qaz812345/TrackNetV3

What it solves

TrackNetV3 is designed to accurately track shuttlecocks in badminton videos, overcoming common challenges like visual interference and temporary obstructions (occlusions) where the shuttlecock disappears from view.

How it works

The system uses two primary modules:

1. Trajectory Prediction: This module locates the shuttlecock by using an estimated background as auxiliary data to filter out visual noise. It also employs mixup data augmentation to improve the model's robustness in complex scenarios.
2. Rectification: When the shuttlecock is obstructed, the system analyzes the predicted trajectory to create repair masks and then uses inpainting to fill in the missing path, ensuring a smooth and continuous trajectory.

Who it’s for

This project is intended for sports analytics researchers and developers building automated badminton tracking systems.

Highlights

• High Accuracy: Outperforms YOLOv7 and TrackNetV2 in accuracy, precision, and F1 score on the Shuttlecock Trajectory Dataset.
• Background Estimation: Uses median images of matches or rallies to help the network distinguish the shuttlecock from the background.
• Trajectory Repair: Specifically addresses occlusions through a dedicated rectification module (InpaintNet).
• Error Analysis Tool: Includes a Dash-based application for detailed evaluation and error analysis of tracking results.