Construction of a point cloud data set, true orthomosaic, and digital surface model using Pix4D software

Part 1 (Introduction):

What is Pix4D?
Pix4D put simply is computer software that can find tie points, or common points between many images.

What products does it generate?
Pix4D can create orthomosaics, digital surface models, and 3-dimensional models such as point clouds or meshes.

Part 2 (Methods):

What is the overlap needed for Pix4D to process imagery?
Pix4D recommends at least 75% overlap in the flight direction and at least 60% overlap from side to side.

What if the user is flying over sand/snow, or uniform fields?
Pix4D recommends using at least 85% overlap in flight direction and at least 70% side overlap due to uniformity.

What is Rapid Check?
Rapid Check allows the software to create tie points quickly for fast results, as opposed to the higher quality results you would get through normal processing, which takes more time.

Can Pix4D process multiple flights? What does the pilot need to maintain if so?
Pix4D can process multiple flights, although the pilot needs to maintain enough overlap between the two flights to ensure there will be enough tie points.

Can Pix4D process oblique images? What type of data do you need if so?
For oblique images, you really should have ground control points or manual tie points between each set of images, as well as enough overlap between the sets of images.

What is the difference between a global and linear rolling shutter?
A global shutter takes the image instantly, whereas a linear rolling shutter does a scan to capture the image.

Are GCPs necessary for Pix4D? When are they highly recommended?
GCPs are not necessary for Pix4D. They are highly recommended when combining nadir images with aerial oblique and terrestrial images. GCPs are also highly recommended if capturing a tunnel system and a multiple tracks of images are not possible.

What is the quality report?
The quality report gives information as to what the software did. For example, it would show you how many images in the set were used, the average number of tie points per image, previews of the results, details about the block adjustment, and various other processing details. Attached below is the quality report associated with this dataset for use as an example.
Note: viewing this document will open the pdf, located on Google Drive, in a separate tab.
Quality Report.pdf

Methods:
In this project, we used Pix4D to process a set of images taken over someone's residence. Ground control points were not used during this project. We started out by creating a project within Pix4D, and then uploading our images to the software. After confirming camera settings, Pix4D began the task of processing our data. The total time for the initial processing was 31 minutes and 11 seconds. After initial processing, Pix4D used 66 out of 67 images.

After the initial processing, we confirmed the data was looking like it should, and began the orthomosaic generation as well as the digital surface model and the 3D models. Upon completion, we had an orthomosaic, digital surface model, as well as 3D files to view and work with.

Part 3 (Results and Discussion):

Below is a table that shows processing times for various segments. Each time is in minutes, and then seconds. Total processing time took approximately 65 minutes to complete. This data is also located in the quality report attached above.

Orthomosaic	15:06
Digital Surface Model	0:43
Point Cloud	27:50
3D Textured Mesh	4:15
Total	65:00

Using the data from Pix4D, a couple maps were created using ArcGIS software. In Figure 15 below, the orthomosaic was used to create a map. One thing to note is the edges of the orthomosaic. Because of the limited overlap around the edges, as well as because of the block adjustment, the edges have abrupt turns and incomplete photos. This is not a concern, just an interesting aspect of the orthorectification process. One interesting aspect about this map though is the trees. Pix4D can process some trees just fine, but struggles with other trees. Along the road are some coniferous trees, which Pix4D was able to process perfectly fine. Behind the house are a bunch of deciduous trees, which upon closer inspection, have some lines where it was obvious they were pieced together, as well as some blurred areas where either the images were not very clear, or where the images were blended in processing. A couple deciduous trees turned out fine, such as the red-orange tree near the road in the center of the orthomosaic. Most likely due to its isolation, this tree was precise, and each individual leaf on the ground was visible and distinguishable.

Figure 15: Pix4D orthomosaic

Moving on to the digital surface model and 3D model, I notice some interesting characteristics straight off. First, there seems to be an incongruity in the digital surface model, near the area to the north of the house. A hillshade effect was added to the DSM to easier identify areas of relief. In the center of the DSM lies an orange circle, which appears to mean that the land the house resides on rises, I am assuming so that water runs away from the house. Near the top and left of that orange circle, though, is a sharp change in the DSM, which I believe to be an error in processing, or an issue with the images, because I see no reason that the land would abruptly change in elevation at that point. This could be related to trees in that area, as trees have been known to cause some 3D and orthomosaicking issues.

It is interesting to see other topographic features of the area though, such as the trees along the road, each represented by a bright red dot, which means that the surface the camera sees is higher than the surrounding terrain. Each tree along the road is very distinguishable. One thing I would like to point out is the lack in elevation change where that red-orange tree should be. Following the curve of the road between the lines of trees, there should be another red dot near the center of the image, but there is not. This is also evidenced in an animation that was created using the Pix4D software, which I will point out momentarily, but I believe this is because of the issues Pix4D can have with trees.

Figure 16: Pix4D DSM and 3D Model

Below I have included an animation that was created using Pix4D. Figure 17 shows the trail that the camera followed in creating this animation, but the short clip below highlights another feature of this software.

Figure 17: Animation Trail

As I mentioned before, though it may be hard to detect on the first play through, the area on the ground where that red-orange tree should be is completely flat. Pix4D seems to have completely edited out or flattened that surface, I am assuming because it could not figure out what to do there, or the data was not clear enough that an elevation or surface change was present in that location. Either way, this is one issue that is frequent with orthorectification. 

Part 4 (Conclusions):

Pix4D is important for processing UAS data due to its capabilities and benefits. With a few clicks and some cash, any user can take data capture from an unmanned system and with the right tools, can create accurate maps of a feature, digital surface models, 3D point clouds, or even animations to showcase at their next meeting. This powerful software does have its drawback, some mentioned earlier with the issues of trees, but others being required time and processing power. This is no small task creating and processing this data. As given above, this data alone, with only 67 images to process, took all of 65 minutes. A project or data set with several hundred images will take significantly more time. Not only that, it takes a powerful machine to achieve these processing times. The machine this data was processed on contained an Intel i7-6700T and 32 gigabytes of RAM. This data processing is very heavy on the CPU, and without a powerful enough CPU, could cause the program to crash, or take an indeterminable amount of time, which is definitely something to consider for UAS data processing.