Building Detection Assisted by HYDICE Cues

Introduction

The Sensor Profile

Thematic Maps

After designating the training areas, a feature extraction algorithm is applied to determine a feature subspace that is optimal for discriminating between the specific classes defined. The algorithm used is called Discriminate Analysis Feature Extraction (DAFE). The result is a linear combination of the original 171 (out of 210) bands to form 171 new bands that automatically occur in descending order of their value for producing an effective discrimination. From the MultiSpec output, it is seen that the first 15 of these new features should be adequate for successfully discriminating between the classes.

Having defined the classes and the features, next a classification is carried out. The algorithm in MultiSpec used was the standard Gaussian maximum likelihood algorithm in which the mean vector and covariance matrix for each class are estimated from the training samples. These estimates then allow calculating the likelihood of each class for a given pixel. The label of the most likely class is assigned to the pixel.

Geometric Correction

Geometric Rectification of Thematic Maps

Prof. James Bethel of Purdue University produced this rectified Thematic Map of a portion of the Fort Hood Scene. Note the straight roads. The waviness of the image boundaries gives an idea of the rectification required.

Sensor Registration

Sensor Geo-registration

At USC we approximate an overhead (nadir) viewpoint sensor which we geo-register with the PAN Images. Here two portions of a 3-D Building model constructed from the PAN images are projected on the thematic map to test for support.

Cue Extraction

Thematic Maps are from Professor Ed Mikhail and colleagues from Purdue University

	Consider an example from Fort Hood, Texas:
	This is the corresponding portion of the geometrically corrected and georeferenced thematic map. The geometric correction was provided by Professor Jim Bethel of Purdue University
	We first extract the roof pixels from the thematic map. Many pixels in small regions are misclassified or correspond to objects made of similar materials as the roofs
	Then we extract connected components of reasonable size. Except for one region, these components clearly cue building roofs.

Experiment using HYDICE Cues

HYDICE assisted and PAN only Building Modeling Results

	Line segments from one of the two PAN images used in this experiment. Compare these with the ones that follow.
	Filtered linears. These correspond to those linear segments "near" the HYDICE cues, an 84% reduction in the number of lines used to start the geometric analysis porcess.
	3D model constructed automatically from the lines above. The height of the objects is approximated from shadow and visible wall evidence collected from all multiple (two in this case) views used by the system. Note the total absence of false alarms.
	The 3-D Model constructed without HYDICE cueing contain less models. Although hypotheses for these are actually made, the evidence to support then is weak and thus not reported. The presence of nearby structures such as sidewalks and roads lead to similarly shaped false alarms.
	Results from an integrated process between PAN and HYDICE sensors allows to improve the PAN-only result and also to provide feedback to the HYDICE classifier. The new roof class is better delineated and helps remove non-roof missclassified pixels.

References