SportTrack

tracking players and ball in sport

Datasets Available

detect football and players

• https://universe.roboflow.com/football-otrsl/football-player-detection-kucab-ofgzn

• https://universe.roboflow.com/university-of-oran-algeria/football-detection-vjuxg

• (SoccerNet v3 H250)[https://github.com/kmouts/SoccerNet_v3_H250]

football player referee and goalkeeper

• https://www.kaggle.com/datasets/theoviel/soccernet-tracking
• https://www.kaggle.com/datasets/atomscott/soccertrack

SportsMOT

players of 3(football, basketball and volleyball) https://www.kaggle.com/datasets/ayushspai/sportsmot

TeamTrack

soccer, basketball, and handball games multi view tracking https://www.kaggle.com/datasets/atomscott/teamtrack

Approach

• Step 1: Initial Training Train a YOLO model on football players and balls.

• Step 2: Prepare Mixed Player Data Use a dataset containing basketball, volleyball, and football players without ball annotations.

• Step 3: Freezing Layers Freeze the layers responsible for detecting footballs and players. Train the unfrozen layers with mixed player data to generalize across all sports.

• Step 4: Validate Test on football to confirm detection quality. Evaluate on basketball and volleyball players for generalization.

Metrics at first dataset:

Class	Precision (P)	Recall (R)	F1-Score	mAP50
Ball	0.88976	0.49999	0.64022	0.6127
Player	0.94752	0.96777	0.95754	0.9694
Referee	0.88898	0.85957	0.87403	0.9232

Metrics at 2nd and 3rd dataset:

Class	Precision (P)	Recall (R)	F1-Score	mAP50
Ball	0.91025	0.77554	0.83751	0.83682
Player	0.96357	0.97938	0.97141	0.98788

Tracking video:

sample image of first model

sample image of 2nd model

WBF diagram:

result using WBF

result modified