Project "Future vision transport"

Context: Autonomous Car Goals: Solutions for video stream images interpretation Repository of OpenClassrooms project 8' AI Engineer path

Future vision transport

Our role is to participate to the conception and development of the autonomous car. To do so, we need to :

• Train a segmentation model, able to differentiate human, road, vehicle, ... from Images (8 Classes)
• Develop API that predict 8 classes segmentation from Image input
• Deploy solution on Cloud

To do so, we rely on Cityscapes dataset.

You can see the results here :

• Presentation

• Technical Note

• Vidéo

• Notebook 1 : datagenerator

  • Data Analysis: Visualisation, Modification using OpenCV
  • Data Generator: Batch step by step ; Custom Data Generator - En Français
  • Masks Colors and 8 classes Labels: Source GitHub
  • Augmentation des Images : albumentation

• Notebook 2 : Trained Models

  • UNET Architecture - Neural Networks : Keras Tensorflow, CNN, Convolution, Max Pooling, GPU
  • Transfer Learning & Bacbones : Librairy Segmentation Models
  • Loss and Metrics evolution : DICE vs IOU ; Segmentation Metrics

• Best Models

• Web Application Deployment - Prototype

  • FastAPI Framework: uvicorn, Server, Swagger Interactive interface
  • Streamlit
  • Heroku