saifrais/w210-accident-detection

Deep Learning at the Edge to build a model to assist surveillance cameras to detect accidents, as they happen.

w251-accident-detection: Deep Learning at the Edge

Contents

• Business Case
  • Problem Statement
  • Solution
• Dataset
  • Road Collision Videos
  • Data Engineering and Annotation

Business Case

Problem Statement

• Survival rate dependent on Emergency Response
• Hit-and-run incidents rising with higher fatalities
• More traffic cameras means less attention span
• Automated notifications can assist monitoring
• Improve emergency response time

Solution: Build an accident-detection algorithm for surveillance cameras and deploy a prototype on an edge device.

Dataset

Road Collision Videos

• Sourced from Youtube
• Collated by Ankit et al, Carnegie Mellon University
• 5.2 hours of footage
• 45GB
• 1,416 videos
• 518,256 extracted video frames
• Variety of weather and lighting amongst frames

Data Engineering and Annotation

Sample Data

Training