Real-Time Tennis Analytics with Multi-Modal Vision on Edge Devices

Project overview

Area	Details
Product context	Portable, mounted-camera tennis analysis
Core constraint	Run the complete perception stack on an edge device
Vision tasks	Court keypoints, ball and player tracking, pose, and person re-identification

I built a multi-modal vision stack for live tennis analysis without a cloud dependency. The work combined multiple perception tasks into one on-device pipeline, where latency, compute use, and consistency between modules mattered as much as standalone model quality.

My responsibilities

• Developed court keypoint detection to estimate court geometry and camera perspective.
• Built multi-frame visual models for tennis-ball and player tracking.
• Integrated human-pose features for tennis shot action recognition.
• Extracted ReID features to improve player identity consistency during tracking.
• Optimized and ran the combined pipeline on the mounted-camera edge device.

System architecture

Module	Purpose
Court keypoint detection	Court geometry and spatial calibration
Ball and player detection	Multi-frame detection and tracking
Human pose extraction	Shot action recognition features
ReID feature extraction	Identity consistency during tracking
Edge deployment	Portable real-time inference on the mounted-camera board

Product use

The unified output supports smart coaching, match analysis, automated highlights, and AI-assisted sports training. Keeping inference on-device reduces reliance on network availability and makes the system suitable for portable court-side operation.