Keywords: SNAKE, Fitting, High resolution image, Residential maps update
Abstract
It is expected that spatial data with various forms and content will be used in city in the future. It is advantageous to integrate various source data to generate urban spatial information. For instance, laser scanner has only limited resolution, though it can acquire three dimensional data directly. On the other hand, CCD image has very high resolution and conveys useful information for recognizing object though stereo matching methods are necessary for the extraction of 3D shape and the reliability is not sufficiently high for the automation. The integration of laser scanner and CCD sensor may enable to automate three dimensional spatial data generation with high resolution.
Model fitting methods attract attention as a method of integrating various sensor data for object recognition. The model fitting methods usually prepare models which describe the characteristics of the object with several parameters to generate a hypothesis, and selects a model which maximizes the agreement to obtained multi-sensor data, and simultaneously decides the parameter values of the model. However, enough examination is not performed on the characteristics of various model fitting methods, though a variety of methods are proposed in terms of the definition of an objective function necessary for deciding the model parameter and those of constraint conditions.
In this research, the building extraction from the high-resolution satellite image was assumed to be an example, and the techniques of the model fitting were compared under different conditions in terms of accuracy and reliability, etc.
1 Introduction
Model fitting methods attract attention as a technique to extract an object from various data of the image etc. The model fitting methods usually apply models which describe the characteristics (e.g. shape, texture) of the object with several parameters to generate a hypothesis, and selects a model which maximizes the agreement to obtained multi-sensor data, and simultaneously decides the parameter values of the model. However, enough examination is not performed on their characteristics of various model fitting methods, though a variety of methods are proposed in terms of the definition of an objective function necessary for deciding the model parameter and those of constraint conditions.
The purpose of this paper is to compare and to characterize model fitting methods ay applying them to high-resolution satellite image data to extract building outline. Success rate of extracting the building outline is used as an indicator.
First of all, the building samples in an urban area are chosen in this image, and those samples are classified into two classes; buildings that can be extracted successfully and the other that fail to be extracted (Chapter 2). For the buildings that cannot be easily extracted, the initial model position/shape is changed so that the initial value dependency of the fitting results can be analyzed (Chapter 3). In addition, range of parameter values which make model fitting succeed is searched for by changing the parameter and forms of the energy function and the constraint conditions of SNAKE, and generality of the sensitibity analysis is examined.
1-1 Test data
In this research, high-resolution satellite image (IKONOS) and the existing residential maps (Zenrin Map; vector data) of Kobe City are used. This is a typical case with the urban area in Japan.
- Figure 1-1: IKONOS image(1999)-
- Figure 1-2: Zenrin Map-
1-2 SNAKE
Here, Polygonal SNAKE which one of the authors proposed based on [Fua 1996] is used. SNAKE is initialized at first by using the map data. Next, each line is moved in parallel to maximize the average of edge intensity along SNAKE lines. The edge intensity is defined as a cross-sectional gradient of the gray value of the edge-enhanced image. The following functions are used for SNAKE.
Einternal ; Paragraph that Internal energy is evaluated
Eexternal ; Paragraph that External energy is evaluated
w1 ; Weight of internal energy
w2 ; Weight of external energy
w1 =1.000000 and
w2=0.250000 are set from the experience as default value. Besides this, it is set as Scale Space=5 and Search Segment Number=5, and assumed default value. Scale space is distance between SNAKE's points. And Search Segment Number is range where SNAKE searchs optimum point. Basically, the fitting is done by these values. These parameter values are changed when the fitting fail with the default value, and the value which the fitting success is recorded.
1-3 Evaluation of fitting errors
Quantitative evaluation of the fitting errors was not made because the purpose of this study is to identify conditions affecting the success or failure of the model fitting. The success or failure is determined by visual inspection.
2 Experiment of model fitting (SNAKE) using existing map data as an initial value
Building in dense urban areas are chosen as samples, and the model fitting (SNAKE) using an existing map to set up initial models is conducted. IKONOS image was geocoded to overlay with the existing 1/2500 digital maps. Example results of the fitting experiment are shown in Figure 1-3~Figures 1-7. Factors greatly affecting the fitting success in the experiment are summarized as follows.
- Distance between initial location/shape of the model and edge location in the image
- Contrast or intensity of the edge
- Three dimensional complexities of buildings
