A Robust Registration Algorithm for Point Clouds From Uav Images for Change Detection

Abstract

Commission I/Vb, ICWG I/Vb KEY WORDS: landslide dynamics, normal distance, bundle block adjustment with self-calibration, structure from motion, 3D dense surface reconstruction, Unmanned Aerial Vehicle (UAV) ABSTRACT:Landslides are among the major threats to urban landscape and manmade infrastructure. They often cause economic losses, property damages, and loss of lives. Temporal monitoring data of landslides from different epochs empowers the evaluation of landslide progression. Alignment of overlapping surfaces from two or more epochs is crucial for the proper analysis of landslide dynamics. The traditional methods for point-cloud-based landslide monitoring rely on using a variation of the Iterative Closest Point (ICP) registration procedure to align any reconstructed surfaces from different epochs to a common reference frame. However, sometimes the ICP-based registration can fail or may not provide sufficient accuracy. For example, point clouds from different epochs might fit to local minima due to lack of geometrical variability within the data. Also, manual interaction is required to exclude any non-stable areas from the registration process. In this paper, a robust image-based registration method is introduced for the simultaneous evaluation of all registration parameters. This includes the Interior Orientation Parameters (IOPs) of the camera and the Exterior Orientation Parameters (EOPs) of the involved images from all available observation epochs via a bundle block adjustment with selfcalibration. Next, a semi-global dense matching technique is implemented to generate dense 3D point clouds for each epoch using the images captured in a particular epoch separately. The normal distances between any two consecutive point clouds can then be readily computed, because the point clouds are already effectively co-registered. A low-cost DJI Phantom II Unmanned Aerial Vehicle (UAV) was customised and used in this research for temporal data collection over an active soil creep area in Lethbridge, Alberta, Canada. The customisation included adding a GPS logger and a Large-Field-Of-View (LFOV) action camera which facilitated capturing high-resolution geo-tagged images in two epochs over the period of one year (i.e., May 2014 and May 2015). Note that due to the coarse accuracy of the on-board GPS receiver (e.g., +/-5-10 m) the geo-tagged positions of the images were only used as initial values in the bundle block adjustment. Normal distances, signifying detected changes, varying from 20 cm to 4 m were identified between the two epochs. The accuracy of the co-registered surfaces was estimated by comparing non-active patches within the monitored area of interest. Since these non-active sub-areas are stationary, the computed normal distances should theoretically be close to zero. The quality control of the registration results showed that the average normal distance was approximately 4 cm, which is within the noise level of the reconstructed surfaces.

Automated co‐registration and calibration in SfM photogrammetry for landslide change detection

“…Least squares surface matching has been applied to glacier monitoring (Fieber et al, ) and airborne photogrammetric studies (Gneeniss et al, ). The ICP algorithm, meanwhile, has been implemented with TLS observations (Wujanz et al, ) and SfM‐MVS point clouds derived from UAV imagery (Al‐Rawabdeh et al, ). A major limitation of both approaches is the requirement for a good starting approximation to the transformation parameters between control and matching datasets.…”

“…For instance, Al‐Rawabdeh et al . () refined the coarse orientation from DG with consumer‐grade UAV, with the inclusion of GCPs into the typical SfM‐MVS pipeline prior to ICP application. Wujanz et al .…”

Automated Registration of SfM‐MVS Multitemporal Datasets Using Terrestrial and Oblique Aerial Images

“…Depending on the application, a range of strategies are generally adopted to minimise errors in the similarity transformation (rotation, translation and scaling) applied to the multitemporal 3D models (for example, Zang et al, 2019; Cucchiaro et al, 2020). A detailed description of registration approaches is provided by Al‐Rawabdeh et al (2016).…”

“…Another option could be direct georeferencing, which involves the measurement of camera position/orientation to georeference the generated 3D models (Turner et al, 2014; Armstrong et al, 2018). Al‐Rawabdeh et al (2016) proposed a robust registration option for UAV imagery that relies on the knowledge of the coarse image position at each exposure. However, for terrestrial applications, measuring ground‐based camera positions/orientations repeatedly over time would be time‐consuming and hence costly.…”

“…This innovative 3D registration method was first identified as “Time‐SIFT” by Feurer and Vinatier (2018) who demonstrated its potential for producing DEMs of difference (DoD) from multitemporal aerial archive images, suitable for landscape change analysis. Other analogous recent studies have highlighted the potential of such an approach to co‐register multitemporal images captured from UAVs (Al‐Rawabdeh et al, 2016; Cook and Dietze, 2019) and terrestrial camera stations (Filhol et al, 2019). However, this registration approach was not adopted to investigate the ability to register multitemporal epochs captured from both terrestrial and UAV‐based cameras into a common reference system.…”