The ability to troubleshoot your data is a valuable skill that takes a bit of practice. Some consider our ability to identify the cause of errors in data to be unique, however, the procedure we follow is quite basic and uses tools available within Theia3D. The goal of this blog post is to present the procedure that we follow to enable you to find issues within your own data, and by contacting us here you can get a checklist of your own to follow.
Calibration
The first part of the troubleshooting process is to verify the calibration - which is where we find the majority of our issues. Start off by opening the trial in Theia3D, loading the video, calibration, and preferences and then running the full analysis. What do you notice?
With the skeleton turned off, toggle each view individually and verify that the reference frame is located at the same position in each view. You may find that one particular view is off and sometimes it’s more obvious than others (this occurs typically because of a bumped camera which can result in an origin location that varies from view-to-view). If this is the case for a lot of data, you may need to consider whether the data is worth keeping, however, if there are many cameras and you aren’t worried about losing one particular camera, you can deactivate that view and reanalyze (Tools -> Toggle Views), which will remove that view from the calculations. The example images for this were selected intentionally; because the calibration was not aligned with a landmark on the floor, it makes it quite difficult to verify the location across multiple views. For this reason, we recommend aligning the calibration board with a landmark on the floor that you can see in every view, so you can perform this step a bit more easily.
In this image you can see that the origin is properly aligned with the edge of the treadmill belt (a very obvious landmark) across multiple camera views.
The second aspect of the calibration check is to verify the lens parameters. Sometimes incorrect parameters are loaded for a calibration - for instance, loading the default Sony parameters with a non-Sony camera system. The volume in some cases may calibrate with less than ideal residuals, however there will be downstream consequences in the data. To review this, with the same trial opened and analyzed as above (now with the skeleton turned on), watch the participant as they travel through the volume. If the skeleton is tracked well near the center of the images (or close to the origin), and looks very much off near the edges of the camera, this often means that there is an issue with the lens parameters that were loaded. To correct this, you need to open the extrinsics calibration trial, load the correct lens parameters, re-calibrate the volume, and re-perform the analysis. After running this process, you should see better camera calibration residuals and the person should track consistently independent of their location within the volume.
Filter
An incorrect filter for the trial is the second most common issue that we see (check out our blog posts on filters here and here). If you have accidentally selected a filter frequency that is too low for the motion being analyzed (i.e. a pitch with a filter of 8Hz) you may see segments that are either attenuated in the tracking or segments that drop out during the fast part of the motion. Correcting for this is simple - adjust the filter within the preferences to a more appropriate value and reanalyze. This error happens frequently because Theia3D is used to analyze many different movements and it’s easy to forget to change the filter back to the correct value. To avoid this - if you are using TMBatch to analyze your data - having a standard operating procedure for each motion (i.e. a checklist) that has this as one of the checks is best practice, and within Theia3D, we verify that value before running any analysis.
Wrong Person Identified
Because Theia3D does multi-person tracking by default, the person of interest may not be identified correctly when there are multiple people within the capture volume. This one is a bit more subtle but is pretty easy to fix. There is a setting that defines how to order the people being tracked (this blog post will give you some insight), and selecting the right preference can help avoid this. If there are background people jumping in and out of tracking, adding a capture volume box to the settings will likely remove them.
In some cases, the settings are correct but the person of interest is still not correct. For example, in a situation where a participant is walking with another person as an aide, and the aide walks closest to the origin, they may be selected as the subject incorrectly. In this case, the aide will be identified as the primary person, and changing this requires manual intervention using Tools->Modify People IDs. In this dialog, you can drag and drop the person order, and then re-save the C3D file.
Navigate to the Tools section of Theia3D and you can find many settings to fine-tune your collection process and ensure the correct subject is identified.
Visibility/Video Quality/General Product Fit
Sometimes there isn’t a whole lot we can do - the camera setup and motions may not be conducive to high fidelity tracking. This can occur because of significant occlusions within the volume, the video quality is low, the camera system can’t record at an adequate frame rate, the camera system isn’t set up correctly (too far away, cameras too high), or the motions being tracked are generally not recommended. Though this can be frustrating - it’s pretty easy to avoid during experimental setup (check out this blog post for tips regarding our recommended setup procedures). I try to reinforce this as much as I can - running a pilot study should be the first step of any scientific experiment and will help you determine whether the system can measure the desired motion. It may sound like extra work - but it is far better than recruiting, collecting all of your data, and realizing after the fact that it isn’t going to produce results in line with your expectations. One of our customers recently did just that and sent us over some data - it was a study on two participants hugging. Given the camera setup and the motion - tracking the people individually was not possible and we didn’t recommend it moving forward! However, the customer in this case is very experienced and this issue was addressed at the very beginning during a pilot study. Because the motions were not feasible, they moved on to other questions and wasted little time and resources on this question.
As you can see, the process is fairly simple and can be performed at your site. Following these four procedures will allow you to find issues within your data, and hopefully generate some standard operating procedures so as to avoid them in the future. We are always happy to help with data troubleshooting - so if you are having issues after troubleshooting with this guide, let us know and we are happy to provide further assistance and detail on what is occurring.
To learn more about Theia or receive a checklist of the above methods to aid in your own troubleshooting, contact us here!
Comments