Hi All,
I've been trying to decide how one would describe the accuracy of features extracted from pointcloud data. It is quite easy to summarize the accuracy of the pointcloud itself based on registration results, checks to known control etc. However once we start extracting data from that pointcloud does it still maintain that same level of accuracy? Surely interpretation of pick point locations either automated or manual will cause a deviation.
I have recently encountered two datasets in the same area where we had scanned using TLS and extracted topographic features from the cloud. (Kerbs, breaklines, walls and fences). By coincidence we had conventional total station pickup in the same area off the same control network. In this area I have seen differences in features of between 40-180mm. Is this the difference between 'relative' and 'absolute'?
Does anyone out there quantify the accuracy of their extracted data or provide some form of disclaimer. I understand that pointclouds in industrial plant environments will be quiet different. I guess I'm aiming this at people who work in construction, roads, rail and infrastructure.
Thanks
Accuracy statement of features extracted from point clouds
-
- I have made <0 posts
- Posts: 2
- Joined: Mon Apr 24, 2017 6:50 am
- 6
- Full Name: Hugh
- Company Details: Aurecon
- Company Position Title: Surveyor
- Country: Australia
- Linkedin Profile: Yes
-
- V.I.P Member
- Posts: 158
- Joined: Fri Nov 13, 2009 10:12 am
- 14
- Full Name: simon judd
- Company Details: Precision Air Surveys Limited
- Company Position Title: Owner
- Country: United Kingdom
- Been thanked: 3 times
Re: Accuracy statement of features extracted from point clouds
Hugh
As a surveyor you are providing your client with information that represents the survey area requested and that it should conform to some form of survey accuracy provided in a specification. This allows the client to have some form of confidenc when using the information provided.
If you are issuing a topographic plan to a client who has specified an accuracy level, 1:200, for example, then the information contained within that plan should stand up to the normal 1 sigma and 2 sigma accuracy levels for that designated scale, regardless of the source data.
Therefore, a survey undertaken by TLS with features extracted from the registered point cloud and the same survey undertaken by conventional Total stations (using the same survey control network) should both conform to the plan accuracy for the scale as defined in the project specification.
The point cloud at the end of the day is still your raw observation data set, just a great deal richer and just because it has been registered it does not mean that it is correct, just as any data set, redundancy and checking are vital to have a set of observations that you can rely on to give your client what it is that they need.
One of the big problems that may arise is that you only provide a pointcloud to the client and let them derive the information. You can warrant that the pointcloud is suitable to derive the necessary features at the required plan accuracy, But and it is a BIG But - you cannot warrant what the client derives as you will have no knowledge of their capabilities to undertake that process, without then checking what it is that they have extracted.
I hope that helps in some way.
Simon
As a surveyor you are providing your client with information that represents the survey area requested and that it should conform to some form of survey accuracy provided in a specification. This allows the client to have some form of confidenc when using the information provided.
If you are issuing a topographic plan to a client who has specified an accuracy level, 1:200, for example, then the information contained within that plan should stand up to the normal 1 sigma and 2 sigma accuracy levels for that designated scale, regardless of the source data.
Therefore, a survey undertaken by TLS with features extracted from the registered point cloud and the same survey undertaken by conventional Total stations (using the same survey control network) should both conform to the plan accuracy for the scale as defined in the project specification.
The point cloud at the end of the day is still your raw observation data set, just a great deal richer and just because it has been registered it does not mean that it is correct, just as any data set, redundancy and checking are vital to have a set of observations that you can rely on to give your client what it is that they need.
One of the big problems that may arise is that you only provide a pointcloud to the client and let them derive the information. You can warrant that the pointcloud is suitable to derive the necessary features at the required plan accuracy, But and it is a BIG But - you cannot warrant what the client derives as you will have no knowledge of their capabilities to undertake that process, without then checking what it is that they have extracted.
I hope that helps in some way.
Simon
- Joon
- V.I.P Member
- Posts: 335
- Joined: Wed Aug 21, 2013 8:01 pm
- 10
- Full Name: Sung Joon Ahn
- Company Details: CurvSurf
- Company Position Title: Founder + CEO
- Country: Republic of Korea
- Linkedin Profile: Yes
- Location: Seongnam-si, Korea
- Has thanked: 7 times
- Been thanked: 35 times
- Contact:
Re: Accuracy statement of features extracted from point clouds
We assume that a measurement point is a probable observation of an unknown object point lying next to the measurement point. Then, the measurement point substitutes an unknown object point.
