I do not aware of object tracking application; however, in general, as in other applications, the toplogy should depend on the input and output dimensions and semantics of the given neural network, and I believe each neural network type for object tracking would differ. I think the gst-nvtracker is for a specific instance of obj-trk neural networks.

1. If your object tracking neural net requires to have a seqeunce of video frames fed simultaneously as in "activity recognition", you need to create a batched frame from a sequence of frames in a stream, use tensor_aggregator

   - refer to ICSE 2021 SEIP paper of nnstreamer

   -  https://github.com/nnstreamer/nnstreamer/blob/main/gst/nnstreamer/elements/gsttensor_aggregator.md

2. If your object tracking neural net requires to have a previous result of the net fed along with the current video frame, you need to create one with tensor-mux, tensor-reposrc, and tensor-reposink.

   -  

Cheers,

MyungJoo