Created by: Spencer Hallyburton
To test networking without real sensors, you can download image data to playback. The Makefile is by default configured to use ADL-Rundle-6 from the MOT15 challenge. To download the data, run:
./download_tracking.sh
from inside the data folder. NOTE: if you system does not allow you to make a folder at /data/tracking, then pass an argument to the call above with your custom download folder, e.g., ./download_tracking.sh ./data/tracking aka locally. You can use local downloads of any AVstack-compatible dataset such as KITTI or nuScenes.
Ensure the symbolic links attached appropriately to the location that the data was downloaded.
First, ensure that the submodules (in the folder third_party) are initialized. You can do this with git submodule update --recursive. If doing this for the first time, add an --init flag as well.
The only configuration of third party libraries you'll need is to download perception models. You will need to do this in the third_party/lib-avstack-core/models folder. Follow the installation instructions for AVstack for more clarity.
After the third party liaries are initialized, we'll use poetry to manage a python environment. To get started, ensure poetry is installed on your machine by following the installation instructions here.
NOTE: if the installation is giving you troubles (e.g., in my experience, with pycocotools in aarch64-based chips), try the following:
poetry config experimental.new-installer false
to get through the problem packages, then reactivation with true.
Once poetry is installed, you can set up the poetry environment with assistance from the Makefile. Specifically, try running:
make install
If all goes well, you will have a working python environment!
After we build the jetson-inference library, we need to link the site packages by running:
./link_site_packages.sh
You will just need two terminal windows to run a replay system. The first will run our controller, the second will replay the data. Specifically, run:
make controller CCONF=camera
in the first window (waiting for the perception models to say they are initialized) and
make mot15_replay # or make kitti_replay, e.g.
in the second window. If all goes well, you will see the sensor data replayed on the Qt window with tracks and labels.
This demo platform can also be run with physical sensors over an ethernet or serial connection. We demonstated this capability at ICCPS 2023. We plan to provide more details on how to accomplish this in the future. If particularly interested, please reach out to us, and we can schedule a consultation.
[Creating virtual network interfaces][virtual-network], run sudo ip addr add 192.168.1.11/24 dev eth0