A four evening Tensorflow and Flask app to detect vehicles on live LTA cameras using deep learning-based object detection. Goals: Lightweight, simple (maybe not scalable) and likeable visuals.
A barebones Tensorflow and Flask app that you can deploy! The app downloads images from cameras available on LTA's DataMall, runs fast and accurate a Tensorflow implementation of a trained SSD-Mobilenet object detection model that finds vehicles, and visualizes the results on a main map:
For each camera, the most recent detection results are also shown, along with a chart tracking the last few detections.
Preferably, you need a Linux or Mac machine (Windows is perfectly fine, but you have to find the equivalent instructions). First, set up a virtualenv:
cd traffic-net
python -m venv traffic-net
source traffic-net/bin/activate
Install required packages by using:
pip install -r requirements.txt
You need to signup for a Mapbox account to get a Mapbox Access Token to display maps, and also you need to obtain a LTA DataMall token. Both are free to get, and you need to input the former in app/visualize.py and the latter in app/download_image.py. Then, use a screen (like tmux) and run the following command in a screen to run flask.
export FLASK_APP=traffic-net_flask.py
flask run
Then, run the following to start scraping and detecting!
cd app
python temp_scraper.py
And use a separate screen to start visualizing:
cd app
python visualize_loop.py
And head over to localhost:5000 to see live traffic conditions in Singapore
For training the SSD object detection model and visualizing the results, I used Tensorflow's Object Detection research repository. Flask and Plotly were used as the web framework and visualization library respectively.