Hi,
For running deep learning inference, We would suggest use DeepStream SDK. Here is the announcement of latest version 5.0.1:
Announcing DeepStream 5.0.1
The implementation is based on gstreamer.
And you may also try jetson_multimedia_api. The frame capture is through v4l2 interface. We have two samples:
/usr/src/jetson_multimedia_api/samples/v4l2cuda
/usr/src/jetson_multimedia_api/samples/12_camera_v4l2_cuda
If your source supports UVC driver, you may try the samples. We don’t have experience in capturing through libusb. Maybe you can check the existing samples and do integration.
Document of jetson_multimedia_api:
https://docs.nvidia.com/jetson/l4t-multimedia/index.html