Fuzzy Point Clouds

[JEM3345]

How do you improve your point clouds pre and post processing. What flight settings, camera and gimbal settings, altitude, flight app, etc?

We use UAV photogrammetry point clouds to keep our survey crews out of busy roads that we need to profile, but often have to do a lot of filtering and cropping where points are generated above and below the road surface. I'm always trying improvements, but the processing time makes testing different methods very time consuming. Most recently I'm trying to see what difference using MTPs in the road (mostly end of paint strips) will make.

I know that oblique images provide more detail on vertical surfaces, but haven't been able to determine if it helps or hurts with overall point cloud sharpness on hard surfaces.

Any help or pointing me in direction of training/software you'd recommend would be greatly appreciated. I've been contemplating taking a Pix4d course, but not sure which one would be most beneficial. Mostly learning from online tutorials and videos.

 

[JEM3345]

I should also mention that I've been using Trimble business center to classify the ground region, and then sample it down a second or third time by re extracting the ground again, but it still sees potholes and piles where there are none so the only solution I've found is to manually crop those areas. I normally just hand off the processed point cloud to the project manager, but not all of them are familiar enough with working with point clouds. For now, my best option is just to create the points needed from the point cloud along the road crest and export those to project manager.

I generally fly at 200' where possible and fly separate flights up to the road on either side to get enough overlap.

 

[Bigriver]

Hey JEM3345. I am no expert, but I have spent quite a bit of time working toward the same goal. I’ll share what I have found and have learned from other experienced mappers on this forum.

Extracting reliable elevation data from uav mapping is not super easy. But it can be done with proper planning and effort. As far as time, it takes a lot of planning and processing time on my part to get good data. But if it keeps my guys out of traffic, I think it is worth the effort.

There are a few things that you need to consider. Elevation data is two to three times as unreliable as horizontal data from the point cloud.

Fly as low as you can. 200-250 feet is okay, But if you can drop down to 100-150 feet, that’s all the better. Trees and buildings often prevent this. You will also need plenty of GCP’s. My rule of thumb is to keep GCP spacing about twice the flight altitude. So if you fly at 150 feet, set GCP’s at 300 feet. Better yet, set them at 150 feet, and use half as check points to verify your model. Then reset the checkpoints to GCP’s to tighten up the model more. And make sure the vertical accuracy of your control points is tight.

If your flight plan is along (longitudinal) the roadway, I recommend a minimum of four paths. 80% frontal and 75% side overlap is good.

Also, do the best you can to fly in overcast conditions. This is the hardest to plan for, especially on projects far from home base. But it is critical for accurate mapping. Casting shadows wreak havoc on the elevations. Just look at a point cloud with heavy shadows. You can see the dips it causes in the point cloud. Sometimes you can clean up the point cloud by clipping out the shadowed areas. But this is tedious, and results are not great. An overcast condition provides a more homogenous lighting condition and better point cloud. Also, pay attention to lighting on your GCP’s. Are some in full sun and others in shadows? That can mess up the model. If the roadway is completely open, with no shadows, full sun conditions are okay. But it is rare, as usually something casts a shadow across the roadway.

Now I am not trying to push one processing program over another. But I used Pix4D exclusively for three years. Contouring seemed adequate. Until I tested Metashape. Metashape allows so much more control in producing contours and DEMs. Contours produced from Metashape are much more finely tuned. It also allows you to create contours and dem’s from smoothed meshes, which helps average and smooth out the elevations. You can vary the “smoothing” effect as much as you find necessary. But be ready for a difficult learning curve transitioning from P4D to Metashape. It is different in many aspects.

As for the elevation models, I started out importing contours into Microstation/Geopak to create a tin model. I have found that loses a lot of data in flatter areas. Exporting the DEM instead of contours gives you a much better triangulation model.

Another thought. If you are extracting profiles from edges of pavement, where a grade change (like drop from asphalt to shoulder) may be present. Cut your profile a foot or so into the pavement. Elevation differences here should be negligible, but it will help eliminate problems with the software trying to map across a grade change.

As far as mapping oblique vs. nadir, I commonly shoot at 75-80 degrees to help with getting images with more info (like sign faces), but as far as accuracy, I have not seen a difference in the point cloud.

Finally, I would recommend testing this on a section of roadway that has been surveyed conventionally. Extract elevations from your DEM at locations that were shot conventionally. Compare the elevation differences. Don’t expect all elevations to be 0.04 or 0.06 feet different. But overall, you should see a RMS of about the same. Is a tenth of a foot adequate? Maybe 0.15 feet? I think on most of our projects that some variation in this range would be fine. Cut a profile along your conventionally surveyed points, and compare it to a profile cut along you uav mapped model. Prove it to yourself before accepting the results of your mapping unverified.

Hope this helps. Try different things. Verify results with field data. And keep us posted if you discover ways to improve accuracy of your mapping. We are always interested in improving our results.

 

[JEM3345]

Hey JEM3345. I am no expert, but I have spent quite a bit of time working toward the same goal. I’ll share what I have found and have learned from other experienced mappers on this forum.

Extracting reliable elevation data from uav mapping is not super easy. But it can be done with proper planning and effort. As far as time, it takes a lot of planning and processing time on my part to get good data. But if it keeps my guys out of traffic, I think it is worth the effort.

