Case Study

Basketball Shot Detection

Category:

Sports Analytics & Computer Vision

Impact:

10 Weeks | 85% Accuracy

Background

Basketball performance analysis is often manual and resource-intensive, requiring human observers to annotate video footage of games. Traditional methods are time-consuming and prone to error, limiting the ability of coaches, analysts, and players to access objective insights. An AI-powered, automated system was needed to detect shots, classify attempts, and provide accurate, real-time performance analytics directly from video footage.

Project Goals

Detect basketball shots in live and recorded footage
Track the ball, hoop, and players with high accuracy
Identify the shot taker and classify shot type (e.g., layup, jump shot, three-pointer)
Improve efficiency and reduce false positives in shot recognition
Enable scalable video analytics for training and sports strategy

Our Approach

Object Detection & Tracking

Utilized YOLOv8 for high-speed ball, hoop, and player detection. Integrated pose estimation to accurately link shots to specific players.

Semantic Segmentation

Applied SegFormer for robust court and object segmentation, improving context awareness.

Custom Algorithm Development

Designed a false-positive removal module, boosting reliability by 15%.

Model Training & Testing

Trained models on diverse in-game and practice footage datasets. Validated with real-world basketball videos, achieving scalable and practical accuracy.

Key Results

~85% accuracy achieved on real-world basketball footage
Automated recognition of shot types and shot takers
15% improvement in detection reliability with custom filtering
Enabled scalable, AI-powered performance analytics for coaches and players

Technologies Used

YOLOv8

SegFormer

Python

PyTorch

OpenCV