Research On Spatial Information Extraction And Error Analysis From Single Close Range Image

With the development of the information technique and the popularity of image acquisition equipments, the amount of image are growing rapidly. If these images, especially for the images with spatial geographic information are used properly, the data sources of geographic information will be increased significantly. It is generally known that the image reflects the real-world scenario, which not only includes abundant superficial information, e.g., spectrum, texture and geometry but also contains the in-depth spatial information, such as size and depth of various objects within images.Spatial information extraction from the images has significant application values in the fields of image-based measure, target identification, robot navigation and emergency relief. Additionally, in some occasion, e.g., historical scene or accident scene, there exist only some limited uncalibrated images, thus the approach for extracting spatial information from a single uncalibrated image has become the only choice. In this context, this thesis will mainly focus on three aspects:the distance measurement, image scene depth recovery and the corresponding accuracy evaluation, respectively.For geometrical measurement, this thesis first regards the rectangles as the given conditions and the approach based on cross-ratio principle are employed to obtain the geometrical measurement results and the corresponding accuracy evaluation is performed. In addition, the error correction model is built to improve the measurement accuracy. Finally, the error distribution has been predicted. The major contributions are described as follows.(1) Realizing cross-ratio and cross-plane based single image measurement. This thesis study the approach cross-ratio based single image geometrical measurement, in which collinear distance measurement and coplanar distance measurement models are extended into cross-plane distance measurement model, thus the number of the given conditions are reduced dramatically.(2) Analyzing the sources of the geometrical measurement error and proposing the corresponding error correction methods. This thesis sufficiently analyzes the sources of geometrical measurement error (e.g., lens distortion, vanishing points calculation and error propagation, pixel pickup deviation, the number of the given conditions, and the distribution of the given conditions) and the appropriate solutions to minimize the error. The conclusions are illustrated as follows.1) The measurement error of unknown line segment has closely related to the distance apart from the given conditions;2) When the number of given conditions are over three, the mean of measurement results can significantly improve the measurement accuracy;3) No significant relationship between the distribution of known conditions and the measurement error.(3) Establishing the error correction and prediction model and realizing the quantitative description of measurement error. The single image geometrical measurement error correction model has been built using the relation between measurement error and distance. Thus, the measurement results are improved significantly and the corresponding error quantitative description are given as well.The geometrical measurement of single image has only recovered the measure information. However, the depth knowledge contained in single image is indispensable parts of the spatial information. Therefore, the traditional methods of image depth calculation are improved and the evaluation accuracy of the depth results are given simultaneously, which are rarely involved in traditional recovery methods.The major contributions are described as follows.(1) Improving the image depth calculation model and achieving depth calculation or estimation for each pixel by interactive processing. In this thesis, the depth is calculated or estimated using the different solutions or criteria, i.e., image segmentation results and spatial relations among different objects. The proposed image depth calculation method is a significantly development of the traditional image calculation depth model, i.e., the ideal pin-hole imaging model.(2) Proposing the height of camera recovery method based on depth difference. In this thesis, we found that the camera height difference is the most important coefficients which affect image depth calculation accuracy. In light of the finding, the method combing ground pixels and depth difference is proposed, which can improve the reliability of calculation results dramatically.