became dead: http://isit.u-clermont1.fr/~ab/Classes/DIKU-3DCV2 Course Description The course will cover essential aspects of geometry based Computer Vision. Among the topics are the use of projective and affine projections, homographies, epipolar geometry, the fundamental matrix, and 3D structure computation.  Methods for camera calibration will be touched. As the title of the textbook suggests the course will focus on geometry. Practical problems, such as robust estimation for data containing blunders, are covered too. Some direct applications of these techniques that will be taught are: How a 3D model is built from 2D images and how synthetic images showing a new views of the scene can be rendered. How an artificial object may be inserted in a movie sequence such that the mixed video looks correct. How to construct a panorama from multiple images.

The study of structural characteristics of Unmanned Aerial Systems (UASs) continues to be an important field of research for developing state of the art nano/micro systems. Development of a metrology system using computer vision (CV) tracking and 3D point extraction would provide an avenue for making these theoretical developments. This work provides a portable, scalable system capable of real-time tracking, zooming, and 3D position estimation of a UAS using multiple cameras. Current state-of-the-art photogrammetry systems use retro-reflective markers or single point lasers to obtain object poses and/or positions over time. Using a CV pan/tilt/zoom (PTZ) system has the potential to circumvent their limitations. The system developed in this paper exploits parallel-processing and the GPU for CV-tracking, using optical flow and known camera motion, in order to capture a moving object using two PTU cameras. The parallel-processing technique developed in this work is versatile, allowing the ability to test other CV methods with a PTZ system using known camera motion. Utilizing known camera poses, the object's 3D position is estimated and focal lengths are estimated for filling the image to a desired amount. This system is tested against truth data obtained using an industrial system.

Introduction.  The analysis of approaches to tracking the human body identified problems when capturing movements in a three-dimensional coordinate system. The prospects of motion capture systems based on computer vision are noted. In existing  studies  on  markerless  motion  capture  systems,  positioning  is  considered  only  in  two-dimensional  space. Therefore, the research objective is to increase the accuracy of determining the coordinates of the human body in three-dimensional  coordinates  through  developing  a  motion  capture  method  based  on  computer  vision  and  triangulation algorithms. Materials and Methods .  A  method  of  motion  capture  was  presented,  including  calibration  of  several  cameras  and formalization of procedures for detecting a person in a frame using a convolutional neural network. Based on the skeletal points obtained from the neural network, a three-dimensional reconstruction of the human body model was carried out using various triangulation algorithms. Results.  Experimental studies have been carried out comparing four triangulation algorithms: direct linear transfer, linear least squares method, L2 triangulation, and polynomial methods. The optimal triangulation algorithm (polynomial) was determined, providing an error of no more than 2.5 pixels or 1.67 centimeters. Discussion and Conclusion . The shortcomings of existing motion capture systems were revealed. The proposed method was aimed at improving the accuracy of motion capture in three-dimensional coordinates using computer vision. The results obtained were integrated into the human body positioning software in three-dimensional coordinates for use in virtual simulators, motion capture systems and remote monitoring.

This line of research seeks to increase knowledge of a tracked target using the particle filter, also known as Sequential Monte Carlo (SMC) methods. The target is tracked using vision based observations. These observations were simulated using both dual cameras and a single camera. If only a single camera tracks the target, depth cannot be determined directly and is considered an unobservable state. Filters can estimate this unobservable state using a dynamic model and data from the image. However the movement of the target is nonlinear which eliminated filters traditionally used to track motion such as the Kalman filter and its variants. The particle filter is an alternative that can track nonlinear motion, but was not feasible until recently due to its computational requirements. Simulations of nonlinear target movement, first in two dimensions, then three, evaluated the particle filter's feasibility and performance. Subsequent simulations evaluated the particle filter's ability to track a target using dual and single camera observations. Evaluation tests were devised to characterize the performance of each filter. Analysis metrics were produced to analyze the results of these tests. Linear and Kalman filters were also devised to serve as additional comparisons to the particle filter. Results for dual camera observations demonstrated the filter could track the target and determine unobservable states, however results for the single camera observations indicated the filter was problematic since it could not return accurate depth estimates and suffered from severe weight collapse.

Top: 3D Computer Scene to Line drawing video. Bottom: 3D Real Life Scene to 2D Color Video to Line drawing video. Different programs, same "Center Surround / Mexican Hat wavelet" I've crossed a 3D to 2D, Computer vision process / human vision bridge four ways at once now.