Dash cam danger detection

Requirements

Python >= 3.8 (and maybe >=3.7 could also works but I'm not sure).

Install them, for example via pip in a terminal:

pip install --requirement ./requirements.txt

Or any virtualenv or something that you want.

if you have trouble installing scikit with pycharm :

python -m pip install scikit-learn==0.23.2

Data

Install the Traffic sign dataset, unzip and put it in the folder ./data/traffic-sign.

Then, run the script traffic_sign_data_cleaning.py. This script will restructure the dataset to allow our models to work easier (it does some label to folder mapping).

Tensorflow

Follow instructions here: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf2.md

The base tutorial: https://github.com/tensorflow/models/blob/master/research/object_detection/colab_tutorials/object_detection_tutorial.ipynb

Make sure that you have Tensorflow 2 installed, Tensorflow 1 won't work. If have followed the previous step, then it should be fine.

Models and Data Provenance

Coco-SSD

• Description
  • This model uses 80 categories from the COCO image dataset to detect objects.
• References
  • Tensorflow Object Detection API.
  • Paper by Microsoft: COCO: Common Objects in Context
  • COCO website.
• Sources
  • We use a detection model pre-trained on the COCO 2017 dataset by Tensorflow. You can find a list of models here.
  • We use the SSD mobilenet v2 coco17 because only SSD models are supported by Tensorflow Lite, which could be handy in the future and this is a very performant model.

Traffic sign

We trained a model to recognize traffic signs. The model is a MobilenetV2 trained from scratch. That means we didn't use transfer learning based on the COCO dataset (or anything else).

• Description
  • This model uses 43 traffic sign categories from the GTSRB dataset.
• References
  • Tensorflow 2.
  • MobilenetV2 model.
• Sources
  • You can find the dataset here.
  • You can find labels here. It is a csv file that map a number (used by the model) to a class name, used by humans.
  • You can find the trained model here.