Kaggle NFL Health & Safety - Helmet Assignment 64th/833

Thanks for the contribution of my teammates @shaoeric and @Leonrain-Liu. For more information, please visit 刘兄的复盘总结 和 我的复盘总结

Competition introduction

• competition link: NFL Health & Safety - Helmet Assignment
• competition dataset: Dataset
• competition task: In this competition, you’ll identify and assign football players’ helmets from video footage. In particular, you'll create algorithms capable of assigning detected helmet impacts to correct players via tracking information.

Dependence

$ pip install easydict
$ sudo apt-get install ffmpeg

Idea and contributions

• A frame edge cropping method to remove the interference of the players and spectators outside the court
• Camera position prediction methods
  • Sideline camera position prediction by rotating mapping on the original and another-side ngs tracking
  • Endzone camera position prediction by Jersey number plate recognition with OCR
• X/Y axis rotating mapping selection
• Dual deepsort trackers on helmets and players fusion strategy

Results

• Sideline camera position: correct 59/60 total videos
• Endzone camera position: correct 51/60 total videos
• XY mapping 24 validation videos: 0.23
• deepsort postprocessing and fusion: 0.529