video analytics software architecture

I am new to Nvidia HW and eco system, looking for some advice on the video analytics system I am trying to build.

The system will be based on TX1, with multiple camera sensors input. I need one software module performing the DVR function, multiple modules for analytics, say one for vehicle detection on cam A, another one for pedestrian detection on cam B. the analytics will use Caffe inference. I would like them to be independent so that I can later add a new VA function without changing existing code.The configuration should be flexible enough to change input/output channels.

Should I use a gstreamer pipeline to do this, or standalone c applications? Should I implement our own plugin? I have only superficial understanding of gstreamer. so please be specific when you kindly give me some suggestion.

Hi hongfeng,
PLease refer to the sample 11_camera_object_identification

It is similar to your case. You can install the samples via Jetpack.

Thanks for the reply.
I did check that particular sample, my question is more about how I should incorporate that with other functions and make it flexible enough. I image I should use a gstreamer pipeline, and the VA module uses appsink to pull frames. I want to make sure this is the optimal way to achieve it.


Hi hongfeng,
Since 11_camera_object_identification is very close to your usecase, we would suggest you begin with it.

We don’t have samples exactly for gstreamer + Caffe. Please refer to gstreamer guide and check how to implement your usecase.