There are a few things that you need to consider. Elevation data is two to three times as unreliable as horizontal data from the point cloud.

Fly as low as you can. 200-250 feet is okay, But if you can drop down to 100-150 feet, that’s all the better. Trees and buildings often prevent this. You will also need plenty of GCP’s. My rule of thumb is to keep GCP spacing about twice the flight altitude. So if you fly at 150 feet, set GCP’s at 300 feet. Better yet, set them at 150 feet, and use half as check points to verify your model. Then reset the checkpoints to GCP’s to tighten up the model more. And make sure the vertical accuracy of your control points is tight.

If your flight plan is along (longitudinal) the roadway, I recommend a minimum of four paths. 80% frontal and 75% side overlap is good.

Also, do the best you can to fly in overcast conditions. This is the hardest to plan for, especially on projects far from home base. But it is critical for accurate mapping. Casting shadows wreak havoc on the elevations. Just look at a point cloud with heavy shadows. You can see the dips it causes in the point cloud. Sometimes you can clean up the point cloud by clipping out the shadowed areas. But this is tedious, and results are not great. An overcast condition provides a more homogenous lighting condition and better point cloud. Also, pay attention to lighting on your GCP’s. Are some in full sun and others in shadows? That can mess up the model. If the roadway is completely open, with no shadows, full sun conditions are okay. But it is rare, as usually something casts a shadow across the roadway.

Now I am not trying to push one processing program over another. But I used Pix4D exclusively for three years. Contouring seemed adequate. Until I tested Metashape. Metashape allows so much more control in producing contours and DEMs. Contours produced from Metashape are much more finely tuned. It also allows you to create contours and dem’s from smoothed meshes, which helps average and smooth out the elevations. You can vary the “smoothing” effect as much as you find necessary. But be ready for a difficult learning curve transitioning from P4D to Metashape. It is different in many aspects.

As for the elevation models, I started out importing contours into Microstation/Geopak to create a tin model. I have found that loses a lot of data in flatter areas. Exporting the DEM instead of contours gives you a much better triangulation model.

Another thought. If you are extracting profiles from edges of pavement, where a grade change (like drop from asphalt to shoulder) may be present. Cut your profile a foot or so into the pavement. Elevation differences here should be negligible, but it will help eliminate problems with the software trying to map across a grade change.

As far as mapping oblique vs. nadir, I commonly shoot at 75-80 degrees to help with getting images with more info (like sign faces), but as far as accuracy, I have not seen a difference in the point cloud.

Finally, I would recommend testing this on a section of roadway that has been surveyed conventionally. Extract elevations from your DEM at locations that were shot conventionally. Compare the elevation differences. Don’t expect all elevations to be 0.04 or 0.06 feet different. But overall, you should see a RMS of about the same. Is a tenth of a foot adequate? Maybe 0.15 feet? I think on most of our projects that some variation in this range would be fine. Cut a profile along your conventionally surveyed points, and compare it to a profile cut along you uav mapped model. Prove it to yourself before accepting the results of your mapping unverified.

Hope this helps. Try different things. Verify results with field data. And keep us posted if you discover ways to improve accuracy of your mapping. We are always interested in improving our results.

Thanks for taking the time to give such a detailed reply. Many of the things you mention as causing problems with elevation I have been suspecting. I have more time now to test the accuracy of new methods different variables with the virus slow down.

One immediate take away is that I'll start trying to fly as low as possible for 3d accuracy. 160' was kind of my default AGL, but I have crept up to 200 and higher because it often allows me to set up my flight lines (I use Drone Deploy's flight app) parallel with most of roads to avoid flying over moving traffic. (My other trick is to set flight lines perpendicular to a road if it's not too busy and just pause the flight when there is traffic.) 200' and higher works just fine for most sites if all I'm going for is the orthoimage.

Do you find that wet pavement and oily patches in the pavement can cause false low areas? Do you correct this using MTPs?

I'll be sure to look into Metashape. The only other processing software I have had any exposure to besides Pix4d is 3DSurvey. I only had a 2 week trial version, and was not experienced enough to really compare results. I went to a workshop with the developers of the software (from Croatia, I think) and from what I remember their point clouds were very crisp. Also, they said they typically fly 200', and I think that's when I started flying at that elevation. Also the software developers were surveyors by trade.

Thanks again, and I'll be happy to share anything I learn.

 

[Bigriver]

I'd definitely avoid trying to map wet pavement, as it would cause all kinds of problems I would think. Oily or dark patches can affect vertical accuracy too, like the low areas you mention. Correcting with MTPs may work, but would be very tedious. But if you are able to fly your missions during overcast conditions, the effects should be small. I can't overemphasize that enough. Either fly in overcast, or all sunlight with no shadows. It just makes the elevations more uniform.

On projects that it is practical, I will fly in both sun and in overcast conditions. I use the overcast mission to help extract good ground elevations. And I use the sunlight missions to help isolate locations of signs, poles, etc. by using their shadows to help identify their true location. Having both models is double the effort (except for using the same GCPs), but it has helped me considerably on many projects. A pole or sign that is invisible on an ortho created from an overcast mission usually stands right out in a ortho created from a sunlight mission, just by the shadow.

Flying lower is helpful, but there is the trade-off of a larger processing set. So its all about balancing what will work for the project size. I often have to break large jobs into sections to make the processing time reasonable.