Détection Des Changements À Partir de Photographies

Résumé du livre

This work deals with change detection from chronological series of photographs acquired from the ground. This context of consecutive images comparison is the one encountered in the field of integrated geography where photographic landscape observatories are widely used. These tools for analysis and decision-making consist of databases of photographic images obtained by strictly rephotographing the same scene at regular time intervals. With a large number of images, the human analysis is tedious and inaccurate. So a tool for automatically comparing pairs of landscape photographs in order to highlight changes would be a great help for exploiting photographic landscape observatories. Obviously, lighting variations, seasonality, time of day induce completely different images at the pixel level. Our goal is to design a system which would be robust to these insignificant changes and able to detect relevant changes of the scene. Numerous studies have been conducted on change detection from satellite images. But the utilization of classic digital cameras from the ground raise some specific problems like the limitation of the spectral band number and the strong variation of the depth in a same image which induces various appearance of the same object categories depending on their position in the scene. In the first part of our work, we investigate the track of automatic change detection. We propose a method lying on the registration and the over-segmentation of the images into superpixels. Then we describe each superpixel by its texture using texton histogram and its gray-level mean. A distance measure, such as Mahalanobis distance, allows to compare corresponding superpixels between two images acquired at different dates. We evaluate the performance of the proposed approach on images taken from the photographic landscape observatory produced during the construction of the French A89 highway. Among the image segmentation methods we have tested for superpixel extraction, our experiments show the relatively good behavior of Achanta segmentation method. The relevance of a change is strongly related to the intended application, we thus investigate a second track involving a user intervention. We propose an interactive change detection method based on a learning step. In order to detect changes between two images, the user designates with a selection tool some samples consisting of pixel sets in "changed" and "unchanged" areas. Each corresponding pixel pair, i.e., located at the same coordinates in the two images, is described by a 16-dimensional feature vector mainly calculated from the dissimilarity image. The latter is computed by measuring, for each corresponding pixel pair, the dissimilarity of the gray-levels of the neighbors of the two pixels. Samples selected by the user are used as learning data to train a classifier. Among the classification methods we have tried, experimental results indicate that random forests give the better results for the tested image series.