← Back to Work Projects
02 / Video Intelligence · Tracking Internal · Not Public

Ball & Shuttle Intelligence System

TrackNet-based tracking that produces frame-wise trajectories of the cricket ball and badminton shuttle under real broadcast conditions — occlusion, motion blur, lighting changes, and multi-angle cameras.

What

A computer-vision system that takes raw match video and outputs per-frame (x, y) coordinates of the ball or shuttle, along with confidence scores and physically-consistent trajectories.

Output feeds every downstream tactical layer — line and length charts, shot speed, smash trajectory analysis, and contact detection.

Why

Ball / shuttle position is the atomic unit of cricket and badminton analytics. Without reliable per-frame tracking, nothing downstream works — no line-length, no swing analysis, no smash speed.

Off-the-shelf detectors fail on small, fast-moving objects with motion blur. We needed a tracker that survived real-world broadcast feeds, not staged conditions.

How
  • TrackNet architecture as the detection backbone, trained on internal datasets curated through the annotation engine.
  • PyTorch training pipeline with sport-specific augmentation — motion blur, lighting shifts, occlusion simulation.
  • OpenCV post-processing for trajectory smoothing and physically-consistent interpolation across missed frames.
  • Failure-mode telemetry feeding back into the annotation engine for targeted retraining on weak frames.
Difficulties
  • Occlusion — the ball disappears behind the batter, the shuttle behind the racquet; the model has to recover trajectory after multi-frame gaps.
  • Motion blur — at smash speeds the shuttle becomes a streak across several frames; bounding-box detectors fail, point-based detection wins.
  • Multi-angle broadcasts — the same match cuts between cameras; the system has to detect camera changes and not extrapolate trajectories across them.
  • Generalization across venues — different lighting, different sightscreens, different broadcast quality. Robustness mattered more than peak accuracy.
Numbers
Frame
Per-frame trajectory output
2
Sports Supported
Multi
Camera angles handled
↻
HITL retraining loop