LumenSport
SPORTS VIDEO INTELLIGENCE · SEP. 2025 — MAY 2026
LumenSport
An end-to-end system that turns raw sports footage into structured analysis and editable video narratives.
Problem
Long sports videos contain many different temporal structures: rallies, possessions, tactical events, individual actions, and emotionally salient highlights. A useful editing system must therefore do more than detect shots—it must interpret the sport, locate meaningful units, preserve temporal context, and assemble clips around a clear editorial intent.
LumenSport organizes these requirements as a layered workflow instead of relying on a single model to make every decision.
System pipeline
Classify
Identify the sport and select sport-specific processing logic.
Segment
Split long footage into rounds, rallies, possessions, or action units.
Understand
Evaluate highlights, tactical events, and player-centric moments.
Orchestrate
Arrange selected material using one of eight editing narratives.
Technical design
Perception and temporal modeling
The repository integrates Swin Transformer for sport/scene classification, SlowFast for round or action segmentation, ConvNeXtV2 for in-round event assessment, VideoMAE for football action recognition, and YOLO with ByteTrack for player detection and tracking.
Editing intelligence
The orchestrator supports highlight, funny, tactical, technical, comparison, knowledge, match-replay, and training narratives. These modes provide explicit editorial objectives for selecting and arranging candidate clips.
Engineering
I built the Python processing backend and a React, Vite, and TypeScript interface. The repository separates video processing, orchestration, player highlights, model utilities, storage integration, and frontend modules.
My contribution
- Led the system architecture and primary implementation.
- Designed sport classification, temporal segmentation, highlight scoring, tactical analysis, player tracking, and reel-generation stages.
- Integrated heterogeneous video, detection, tracking, and multimodal components into one workflow.
- Implemented the backend and the editor/analytics-oriented web interface.
VERIFIED OUTCOME
From research prototype to competition work
The associated “Yuelai Sports · Intelligent Football Platform” received First Prize in the Football Application Innovation Track at the First National Football AI Innovation Exchange Conference.
No additional accuracy, latency, deployment-scale, or user-impact claims are presented because they are not verified in the current project record.