Speeding up Point Tracking

I'm using a vision.PointTracker to track feature points over 3 images. The images tracked points are stored in an imageviewset to generate 3D-points using triangulateMultiview. The goal is to do this process mostly realtime. With increasing camera resolution(1920x1080), the runtime increases significantly with each tracking step taking more than 0.1s resulting in an overall output of about 2FPS.

Is there a way to set up GPU support for the Tracker or a different, faster approach to Pointtracking?

Thanks in advance

Answer 1

[Udit06]

Hi Tazzmani,

You can try the following approaches to make the point tracking faster:

1) Reduce the "NumPyramidLevels" property of "vision.PointTracker" which is by default set to 3.

2) Reduce the "BlockSize" property of vision.PointTracker which indicates the size of neighborhood around each point being tracked. "BlockSize" property is directly proportional to the computation time.

3) You can also reduce the "MaxIterations" that KLT algorithm performs to reduce the computation time. By default, it is assigned as 30, but the algorithm typically converges within 10 iterations.

You can refer to the following MathWorks documentation to understand more about the "vision.PointTracker" properties

https://www.mathworks.com/help/vision/ref/vision.pointtracker-system-object.html#:~:text=points%20for%20tracking.-,Properties,-expand%20all

I hope this helps.