A new way to decrease the distortion of the edges of buildings on true orthophotos

Message:
Article Type:
Research/Original Article (دارای رتبه معتبر)
Abstract:
Introduction

On true orthophotos, there are some distortions on the structural edges of buildings, which is due to defects in these areas in the point cloud used in the digital surface model. This problem is greater for orthophotos that have been made from UAV images in urban areas because of their lower altitude. Before interpolation of the point cloud and preparation of the digital surface model and then preparation of orthophotos of it, it is necessary to complete the point cloud in areas with defects. Some studies have shown that adding edge points has the effect of decreasing the distortion of true orthophotos. In this study, a new method for completing point clouds using a trained deep learning network is proposed, which includes steps: 1) Preparation and normalization of point cloud data, 2) completion of the point cloud by learned networks; 3) reversion of the completed point cloud to real-world coordinates and, 4) integration with the existing original point cloud and preparation of the digital surface model and generation of true orthophotos.

 Materials & Methods

In this study, the imaging of the Yazd region was done with a Phantom 4 drone equipped with a DJI camera. The SfM algorithm has been used to calibrate the camera, estimate the internal and external camera parameters, and produce images without distortion and low-density point clouds, and SGM has been used to produce dense point clouds. In the proposed method, the trained SnowflakeNet network is used to complete the incomplete roof points of the building. Assuming that the points on the roof of each building are predetermined, without noise, and have incomplete edges, these point clouds were introduced as inputs to the network to complete. Points related to edge points were extracted for each roof and added to the existing point cloud after increasing the density and returning to the actual coordinates. The final point cloud was used in the preparation of digital models to produce irregular and then regular surfaces and in the preparation of true orthophotos using camera parameters and undistorted images. One of the images with buildings marked as numbers 1 to 4 was selected to perform tests and prepare orthophotos.

Results & Discussion

The lack of structural edge points on any roof, which is the distance between severe height differences between levels, causes the greatest amount of distortion on the edge of the roof and around it. Adding these points with edge line recognition and reconstruction algorithms to the point cloud improves the resulting digital surface model. Since the quality and accuracy of the digital elevation model directly affects the resulting orthophoto, using a more accurate digital elevation model improves these images. In the proposed method, these point clouds are complemented by the deep learning method, and quantitative and qualitative comparisons show better results in reducing distortion in most of the buildings tested. The reasons for the superiority of the proposed method over previous methods include determining and calculating a more complete and integrated form of the roof of each building instead of multiple line segments and considering the outermost edges of the buildings.

Conclusion

In this study, a new method was introduced to improve the quality of true orthophoto edges by using a deep learning network to complete the point cloud, which was tested on several building images and compared with the results of previous methods. In this study, in addition to the fact that, for the first time, a deep learning network was used to improve point clouds to produce orthophotos, Compared to the previous method, the amount of distortion on the selected edge of four buildings has been significantly reduced and the success of the results with the latest proposed method of true orthophoto enhancement indicates an improvement of about 62% and 55% in the distortion decreasing of the structural edges and maintaining their coordinate accuracy. Despite the reduction of distortion on the selected structural edge using the proposed method, this value is increasing in curved areas as well as the corners of the roofs due to the type of network training and network output error. However, this can be reduced by improving the structure of the deep learning network and increasing the training data to a variety of roof modes with curved walls.

Language:
Persian
Published:
Journal of of Geographical Data (SEPEHR), Volume:31 Issue: 123, 2022
Pages:
89 to 106
magiran.com/p2516486  
دانلود و مطالعه متن این مقاله با یکی از روشهای زیر امکان پذیر است:
اشتراک شخصی
با عضویت و پرداخت آنلاین حق اشتراک یک‌ساله به مبلغ 1,390,000ريال می‌توانید 70 عنوان مطلب دانلود کنید!
اشتراک سازمانی
به کتابخانه دانشگاه یا محل کار خود پیشنهاد کنید تا اشتراک سازمانی این پایگاه را برای دسترسی نامحدود همه کاربران به متن مطالب تهیه نمایند!
توجه!
  • حق عضویت دریافتی صرف حمایت از نشریات عضو و نگهداری، تکمیل و توسعه مگیران می‌شود.
  • پرداخت حق اشتراک و دانلود مقالات اجازه بازنشر آن در سایر رسانه‌های چاپی و دیجیتال را به کاربر نمی‌دهد.
In order to view content subscription is required

Personal subscription
Subscribe magiran.com for 70 € euros via PayPal and download 70 articles during a year.
Organization subscription
Please contact us to subscribe your university or library for unlimited access!