RmsT^2 = RmsM^2 + RmsS^2, where
RmsT: total error rms,
RmsM: measurement error rms,
RmsS: software error rms, then
the measurement accuracy can be reserved to the total accuracy only if the software accuracy is at least 5~10 times higher than the measurement accuracy.
E.g., RmsM = 3,
RmsT = sqrt ( 9 + 1 ) = 3.162, if RmsS = 1 (3 times higher accuracy than RmsM),
RmsT = sqrt ( 9 + 0.36 ) = 3.059, if RmsS = 0.6 (5 times higher),
RmsT = sqrt ( 9 + 0.09 ) = 3.015, if RmsS = 0.3 (10 times higher).
Also, crucial to point cloud processing is how to sort out the inliers from point cloud (segmentation). Furthermore, how to automatically determine the shape of object surface is the current issue of point cloud processing (automatic feature extraction from point cloud).
An object surface has shape, size, position, and orientation in 3-D space. How to automatically and accurately determine the values of such parameters is the question.
For more information, please visit https://developers.curvsurf.com/Documen ... cepts.html
Source codes for Revit Plugin, GUI application, Intel RealSense 3-D camera are available at https://github.com/CurvSurf
Joon
If a set of measurement points (point cloud) is obtained by observing unknown object surface points. Then, a model surface is to substitute an unknown object surface while the sum of the squared errors between the model surface and the measurement points is to be minimized (Least Squares Method). The best error measure is the shortest distance between the model surface and a measurement point (Orthogonal Distance Fitting, a particular art of LSM).
Orthogonal Distance Fitting is prescribed as the ground truth algorithm in ISO 10360-6, standard for evaluating the accuracy of data processing algorithms used by coordinate metrology. ODF has a variety of advantages in accuracy, robustness, and segmentation. But, the implementation or development of ODF is not an easy job.
If we assume an error propagation model ofRmsT^2 = RmsM^2 + RmsS^2, where
RmsT: total error rms,
RmsM: measurement error rms,
RmsS: software error rms, then
the measurement accuracy can be reserved to the total accuracy only if the software accuracy is at least 5~10 times higher than the measurement accuracy.
E.g., RmsM = 3,
RmsT = sqrt ( 9 + 1 ) = 3.162, if RmsS = 1 (3 times higher accuracy than RmsM),
RmsT = sqrt ( 9 + 0.36 ) = 3.059, if RmsS = 0.6 (5 times higher),
RmsT = sqrt ( 9 + 0.09 ) = 3.015, if RmsS = 0.3 (10 times higher).
Also, crucial to point cloud processing is how to sort out the inliers from point cloud (segmentation). Furthermore, how to automatically determine the shape of object surface is the current issue of point cloud processing (automatic feature extraction from point cloud).
An object surface has shape, size, position, and orientation in 3-D space. How to automatically and accurately determine the values of such parameters is the question.
For more information, please visit https://developers.curvsurf.com/Documen ... cepts.html
Source codes for Revit Plugin, GUI application, Intel RealSense 3-D camera are available at https://github.com/CurvSurf
Joon
You do not have the required permissions to view the files attached to this post.
Last edited by Joon on Thu Apr 19, 2018 5:26 pm, edited 4 times in total.
- sreed
- V.I.P Member
- Posts: 257
- Joined: Mon Apr 28, 2014 4:31 pm
- 9
- Full Name: Scott Reed
- Company Details: Prologue Systems
- Company Position Title: Principal
- Country: USA
- Skype Name: sabotva
- Linkedin Profile: Yes
- Location: Richmond, VA USA
- Has thanked: 16 times
- Been thanked: 15 times
Re: Accuracy statement of features extracted from point clouds
I'm not sure if this answers your question, but you should look into the US Institute of Building Documentation. They have developed a robust standard for quantifying accuracy for extracted features and modeled components. Let me know if you need more information.
www.usibd.org
It seems to me that nearly every user on this forum should be a member and utilize the LoA documentation/standard.
www.usibd.org
It seems to me that nearly every user on this forum should be a member and utilize the LoA documentation/standard.