Geometric Change Detection and Image Registration in Large Scale Urban Environments
Aparna Taneja · Repository for Publications and Research Data (ETH Zurich) · 2014
3D reconstruction of urban environments has gained a lot of interest in the recent past due to the extensive applicability of 3D models to city planning and monitoring, navigation applications, to name a few. However, cities change with time, with new constructions being built and old taken down. Therefore, to maintain the accuracy of these models, there is a need to update them as changes take place in the environment. Instead of rebuilding the entire model from scratch every time, an efficient technique is proposed to update only those parts of the model which may have undergone a structural change. To this end, an algorithm was devised to detect changes in the geometry of an urban environment using images observing its current state. This ensures that a very simple setup is needed to detect changes on a frequent basis, contrary to the big setups required for 3D acquisition. Keeping the final application in mind, the algorithm was specifically designed to detect only structural changes in the environment, ignoring any changes in its appearance, and ignoring also all the changes which are not relevant for update purposes, such as cars, people etc. Moreover, the algorithm was extended to further deal with the additional challenges introduced with the usage of large scale data, i.e. a city scale image database and 3D model. Images were captured by cars driving around the city, with sensors like GPS and IMU recording their geo-location data. Since this data is typically noisy, an algorithm is also presented to refine this data exploiting the available 3D information. In addition, another method was explored to localize images inside an environment, removing the assumption of the availability of 3D models. In fact, instead of a 3D model, the Google streetview database was used as a reference. In particular, both the streetview images and the graph connecting the locations of these images was exploited, to continuously localize a driving vehicle inside this discrete graph, by comparing them with images captured by a phone mounted on the vehicle.