Session: 02-02-01: Design, Modeling and Systems

Paper Number: 113616

113616 - Toward Position Approximation Using Asynchronous Multi-View Cameras: A 2D Investigation With Probabilistic Considerations 

Multi-view position recovery is common in the field of computer vision, and applications range from commercial motion tracking to 3D reconstruction of static environments. For these applications, one of two assumptions must be true: (1) camera information is synchronized, or (2) the environment being viewed is static. These assumptions ensure a correspondence between features within the field-of-view (FOV) of the various cameras or camera views used. This is essential for approximating the 3D position of features, and is a requirement for 3D tracking or reconstruction algorithms. Deviations from these assumptions produce large errors and result in unreliable 3D approximations.

This work considers a simulation study of multi-view position approximations of a moving target using two or more 2D pinhole cameras. In this use case, images from cameras are asynchronous and the relative time between images can vary. Images returned from 2D cameras are 1D projections of 2D space (i.e. discrete, 1D pixels are spaced along a segment of fixed length defined by the cameras FOV). For a given camera, approximations of intrinsics and extrinsics (relative to a common reference frame) are known, and intrinsic and extrinsic uncertainty is assumed Gaussian, known, and captured in covariance terms.

For a given camera, target position and velocity are approximated in image space using the time associated with each image, and the current and previous pixel locations of the segmented target. Using this information, synchronous pixel coordinates are approximated across all cameras by propagating position information forward or backward in time using velocity approximations. Target position in 2D space (relative to the common reference frame) is then recovered using this approximation in conjunction with intrinsic and extrinsic information.

In simulations, recovered target position varies widely from true target position. Following intuition, target velocity relative to the time between multi-view images directly impacts the accuracy of tracking. To compensate, uncertainty is considered in the position approximation. This uncertainty includes the assumed Gaussian noise on camera intrinsics and extrinsics, as well as Gaussian noise on target position and velocity in image space. Variance for image space position and velocity is defined using simulated camera resolution, image frequency, and variability in previous position samples. Distributions are then combine using Monte Carlo methods to approximate a single probability distribution representing the target’s uncertain position in space.

Results show a probability distribution that reasonably approximates the target’s true position in 2D space; and the relative size of this distribution shifts based on uncertainty in camera parameters and relative timing between camera images. Extensions of this work to 3D may: (1) relax the need for synchronous data capture, and (2) have implications to data fusion in a variety of cyber-physical system applications (e.g. search and rescue).

Presenting Author: Christopher Civetta United States Naval Academy

Presenting Author Biography: MIDN Christopher A. Civetta is an undergraduate student at the United States Naval Academy (USNA) in Annapolis, MD. He is enrolled in the honors program of the Weapons, Robotics, and Control Engineering Department (WRCE). Topics of interest include computer vision, robotic manipulation, and autonomous vehicles.

Authors:

Christopher Civetta United States Naval Academy
Michael Kutzer United States Naval Academy