Traffic survey automator is a software that analyses traffic recordings using deep learning and statistical algorithms. It outputs the numbers of vehicles passing through the defined regions of interest.

It was developed as a part of a master's thesis at Faculty of Information Technology at Czech Technical University in Prague in Spring 2022.

The application consists of a backend and a frontend. The frontend displays information about existing source videos, performed analyses, shows visualizations and results and enables a user to create new task for analysis, including parameters customization.

Backend servers responses to the requests from frontend, schedules the analysis tasks which are executed by celery workers asynchronously.

Models:

• EfficientDet D6
• EfficientDet D5 Adv-Prop

Algorithms:

• Simple Online and Realtime Tracking with constant acceleration Kalman filter modelling
• Deep Online and Realtime Tracking with the same Kalman filter model, VGG16 feature embedding

Technologies:

• tensorflow
• OpenCV
• FastAPI, pydantic, SQLAlchemy
• ReactJS

The common pre-requirement is to download the pre-trained deep learning models for object detection. How we created these checkpoints is described in section Exporting models below. The models have to be downloaded and stored in the MODELS_PATH.

• EfficientDet D5 Adv-prop: https://owncloud.cesnet.cz/index.php/s/IUnmEK9iFol9NaF
• EfficientDet D6: https://owncloud.cesnet.cz/index.php/s/KPeGv4rl3cJKLDj

We provide Dockerfiles for both backend and frontend, plus a docker-compose to bring everything together. The docker-compose configuration describes well what is needed to run the whole application.

PostgreSQL and redis are included for simple and fast start-up. Both can be replaced by an online alternative by changing the environment variables DATABASE_NAME, DATABASE_URL and CELERY_BROKER on tsa-backend. The database URL should be without postgresql:// prefix. This is added automatically by the application as postgresql+asyncpg://.

Four folders are necessary:

• source_files: the source video recordings
• output_analysis: the output JSON analysis data are stored here
• models: source of the deep learning models checkpoints
• postgres_data: place where the docker-composed PostgreSQL stores its data

All of these locations are mounted from the local system and the necessary environment variables are set up. Please, adjust those, if needed.

By default, backend is available on http://localhost:8000 and frontend on http://localhost:3000. You can adjust the ports.

The application's backend runs on Python 3.8 and 3.9 using poetry dependency manager. First, you have to have one of the required versions of Python and poetry installed. Then, follow the steps below.

1. Run poetry install inside the backend/ folder.
2. Download the deep learning models and place them into the models/ folder.
3. Export the necessary environment variables and create the required folders to store data. It is the set of 6 variables defined in the docker-compose.yaml, tsa-backend.
4. Optionally, neptune.ai project name and api key can be set up to monitor the experiments via Neptune. See tsa.config.base how to set these.
5. Activate the virtual environment poetry shell.
6. Apply migrations through alembic upgrade head, if necessary.
7. Run a CLI utility command located in cli/ package. Refer to the documentation of those commands by running python cli/*.py --help.
8. Start the uvicorn server uvicorn --port 8000 tsa.app.app:fast_app
9. There is a set of poe-the-poet commands predefined for quick execution:
   • poe black to format the code,
   • poe sort to sort imports,
   • poe test to run those few tests,
   • poe run to start the backend application on port 8000,
   • poe worker to start the celery worker that processes the analyses queue.

To run the application's frontend:

1. Make sure you have node and npm installed on your machine.
2. Create a copy of .env.example named .env and set the API URL according to where backend is located.
3. Run npm install to install all the dependencies.
4. Run npm start to start the frontend application on port 4000.

The models that we use are pretrained on COCO dataset. These are built from Google's AutoML implementation located at https://github.com/google/automl/tree/0b0ba5ebd0860edd939465fc4152da4ff9f79b44/efficientdet.

These models perform a non-max suppression as a part of their pipeline. Our configuration of NMS parameters minimizes the effect of this operation to minimum so that our own version of NMS can affect the results.

D5 version is trained with AdvProp and AutoAugmentation which should improve model's robustness and ability to generalize on images that are blured.

It is converted from the original checkpoint (https://storage.googleapis.com/cloud-tpu-checkpoints/efficientdet/advprop/efficientdet-d5.tar.gz) using the following command:

python efficientdet/model_inspect.py \
    --runmode=saved_model \
    --model_name=efficientdet-d5 \
    --ckpt_path=original_models/efficientdet-d5 \
    --saved_model_dir=traffic_survey_automation/models/efficientdet-d5-advprop-aa \
    --batch_size=0 \
    --hparams=params.yaml \
    --max_boxes_to_draw=512 \
    --min_score_thresh=0.05 \
    --nms_method=gaussian

This is a standard version of the architecture extracted from checkpoint (https://storage.googleapis.com/cloud-tpu-checkpoints/efficientdet/coco2/efficientdet-d6.tar.gz) using:

python efficientdet/model_inspect.py \
    --runmode=saved_model \
    --model_name=efficientdet-d6 \
    --ckpt_path=original_models/efficientdet-d6 \
    --saved_model_dir=traffic_survey_automation/models/efficientdet-d6 \
    --batch_size=0 \
    --hparams=params.yaml \
    --max_boxes_to_draw=512 \
    --min_score_thresh=0.05 \
    --nms_method=gaussian

The params.yaml file is configured as follows:

image_size: "1280x720"
nms_configs:
  sigma: 0.6
  iou_thresh: 0.97

We made a slight adjustment to the code of efficientdet/tf2/postprocess.py:186 to allow configuration of other than 1.0 value for iou_threshold to:

iou_thresh = nms_configs['iou_thresh'] or 1.0