ACCURACY
~85%
Real-World Footage
Sports Analytics & Computer Vision
ACCURACY
Real-World Footage
RELIABILITY
With Custom Filter
DETECTION
Ball, Hoop & Players
SHOT TYPES
Layup, Jump, Three-Ptr
SEGMENTATION
Court Segmentation
When elite performance is capped by manual observation
Human analysts cannot track every movement across a fast-paced court without error. This reliance on manual review causes critical tactical insights to be lost to fatigue and subjective interpretation.
The Intellema Design Challenge
Basketball performance analysis typically requires labor-intensive manual annotation, which is prone to inconsistencies and delays. This slow process makes it hard for athletes to get the immediate feedback they need to improve their skills.
This project addressed the need for an automated system capable of identifying shots, tracking players, and generating real-time analytics from raw video. The system provides coaches with objective, scalable data to refine training and strategy.
High-speed YOLOv8 architecture for ball, hoop, and player detection, integrated with pose estimation to link shots to specific players.
SegFormer-based court and object segmentation to enhance context awareness and spatial understanding during gameplay.
Specialized false-positive removal module that improved overall detection reliability and tracking accuracy by 15%.
Extensive validation using diverse in-game and practice datasets to ensure scalable performance and accuracy in real-world environments.
