search

OpenPTrack / open_ptrack_v2

OPT v2 "Gnocchi" - adding object tracking and pose recognition
BSD 3-Clause "New" or "Revised" License
220 stars 65 forks source link

Dual RealSense Person Tracking Setup #116

Closed Julian243 closed 4 years ago

Julian243 commented 4 years ago

Hello, I read in issue #51 that tracking results with realsense cameras were pretty bad. Does anyone have experience with tracking results for realsense with v2.2 (released in december)? As I only have one PC that is powerful enough I was also wondering whether dual camera person tracking on a single PC works sufficient enough. My PC is equipped with an Intel i7-6700K with 4 GHz, 16 GB RAM and a GeForce RTX 2080 Ti. If anyone has experience please let me know! Thanks a lot!

DiegoJRR commented 4 years ago

Any updates on this? Are the results with the Intel RealSense D435i also bad?

bsubbaraman commented 4 years ago

We've seen decent results with the RealSense after tuning tracking parameters based on your environment. For example, here is a video I recently took of a single camera RealSense. Before tuning, however, you may seen many erroneous tracks. Over larger distances/bigger spaces than that shown in the video, the RealSense quality decreases due to pointcloud noise.

Running two RealSense from a single machine should work out of the box, but it hasn't been tested as I only have 1 here in the lab. To preemptively provide debugging direction, it might take a bit of editing listener.py to ensure the serial number/unique identification of each Realsense is properly handled. Do you plan on using multiple modules (person + skeleton + object) or just person tracking? I would expect some drop in fps/quality if using multiple modules, but two person tracking instances shouldn't be a problem given your setup.

Let me know if you have any more questions!

Julian243 commented 4 years ago

We've seen decent results with the RealSense after tuning tracking parameters based on your environment. For example, here is a video I recently took of a single camera RealSense. Before tuning, however, you may seen many erroneous tracks. Over larger distances/bigger spaces than that shown in the video, the RealSense quality decreases due to pointcloud noise.

Running two RealSense from a single machine should work out of the box, but it hasn't been tested as I only have 1 here in the lab. To preemptively provide debugging direction, it might take a bit of editing listener.py to ensure the serial number/unique identification of each Realsense is properly handled. Do you plan on using multiple modules (person + skeleton + object) or just person tracking? I would expect some drop in fps/quality if using multiple modules, but two person tracking instances shouldn't be a problem given your setup.

Let me know if you have any more questions!

Thanks a lot for your response!! I wanted to use the skeleton tracking module as well. I think just the skeleton tracking module itself might be sufficient. Do you happen to know what results to expect in terms of accuracy and reliablity?

Thanks a lot for your help!

jburkeucla commented 4 years ago

Performance depends a lot on the circumstances. We've been focusing on skeleton tracking for recognition rather than real-time use, but performance in small spaces with good intrinsic and extrinsic calibration is reasonable but still sort of noisy.