This repository demonstrates how to set up Azure Kinect camera with your Jetson platform, collect and annotate data, train a object ddetection model with the collected data, and finally run a real time object detection model on your development kit.

• Install Sensor SDK on Jetson
  • Install python packages
  • Get pretrained weights
• Collect/Annotate Data
  • Data collection
  • Annotate data
• Train Model

Note in this tutorial we will be installing the SDK on Ubuntu Version 18.04.5 LTS".

To check your distribution/version you can run the following command

cat /etc/os-release

Here is the link to microsoft package repository in case you are using any other distribution/version.
Here are more instructions on how to configure and install the SDK on other platforms Link

curl https://packages.microsoft.com/keys/microsoft.asc | sudo apt-key add -
sudo apt-add-repository https://packages.microsoft.com/ubuntu/18.04/multiarch/prod
sudo apt-get update

sudo apt install k4a-tools
sudo apt install libk4a1.4-dev

• in order to use the Azure Kinect SDK with the device and without being 'root', you will need to setup udev rules Link

• Copy 'scripts/99-k4a.rules' into '/etc/udev/rules.d/'.

• Detach and reattach Azure Kinect devices if attached during this process.

• test the SDK

k4aviewer

Note Here are more instuctions if you were experiencing dificulty with yout setup Link

• install python pip

sudo apt update
sudo apt-get install python3-pip

• install torch 1.7.0. Here is the Link to choose the correct versions.

wget https://nvidia.box.com/shared/static/wa34qwrwtk9njtyarwt5nvo6imenfy26.whl -O torch-1.7.0-cp36-cp36m-linux_aarch64.whl
sudo apt-get install python3-pip libopenblas-base libopenmpi-dev 
pip3 install Cython
pip3 install numpy torch-1.7.0-cp36-cp36m-linux_aarch64.whl

• install torch vision 0.8.1

sudo apt-get install libjpeg-dev zlib1g-dev libpython3-dev libavcodec-dev libavformat-dev libswscale-dev
git clone --branch v0.8.1 https://github.com/pytorch/vision torchvision
cd torchvision
export BUILD_VERSION=0.8.1  # where 0.8.1 is the torchvision version  
sudo python3 setup.py install
cd ../  # attempting to load torchvision from build dir will result in import error
pip3 install 'pillow<7' # always needed for Python 2.7, not needed torchvision v0.5.0+ with Python 3.6

• install requirments

# Make sure to install Protobuf compiler before running pip installation of onnx
sudo apt-get install protobuf-compiler libprotoc-dev

pip3 install -r requirments.txt

• install torch2trt for compiling the tensorrt model, here is a Link

• get the pretrained weight from the following link and place the file in:

cp DOWNLOADED_WEIGHTS PATH/TO/PROJECT/checkpoint/Logistic_ResnetCenterNet_fused.pth

We will discuss data collection and annotation in this section

• Setup JSON_ANNOTATION_PATH and SAVE_DATASET_PATH in pipeline/constants.py file
  • SAVE_DATASET_PATH setups the location to save the dataset. here is how the dataset directory hierarchy is setup by defaul it is set to PATH/TO/PROJ/annotation_json/

  data
    ├── annotation
    ├── casted_ir_image
    ├── depth_image
    ├── ir_image
    └── rgb_image

  • JSON_ANNOTATION_PATH setups the location to store the annotation json files to train the model (i.e train.json, val.json) these files will be created after parsing the annotation (.xml) files. by defaul it is set to PATH/TO/PROJ/annotation_json/

  {"img_path": "/path/to/image.png",
  "chw": [1, 576, 640],
  "boxes": [[345, 191, 384, 235], [424, 185, 467, 223], [309, 401, 341, 430], [152, 430, 198, 483]],
  "labels": [1, 1, 1, 1]}

• We used labelImg to annotate out dataset. you can install and run by running teh follwoing command

pip install labelImg
labelImg

• Open the DATASETPATH/data/casted_ir_image directory to load images and set the Save Dir to DATASETPATH/data/annotation and start annotating

• When finished with annotating the data you need to parse the Pascal VOC format and create json files containging annoation information. run pipeline/parse_pascal.py to create
  • label_map.json
  • train.json
  • val.json

Having the dataset ready We need to train the model (it is recomended to collect and annotate over 3k images for better performance).

• CHECKPOINT_PATH the path to save the model checkpoints (default *PATH/TO/PROJ/checkpoint)
• you can set the to either train with only depth images or with ir and depth fused images (setup in pipeline/constants.py)
  • depth input images: image size (3, 300, 300) dublicate the image for all the 3 input channels
  • fused input images: channel1 ir images, channel 2 depth image, channel 3 mean of the first 2 channels
• loss functions: you can either use Logistic or MSE loss for the heatmap regression (setup in pipeline/constants.py)
• setup max epoch in pipeline/constants.py*
• Model naming convention: LossFunc_ModelName_dataType.pth (i.e Logistic_CenterNet_fused.pth or Logistic_CenterNet_depth.pth)

run training by

python3 train # To start training a model from scratch

python3 train -r True # To continue an already trained model

run training by

python3 visualization.py # To run inference on a validation image

run training by

python3 azure_kinect.py # To run real time inference