Autopilot: An advanced perception, localization and path planning techniques for autonomous vehicles using YOLOv7 and MiDaS

A Project of Self-Driving Car 🚘

Appendix

Autopilot system is designed with motive of fully self driving system. The autopilot system takes the video as input. It can be directly taken from webcam or a saved video of dashcam from internet. That video is then processed frame by frame. Each frame produces the result of both object detection and depth estimation. First the frame is passed on to YOLOv7 state-of-the-art rapid object detector released just a few months ago. The machine learning model is trained and validated from COCO dataset. It is capable of identifying 80 different types of objects including car, trucks, motorcycles, pedestrians, animals, etc. Each obstacle is identified and detected by making bounding box around them. These boxes also provide with probability of object which determines accuracy of the detector with minimum prediction of 75% then frame is passed on to MiDaS which is trained on 12 different datasets and produces inferno colour depth map representing depth of pixels present inside frame. The road has been transformed to bird-eye-view using perspective transform for drivable space and A* routing algorithm has been used for path planning to determines the optimal and quickest route between two points. This plays critical role in decision making process for self driving without human intervention.

Tech Stack

Language: Python

Frameworks: tensorflow, pytorch

Run Locally 💻

Clone the project

  git clone https://github.com/Harsh19012003/Autopilot

Go to the project directory

  cd Autopilot

Install dependencies

  pip install -r requirements.txt

Run

  python detect.py

Features 📌

• Realtime low latency
• Upto 80 different objects
• Highly effective in bad weather
• Real time Depth Perception
• Occupancy Tracker
• Optimal Path Planning

Roadmap 🛣️

• Perception

  • Object Detection (YOLOv7 Algorithm)
  • Depth Perception (MiDaS)

• Localization

  • Top View Transformation (Perceptive Transformation)

• Path Planning

  • A* Routing Algorithm

Output 🏅

Output figure of YOLO Object Detection

Output figure of MiDas Depth Perception

Output figure of Top View Localization

Output figure of Path Planning

Feedback 📝

If you have any feedback, please reach out to us at harshdevmurari007@gmail.com

Contributing 🤝

Contributions are always welcome!

Contact harshdevmurari007@gmail.com for ways to get started.

Support

For support, email harshdevmurari007@gmail.com

Authors 👨🏻‍💻

• @harshdevmurari
• @gauthamkuckian
• @prajjwalvishwakarma