VisionWorks Object Tracking with Vision Primitives API

Autonomous Machines Jetson & Embedded Systems Jetson Xavier NX
1

Hello everyone,
I need to develop a tracking algorithm for objects, I was following the example in the documentation, the following below

/usr/share/visionworks/sources/demos/feature_tracker

./nvx_demo_feature_tracker [options]

but this algorithm tracks all features and not just one object.I saw another example but made with VisionWorks-CUDA where tracking is done on objects.

/usr/share/visionworks/sources/demos/object_tracker_nvxcu/

Is there a possibility to do the same with the Vision Primitives API algorithm with graphs and nodes?

One solution would be to use the mask as input to the nvxFastTrackNode , but in the next frame where I don’t provide the mask the algorithm finds other points and tracks them all, so I should provide the mask every frame but how do I get it?

I want to select points on an object and track only the points of that object while it is moving.

Thanks
I’m waiting for some answers

3

Hi,

VPI only has a feature tracking algorithm:
https://docs.nvidia.com/vpi/algo_klt_tracker.html

A heuristics approach is to set the mask as a window.
And limit its scaling ratio (ex. [0.9, 1.1]) based on your use case to filter out other unwanted features.

Thanks.

4

So you can’t achieve this with VisionWorks?

Does the example of VisionWorks-CUDA not do this?

/usr/share/visionworks/sources/demos/object_tracker_nvxcu/

I ask you this because I’m already using the part of Semi-Global Matching with VisionWorks and I wanted to use this library to get the tracking as well, it’s done for all the points of the image I was wondering if there was some modification I could do to make the tracking on the points of some objects.

Instead of using this library VPI - Vision Programming Interface isn’t it better to use the example with VisionWorks-Cuda Object_tracker_nvxcu?

Thanks for answer

5

Hi,

Is there a possibility to do the same with the Vision Primitives API algorithm with graphs and nodes?

Sorry that I thought you are asking the same functionality in VPI.

You can use the object_tracker_nvxcu in VisionWorks.
It filters out the features based on the motion distance.

For more details, you can find it in our document below:

Thanks.

6

Thank you for your answer
Should I use the VisionWorks or VPI library?
consequently the algorithm object_tracker_nvxcu or VPI - : KLT Feature Tracker ?

I’m using VisionWorks for stereo correspondence, give me some advice.
Thanks

7

Hi,

Based on your use case, it’s more recommended to use VisionWorks.

But please noted that we don’t add new features into VisionWorks anymore.
This will limit you to request for a new algorithm if not available in VisionWorks currently.

Thanks.

8

Thanks for the answer, if I had to migrate everything to VPI, does the stereo correspondence algorithm have the same performance as VisionWorks?
Thanks

9

Hi,

We don’t have a benchmark report for VPI vs. VisionWorks.
But you should get the similar or better performance with VPI since we keep optimizing it.

Thanks.

10

Ok I will consider migrating to VPI thanks for your support and answers