Skip to content
Roshil Shrestha
All projects

Client project for Darkhorse Soccer Analytics

Soccer Detection & Tracking

Detects and tracks the ball, players and jersey numbers in match video at 30 FPS, backed by an annotation team I led.

Computer vision

Overview

A computer vision system for Darkhorse Soccer Analytics that detects and tracks the soccer ball, the players and their jersey numbers in short and long match clips. Models were trained on real photos and frames from actual game video.

The problem

A soccer ball is a tiny, fast object in high-resolution footage, often blurred or hidden behind players. Tracking it reliably, alongside players and their numbers, needs well-labelled data and detection tuned for small objects, while still running in real time.

What I did

  • Worked with the team to build the system that detects and tracks balls, players and jersey numbers.
  • Led an 8-member annotation team, keeping labelled datasets high quality with Roboflow and CVAT.
  • Designed and implemented the ball detection system using SAHI.

How it works

  • SAHI (Slicing Aided Hyper Inference) with 6-slice and 8-slice configurations for small-object detection in high-resolution frames.
  • RT-DETR, YOLOv8 and YOLOX for detecting and tracking the ball.
  • Kalman filters to follow the ball's movement between frames.
  • Post-processing tuned for small-object tracking, improving reliability and efficiency.

Outcome

  • Real-time performance at 30 FPS, balancing accuracy and scalability for sports analytics.