Image analysis – Applications – Range or distance measuring
Reexamination Certificate
1999-11-09
2004-03-30
Dastouri, Mehrdad (Department: 2623)
Image analysis
Applications
Range or distance measuring
C382S154000
Reexamination Certificate
active
06714663
ABSTRACT:
1. INTRODUCTION
1.1 Field of the Invention
The present invention pertains to identifying objects in multi-dimensional imagery data and, more particularly, estimating the ground plane in multi-dimensional imagery data.
2. BACKGROUND OF THE INVENTION
2.1 Acquisition of Multi-Dimensional Imagery Data
Multi-dimensional imagery data is an electronic picture, i.e., image, of a scene. Multi-dimensional data may be acquired in numerous ways. Laser Detection And Ranging (“LADAR”) systems are commonly employed for this purpose. Referring to
FIG. 2
, laser signals are transmitted from a platform
18
onto a scene, e.g., a scanned field of view. Upon encountering object(s)
12
and surrounding environment
14
, varying degrees of the transmitted laser signals, characteristic of the particular scene or portion thereof, are reflected back to and detected by a sensor on the platform
18
. The platform
18
can then process the reflected signals to obtain multi-dimensional data regarding the object
12
causing the reflection. The multi-dimensional data captures the distance between the object
12
and the platform
18
, i.e., range, as well as a number of features of the object
18
such as its height, length, width, average height, etc. The quality and accuracy of the features depends in large part on the conditions prevailing at the time the data is collected, including the orientation of the object relative to the platform (e.g., aspect and depression angles), obscurations, and pixel resolution.
The object
12
may be either airborne or, as shown in
FIG. 2
, on the ground
16
. LADAR data is generally acquired by scanning the field of view to generate rows and columns of discrete units of information known as “pixels.” Pixels are used to generate a two-dimensional “image” of the scanned field of view and are correlated to the third dimension, range information. Data acquisition, and particularly LADAR data acquisition is well known in the art and any suitable technique may be employed. Suitable techniques are disclosed and claimed in U.S. Pat. Nos. 5,200,606; 5,224,109; 5,285,461; and 5,701,326.
2.1 Processing Multi-Dimensional Imagery Data
Since platform
18
typically transmits many laser signals across a general area that may contain one or more objects reflecting the laser signals, it is necessary to examine the reflected data to determine if any objects
12
are present and if so, determine which reflecting objects
12
might be of interest. Automatic target recognition (“ATR”) systems are used to identify objects
12
represented in multi-dimensional data to determine whether they are potential targets. ATR systems are often divided into four subsystems: object detection, object segmentation, feature extraction, and object identification.
Object identification is the final process which takes inputs such as the object features discussed above and establishes an identity for the object based on comparison(s) to features of known objects. The accuracy of the identification depends on several factors including the correctness of the object features used in the comparison and the number of known objects constituting potential identifications.
Feature extraction selects one or more features of object
18
, such as its height, width, length, average length, etc., from the multi-dimensional imagery data. However, preceding identification and extraction, object
18
must first be detected and segmented from the environment
14
as portrayed in the multi-dimensional imagery data. This means that the accuracy of detection and segmentation directly influences the accuracy of extraction and identification.
Object detection is essentially the first sweep through the imagery data. It searches for the presence of one or more objects by processing the image data. The imagery data includes pixel information having either x, y or x, y, z coordinates in multi-dimensional space. Pixel coordinates x, y, represent vertical and horizontal position while the z coordinate represents the range, or depth, of a particular point or area in the scene relative to the platform
18
.
The term “pixel” is derived from the phrase “picture element.” A picture (i.e., an image) is a depiction or representation of a scene. Each pixel in the array of pixels which combine to create a picture depicts a certain amount of space in the scene.
Traditional object detection is generally accomplished by locating pixels with variances in coordinates, relative to other pixels, exceeding predefined thresholds. Common detection methods search for object boundaries, object features, or some combination thereof.
An illustrative method of detection entails the analysis of pixel coordinate data relative to linearly adjacent pixels. This method is disclosed in my commonly assigned U.S. Patent Application by Arthur S. Bornowski entitled “Improved Method and Software-Implemented Apparatus for Detecting Objects in Multi-Dimensional Data” filed Oct. 22, 1999, Ser. No. 09/426,559 hereby expressly incorporated by reference herein for all purposes as if fully set forth verbatim. The method rejects relatively homogeneously sloped pixels as ground or surroundings. If a nonhomogeneous slope exceeds a specified threshold, then the method analyzes the pixel's range discontinuity relative to each linearly adjacent pixel. If the range discontinuity, i.e., edge, exceeds a specified threshold the pixel is designates as part of an object. The method continues with each pixel in the multi-dimensional data. This novel method identifies a significant portion of the upper boundary of objects while it rejects relatively homogenous sloping terrain. This method does not sufficiently define the interface between an object and ground. Thus, there is a need to minimize the errors to segmentation and therefore feature extraction and object identification by better estimating the ground plane.
Object segmentation follows the object detection process. The segmentation procedure separates the entirety of the detected object from its surroundings for feature analysis. Detection may not fully delineate the object. Segmentation involves further analysis of the object and surroundings to accurately identify the entire object prior to feature extraction. Ground plane estimation assists in accurate segmentation.
Traditional ground plane estimation relies on both localized and global techniques. These methods typically employ regression techniques of a linear or quadratic fit, using the range as a function of the rows and columns, about the pixels in the approximation. Global techniques assume that the entire scene is flat, and thus all pixels within the scene would be used in the analysis. This technique works well on rather benign scenes but performs poorly on more dynamic scenes. Localized methods, which attempt to estimate a ground plane about an area of interest, perform better on more dynamic scenes.
2.3 Problems With Prior Art Ground Plane Estimation
A significant problem with prior art ground plane estimation methods is inaccurate determination of the object-to-ground interface. This problem can lead to erroneous object segmentation, erroneous feature extraction, erroneous feature comparisons, and ultimately to incorrect or missed object identifications.
The novel detection method leads to more accurate definition of the upper boundaries of an object. However, utilization of traditional ground plane estimation methods would lead to intolerable directional errors in the estimated ground plane. The improved method minimizes this-problem because it locates more object pixels in the traditional ground plane.
However, several new problems are introduced by the illustrative detector, over and above the problems with traditional detectors. Because the improved detector is able to identify more object pixels than conventional detectors, the extra-identified pixels sometimes cause a directional bias in the existing ground plane estimation process.
Accurate segmentation over complex terrain is highly desirable. Complex terrain exacerbates the inaccurate ground/object inte
Dastouri Mehrdad
Kibler Virginia
Lockheed Martin Corp.
Williams Morgan & Amerson
LandOfFree
Method and software-implemented apparatus for ground plane... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method and software-implemented apparatus for ground plane..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method and software-implemented apparatus for ground plane... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3186884