Problem: Over the last 20 years, image based object and scene digitizing systems have evolved significantly. It is currently possible to digitize the shape of a wide range of objects with sub-millimeter accuracy and resolution. This type of vision-based technology has significantly contributed to various fields such as industrial inspection, prototyping, archaeology, design and arts, etc.
Currently, the predominant challenges faced by designers consist in improving the flexibility of the digitizing systems. This challenge raises fundamental problems in the field of computer vision. A flexible system should be self-referenced based on the content of the images. Such a system deals concurrently with capturing and integrating hundreds and even thousands of images while providing precise position. This suggests an automatic auto-positioning method coupled with a diagnosis system of the positioning quality.
|
Approach: The camera position is recovered from image keypoints matched between different views. In order to recover an accurate position, it is essential that corresponding keypoints also correspond to the exact same physical point in the scene. After an analysis of the bias affecting existing feature detectors, we will propose a detector able to minimize their effect. After analyzing the detection of features stable to view point, their matching between images will be studied.
Feature matching is usually performed by comparing small image patches in the neighborhood of the features. To be robust to deformation of the image content under perspective, a descriptor vector is extracted from the local texture. Such descriptors consist in histograms of intensity levels for example. For the vector to have discriminant power, the region must contain sufficient information. Feature detectors will trigger on image region centers [SURF, SIFT], or corner-junction points [Harris, Forstner]. Region detectors identify keypoints in image regions rich in details, allowing very distinct descriptors to be extracted. However, the position of the keypoints is biased by view point changes. Inversely, corner-junction detectors identify points whose neighborhood contains very little information. However, the localization of such points is very accurate and stable to view point. We believe there is no compromise to be made between obtaining keypoints associated with discriminant local descriptors and accurate keypoints. The complementarity of these two types of detectors will be studied.
During acquisition, the analysis of the 3D positions of detected points will allow the point reliability to be assessed and a diagnostic of the system position accuracy. In addition, a multi-view appearance model of the points will be constructed to allow the system to recognize points already observed as they return in the camera field of view. The recognition of previously observed points is important to prevent the error accumulation in the camera position.
|