ارزيابي راهكارهاي توجيه هندسي تصاوير متوالي اخذ شده از محيط شهري

Assessment of Strategies for Pose Estimation of Image Sequence Captured From Urban Environment

3D modeling of urban environment is growing for wide variety of applications related to urban management. Images can provide visual interpretation in addition to geometric information, so 3D modeling from images is desirable. This paper studies pose estimation of image sequence captured from urban environment which is major part of 3D modeling. Process of pose estimation consists of feature extraction and tracking, outlier detection, key-frame selection and bundle adjustment. In this research, a video is captured by an amateur camera from buildingʹs facade and the goal is to estimate pose parameters (exterior orientation) of the camera in each video frame. Urban environment is considered as complex scene due to moving objects, greens, repetitious features and scene incidental events, therefore, robust algorithms have to be used. First, feature selection is done using Harris operator and features of green areas are filtered out using hue values of surrounding window. Then, pyramidal KLT color feature tracker is used in short and wide base line fashions to investigate their behavior. To detect outliers, RANSAC in conjunction with common relative orientation algorithms (Eight point and Nister five point) and space resection is used to compare their performance. Not having overlap, bulk bundle adjustment is not directly possible in this type of problem and sequential bundle adjustment has to be used to estimate approximate camera poses. Then, global bundle adjustment is performed to refine estimated parameters. According to final accuracy of global bundle adjustment as a criterion, tracking in short base line and outlier detection by RANSAC and space resection shows relatively better performance in the sequence.