This work studies how to explore computer vision techniques to detect urban changes from two high spatial resolution remote-sensing images. The main contributions are summarized as follows: (1) Image registration is usually performed by a local-distortion method in most existing change detection algorithms. This work analyses the limitations of this approach and proposes to register images according to the ground plane by a global transform. This registration scheme preserves the structural feature of urban area and the multi-view geometrical constraints, which are used to extract real urban changes in this work. (2) The imaging process is inevitably coupled by information lost. This work studies the effects of this phenomenon on change detection and proposes the concept of Change Blindness Region (CBR). CBRs are image regions where change detection is not supposed to be performed, because the bi-temporal information losing processes are different for these regions. This work explores the details for CBR, their different types and the methods to detect CBRs. In experiments, a change detection model with CBR can effectively remove the false changes induced by disparity, occlusions and ground shadows of tall buildings. (3) For a real city with complex contents, this work proposes to construct and use different change detection models for different types of urban areas. Change detection models for several types of urban areas are studied and a novel change detection model is proposed for urban building areas. This model first extracts unchanged urban regions and CBRs, and then in the rest of the image detects urban changes by extracting clustered changed line segments. In experiments with real data and CG simulation data, this model still works well even the view angle and illumination undergoes large variations. (4) A multi-model based urban change detection system is proposed. This system extracts urban changes with different models in different types of urban areas. Experiments of Ikonos, Quickbird and aerial images show the effectiveness of this method.
修改评论