Point cloud generation
This step is necessary for generating height information from the UAV images after stitching. Point clouds generated from UAV images are essentially elevation data (i.e. each point has an associated elevation). High density point clouds are automatically generated through image matching (Rosnell and Honkavaara, 2012). The success of image matching, and subsequent generation of the dense point could, depends on factors such as visual content and overlap. Large overlaps and high visual content will lead to better results than smaller overlaps and low visual content. Once image stitching has been completed, dense point cloud can be generated. Once image stitching has been completed, dense point cloud can be generated (Figures 5.10 and 5.11).
- Figure 5.11: Point cloud and super-imposed camera positions in the Pix4D software package (source: STARS AgriSense team).
The generation of dense point cloud is perhaps the most computationally intensive step in processing UAV images. One needs a computer with high specification in order to process a large number of images.
Example, for an area of 0.47 km2, and 97 photos, generation of a dense point cloud took approximately 45 min for a computer with specifications CPU: Intel(R) Core(TM) i5-3340MCPU @2.70GHz, RAM: 16GB, GPU: Intel(R) HD Graphics 4000 (Driver: 220.127.116.1196). In desktop computers CUDA GPU’s devices are advisable to improve the processing power of the PC.
In case you do not have a high specs computer, it may be advisable to perform this task overnight while remotely monitoring the progress (e.g. monitoring your workstation at work with your laptop at home). Performing this overnight ensures that all the computer resources are dedicated to this operation and therefore lower chances of failure due to low computer memory. You have to ensure that there is continuous power supply, though!
Cleaning the generated point cloud is important to improve the resulting orthoimage, especially in areas of drastic changes in height, such as fences, trees and buildings. This can be achieved manually when an analyst observes the generated point cloud and delete points which are outliers to the terrain shape. Low overlap areas outside the planned area need more cleaning work to get a good DSM. For easier viewing and editing of the point cloud, a 3D mesh can be generated in the same processing software (e.g. in Pix4D) (figure 5.11). This is a surface which is formed by roughly connecting the points in the generated cloud. After this cleaning, the project can be re-matched and improved.