Range estimation from Intensity Gradient Analysis

Abstract

Conventional approaches to recovering depth from gray-level imagery have involved obtaining two or more images, applying an "interest" operator, and solving the correspondence problem. Unfortunately, the computational complexity involved in feature extraction and solving the correspondence problem makes existing techniques unattractive for many real-world robotic applications. By approaching the problem from more of an engineering perspective, we have developed a new depth recovery technique that completely avoids the computationally intensive steps of feature selection and correspondence required by conventional approaches. The Intensity Gradient Analysis technique (IGA) is a depth recovery algorithm that exploits the properties of the MCSO (moving camera, stationary objects) scenario. Depth values are obtained by analyzing temporal intensity gradients arising from the optic flow field induced by known camera motion. In doing so, IGA avoids the feature extraction and correspondence steps of conventional approaches and is therefore very fast. A detailed description of the algorithm is provided along with experimental results from complex laboratory scenes.

<title>Fast algorithm for obtaining dense depth maps for high-speed navigation</title>

“…To fit a plane, we need at least 3 depth points to fall within each region while higher order surfaces require considerably more depth points. The second reason is that the stereo algorithm which we have been using [13] is not accurate enough to distinguish between a curved surface and a flat surface. We have chosen to increase the range of our sensor (i.e., the stereo algorithm) while reducing the depth resolution.…”

“…A sparse depth was created using the Intensity Gradient Analysis (IGA) [13]. The IGA is a fast monocular stereo algorithm which produces a sparse depth map given a sequence of images.…”

<title>Fast algorithm for image-based ranging</title>

“…The quantities within the parenthesis in the disparity vector can be computed if the vehicle-camera displacement, pixel location and the camera focal length are known. Substituting for r from (9) in the Taylor series approximation (6), (7) yields two polynomials E1 -122 a1 6 + a2 2 a3 + a4 + (10) 0=b1 +b2 ö2+b363+b4 45+ (11) The polynomial coefficient a1 depends on the vehicle motion parameters and the first spatial partial derivatives of the images. The coefficient a2 depends on the vehicle motion parameters and the second spatial partial derivatives of the two images and so on.…”

<title>Active depth extraction using image streams</title>

“…Most of these approaches do not provide the ability to actively control the imaging parameters to generate the task specific characterization of scene. Recently Skifstad [7] presented a formulation of range estimation based on spatial and temporal gradients of a fixed sequence of images for small image displacements. Though this approach results in a simple and computationally inexpensive formulation of the problem, it suffers from following limitations 1.…”

“…For very small image displacements, the 3D structure estimation problem can be directly formulated in terms of spatial and temporal gradients without explicitly solving for optical flow or tracking features over frames [7].In the presence of known camera motion, the movement of pixels over a sequence of frames is known with reference to the focus of expansion (FOE). Points exhibiting a small fixed disparity are extracted from each frame.…”