Cross camera tracking using VisionWorks + Tensorflow object detction in cpp

I want to do multi object detection and tracking on Nvidia Jetson Tx2. Nvidia’s Visionworks toolkit provides a object tracker sample application in which we need to draw a rectangle around the object to track it. I want to integrate object detction with this tracking code. So I need some help regarding that:

  1. Which object detection code will give the best performance on Nvidia Jetson Tx2 in cpp as I want to integrate it with VisionWorks object tracker code. As tensorflow object detction API is optimized with tensorrt for python.

  2. Can MobileNet-SSD accelerated with TensorRT be used for object detection : https://github.com/Ghustwb/MobileNet-SSD-TensorRT ?

  3. Is it possible to do cross camera object tracking using VisionWorks?

Hi,

1) Please try our DeepStream SDK: [url]https://developer.nvidia.com/deepstream-sdk[/url]

2) Please use this tutorial: GitHub - NVIDIA-AI-IOT/tf_trt_models: TensorFlow models accelerated with NVIDIA TensorRT

3)Could you share more information about the “cross camera” you means?
Do you want to use multiple camera?

Thanks.

Hi,

Thanks for your reply.

  1. Is it not possible to integrate TensorRT accelerated MobileNet-SSD for object detection : https://github.com/Ghustwb/MobileNet-SSD-TensorRT with VisionWorks? As I want to do everything on the edge i.e Jetson Tx2.

  2. The link that you have suggested is in python. I want something in cpp so that it can be integrated with Visionworks. Can you please suggest

  3. Yes multi camera tracking. I want to track in only two video streams for now. Is it possible to do with VisionWorks?

Thanks

Hi,

1.
Deepstream SDK is the current optimal since we do a lot on the acceleration.
We don’t enable VisionWorks <-> TensorRT path but you can try it on your own.
There is an example for VisionWorks <-> CUDA, which should be a good start.

2.We don’t have a UFF parser(TF->TRT) in c++ interface.
It’s recommended to convert the TensorFlow .pb file into uff with python first and launch it with TensorRT with C++.

3.We don’t have a sample for this.
But I think you can just append the camera frame and apply the tracking.

Thanks.