120 fps tracking of 100-500 particles on an AMD 3950X (16C) running in eco mode at 70% CPU utilization. The bottleneck is generally at the Tracking node which performs keypoint optical flow on the 2D camera image to find the corresponding targets between frames. We investigated several methods using CUDA for this section, but so far haven't been able to find a better performing system than what we have already. In theory, the prior knowledge of the target's trajectory can be useful to either initialize the sparse optical flow or replace it entirely for some tracked targets (this is not straightforwards because the tracking is run backwards from current to prior).
Introduction
Animations of it working
Example performance
120 fps tracking of 100-500 particles on an AMD 3950X (16C) running in eco mode at 70% CPU utilization. The bottleneck is generally at the Tracking node which performs keypoint optical flow on the 2D camera image to find the corresponding targets between frames. We investigated several methods using CUDA for this section, but so far haven't been able to find a better performing system than what we have already. In theory, the prior knowledge of the target's trajectory can be useful to either initialize the sparse optical flow or replace it entirely for some tracked targets (this is not straightforwards because the tracking is run backwards from current to prior).
Hardware
Building
Requirements
Git lfs if using git
Layout
MoCapLib
Operators
Building OpenCV