[VisionWorks-Tracking] Proper framework usage

Autonomous Machines Jetson & Embedded Systems Jetson TX2
1

Dear all,

I am developing a surveillance application based on the Jetson TX2 board.
This software has to process frames from an IP camera, detect people in frames and track their movements.
I implemented a deep neural network based algorithm to detect people in frames giving very good mAP running at 3FPS.
I was looking for an object tracker to track detected bounding boxes in images. I came across the nvidia VisionWorks tracker to track rigid objects for short time intervals.

As far as I understood, main ObjectTracker methods are:

  • addObject (adding a bounding box to be tracked while returning a pointer to tracked object)
  • process (using a grayscale image to update tracked objects status and location)

In my SW, I would like to perform the following steps :

1- Detection.
2- Use found bounding boxes in the object tracking.
Iterate:
3- Perform a new detection.
4- Solve the data association problem.
5- Perform object tracking.

After step 2, the objects have an associated ID, a location and a status. I can access them with TrackedObject getId, getLocation and getStatus. However, after data association (step 4), I cannot set neither the location nor the status to a new value (the one computed by the detector).

I thought I could remove the old object and add a new one using the ObjectTracker removeObject and addObject but this action does not preserve the ID.

Do you have any suggestion? Have I misunderstood something?

Thank you.

Attilio

2

Hi,

Please check here for object tracker document:
VisionWorks API > Samples and Demos > Demo Applications > CUDA Layer Object Tracker Sample App

Tracking is applied by tracker->process().
You can add a new object via addObject(), remove a object via removeObject().

For example, if your flow is detect, track, track, detect, track, track, detect, …
t0: Add obj1
t1: track for obj1 → track1
t2: track for track2 → track2
t3: Add obj2, remove track2
…

Thanks.

3

@AastaLLL Thank you for your reply.

I read the suggested example and it is very interesting. However, in t3 I would like to preserve the track identification number, i.e., if a detection is marked as ID 1 in t1, I would like the same detection to be ID 1 after t3. I didn’t understand if the approach you suggested would guarantee this requirement.

Thanks.

4

Hi,

Do you want a multi-object tracking algorithm or a refreshable tracking algorithm?
VisionWorks can support multiple object tracking but can’t manually refresh object location on the fly.

If you want to create a new object, just use addObject to add it.
If you want to refresh the location of a tracked object, it’s recommended to remove the tracked object first and add the object back with the desired position.

Thanks.

5

Hello,

I want to understand and try to use it too. But on my device, I don’t see the ‘/usr/share/visionworks/sources’ folder.

There are only two folders, namely, cmake and docs. Is it available elsewhere?

6

Hi,

This sample is contained in VisionWorks package.
Please remember to install VisionWorks package first.

Thanks.

7

Hi, Thanks for your answer. I figured it out. However, is there any workaround for not installing OpenCV4Tegra? Currently, the VisionWorks API is dependent on it but I already use OpenCV 3.3 and I understand that it is difficult to have two different versions on the board.

Thanks,
Bhargav

8

Hi,

You can purge the OpenCV4Tegra directly.
Most of the VisionWorks API is independent of OpenCV library except some sample to demonstrate the interoperability between VisionWorks and OpenCV.

Thanks.