Edge TPU Multiple Video Streams Demo

This repo contains an end to end demo source code that allows to process multiple video streams with ML inference using 8 EdgeTPUs in one system. It is intended to be used with the ASUS AI Accelerator PCIe Card although it is also possible to use x86 or ARM systems with 8 TPUs (likely combining multiple USB Accelerators and a host like the the Dev Board).

The demo is hardcoded to run 5 separate video streams (defined in the Pipeline class) with different types of inferenceing running on the streams on a total number of 8 TPUs. In addition to the 8 TPUs, a mid to high end GPU is recommended to handle the mixing/rendering done by the GStreamer pipeline.

The window shows 6 video windows as follows:

1. Cars on highway with pipelined inferencing, 4 TPUs
2. Students in classroom with people segmentation, 1 TPU
3. Workers walking around with detection inferencing and a keepout zone, 1 TPU
4. Back yard video with detection inferencing and identification, 1 TPU
5. Bird video with object detection marking up birds and cropping the next window, 0.5 TPU
6. Cropped window of the same bird video as previous with classification of species (This model is cocompiled with the previous bird detection model), 0.5 TPU

Pressing 1-6 brings the selected video to the front and scales it up. Pressing any other key brings back the tiled view.

Building the demo for x86

The demo is only tested on an x86

If you have docker installed you can build it by typing (preferred way):

make DOCKER_TARGETS=demo docker-build

It is also possible to build locally, but you should validate your system has the dependecies listed in the Dockerfile.

make demo

Running the demo

./MultiVideoStreamsDemo

That's all!