Building Recognition and Change InferenceBased on Very High Resolution Satellite Based on Very High Resolution Satellite Images Using Probabilistic Model
The present work was achieved in the context of a 863 project which aimed at providing robust methods and techniques for fast digital map updating in urban area. The specific objective of this Master thesis is to address the issues of building recognition from a single VHR image and building change inference from a pair of images. This thesis gives at first a general review and analysis of existing methods on object extraction and change retrieval for remote sensing applications (Chapter 2). It then presents in details the newly developed techniques and illustrates results on high resolution Quickbird image covering the area of Beijing. The major achievements and contributions can be summarized as follows: 1. Building contour recognition based on probabilistic model (Chapter 3) We define the building object by its probability density function --a logistic function-- and its data features. Building recognition is performed using a newly developed cut-and-merge algorithm. This original framework enables to alternately and iteratively detect building contour candidates, estimate their likelihood, then merge possible conflicting candidates such as to maximise their probability to be a building. The advantage of this algorithm is that any data feature can be fused in the model; the proposed method is not specific to the building and can be adapted to any object class. The application on a large data set demonstrates the robustness and performance of the technique. 2. Buiding change detection based on contour detection (Chapter 4) We propose a new approach for building change detection based on the building recognition algorithm presented in Chapter 3. The main idea is to compute the change's probability as a product of the probabilities of object detection and object non-detection calculated in each of the image pair. This approach does not require accurate image registration. We illustrate its efficiency on synthetic and real images. 3. Structural object change inference based on DRF model (Chapter 5) In order to avoid the preliminary stage of contour detection as required in Chapter 4, we introduce an extension of the existing DRF (Discrimative Random Field) model. The new DRF cope directly with the pair of images and model the change as a discontinuity in time and continuity in space. We performed application and show the potential of the method.
修改评论