Architecture Index

The Anatomy of K-RAID

A complete breakdown of the 5-phase pipeline that powers K-RAID. From raw video ingestion and multi-object tracking to adversarial LLM synthesis, explore how each module transforms broadcast footage into tactical intelligence.

Figure 3.1: End-to-End System Architecture

Phase 01Computer Vision

Perception Pipeline

Tracking objects in extreme chaos

Multi-object tracking using YOLOv8 and DeepSORT with custom occlusion-aware Re-ID to maintain raider identity through dense cluster tackles.

YOLOv8DeepSORTOpenCV

Phase 02Geometry

Spatial Normalization

Eliminating broadcast perspective

Extracting 8 court keypoints to compute a homography matrix, projecting 2D camera pixels into a mathematically sound Bird's Eye View (BEV) space.

HomographyMatrix MathNumPy

Phase 03Orchestration

Router Agent

Intent classification & fan-out

A Llama-3 powered router that acts as the central nervous system, translating user queries into deterministic execution plans for the worker nodes.

LangGraphLlama-3JSON Parsing

Phase 04Specialized Agents

Execution Cluster

Analysis, Simulation, and Retrieval

Four distinct Python workers operating in parallel to evaluate defensive compactness, generate alternative trajectories, and retrieve historical precedents.

STGCNFAISSSimulation

Phase 05Verification

Adversarial Synthesis

Zero-hallucination outputs

A strict Reviewer Agent cross-checks the Synthesizer's natural language output against the raw BEV coordinate math to enforce absolute correctness.

LLM SynthesisAdversarial CheckPython