Résumé:
High-quality 3D city models serve as fundamental infrastructure for smartcities and various applications. However, the presence of moving objects, especiallyVehicles, poses a significant challenge to the automated generation of these models. Moving targets introduce instability in density matching and aerial triangulationprocesses, which can adversely affect the overall quality of the final models.
To address this challenge and faithfully represent the dynamic environment ofcities using discrete still captures, we propose a pre-processing procedure for opticalimagery. This procedure focuses on detecting problematic objects, specificallyvehicles, to pass them later to the elimination phase and to ensures the generation ofaccurate and precise 3D city models without the inconveniences and distortionscaused by these objects.
This research contributes to the fields of Image Processing and ComputerVision by addressing the Object Detection key aspect. We design a modern, fresh andmore flexible Deep Learning-based method to detect moving vehicles that maydisrupt stereo-vision during the 3D extrusion process. The detection models showed apromising result of 98% mAP and 94% mAP, this may not seem impressive but theflexibility and new features of these models make up for the relatively averageaccuracy.
By mitigating the impact of moving vehicles on the 3D city generationprocess, our approach enhances the overall accuracy and realism of the resultingmodels.