K-RAID
Video-Based Agentic Reasoning.
A complete reasoning framework for Kabaddi. We transform raw broadcast footage into actionable tactical insights by orchestrating multi-object tracking, Bird's Eye View mapping, and a 4-agent LangGraph cluster.
Why We Built This
High Occlusion Chaos
Kabaddi raids are short (5-10s) and incredibly dense. During a tackle, 7 defenders collapse onto 1 raider. Traditional tracking systems instantly fail during these extreme occlusions.
Subjective Analysis
Current analysis relies entirely on manual observation by coaches, making it slow and highly subjective.
Agentic Reasoning
Our objective was to build an Agentic Framework that transforms raw pixels into structured data. By utilizing LangGraph, we enable AI to actually reason over complex spatial gameplay dynamics.
02
You Can't Analyze What You Can't See
Most people think sports analytics is just about plotting player coordinates on a screen. But how do you track a player when they are buried under a 7-man tackle in a chaotic 10x13 meter box?
Traditional computer vision fails completely in Kabaddi. If the camera loses the player in a pile-up, the tracking algorithm instantly dies.
What if the AI didn't just track players, but actually reasoned about where they must be? K-RAID guarantees zero-hallucination tactical analysis, even when players are completely obscured.
03
The Architecture of Reasoning
Because traditional YOLO models fail at extreme occlusion, we built a BEV pipeline. But because 2D matrices still lose players in a pile-up, we needed a semantic layer. The result? A massive LangGraph cluster that deduces missing variables using social-temporal physics and game rules.
Technologies
Tech Stack
The architecture that powers K-RAID's real-time video processing, multi-agent orchestration, and responsive live telemetry interfaces.
- Python
- PyTorch
- HigherHRNet
- ST-CGNet
- OpenCV
- LangGraph
- Multi-Agent Systems
- Google Gemini
- LLMs
- Node.js
- Express
- SQLite
- Ngrok
- WebSockets
- Next.js
- React
- TypeScript
- Tailwind CSS
- Framer Motion
- Three.js
05
Built at Anna University
K-RAID was designed and developed as a B.Tech final year project at the College of Engineering Guindy, Anna University, aimed at proving the viability of multi-agent LLM systems in highly occluded sports analytics.
- 01 Yogasimman R
- 02 Manish M
- 03 Kesav Kumar J
- 04 Shanjai K
Anna University
Explore Architecture Deep-Dive
→
Interested in our research or looking to collaborate?
We completed this project during our B.Tech engineering years at Anna University. Feel free to explore the source architecture or reach out to our team directly.