Architecture

Architecture

IncidentGraph separates evidence preparation, heterogeneous fusion, and graph reconstruction so that local recognition can evolve without changing the evidentiary graph contract.

flowchart TB
    subgraph P1[Phase I — Evidence preparation]
      A1[RGB / thermal / depth / other sources]
      A2[Tracks and actor tubes]
      A3[Time · world position · quality · evidence reference]
      A1 --> A2 --> A3
    end

    subgraph P2[Phase II — Heterogeneous fusion]
      B1[Activity recognition]
      B2[Appearance descriptor]
      B3[Modality and clock correction]
      B4[Cross-camera association]
      B5[Event scoring and duplicate suppression]
      B1 --> B3
      B2 --> B3 --> B4 --> B5
    end

    subgraph P3[Phase III — Graph reconstruction]
      C1[Entity clusters]
      C2[Event edges]
      C3[Conflict groups and alternatives]
      C4[Missing-evidence objects]
      C5[Timeline + confidence + provenance]
      C1 --> C2 --> C5
      C3 --> C5
      C4 --> C5
    end

    P1 --> P2 --> P3

Loading

Layered view

A. Evidence layer

Carries source-specific facts:

• camera identifier;
• modality;
• track identifier;
• entity type;
• local start/end time;
• world-space location;
• appearance descriptor;
• observation quality;
• immutable evidence reference.

B. Recognition layer

Produces local evidence such as activity probabilities and appearance features. The current MEVA activity study uses an R3D-18-based classifier.

C. Fusion layer

Combines appearance, time, space, modality and quality using a transparent reference association score. Clock and modality corrections are applied before cross-camera clustering.

D. Reasoning layer

Creates entity clusters and event edges, retains competing claims, attaches source evidence, represents sensor gaps, and exposes a timeline rather than only a terminal decision.

Provenance plane

Provenance is not a post-hoc explanation module. Source identity, temporal support, evidence reference, quality and confidence are carried from ingestion into the graph representation.

Computational profile

The transparent reference association compares cross-camera observation pairs and is worst-case quadratic in the number of observations. Conflict construction is also pairwise in the number of event proposals. Temporal blocking, camera-neighbor constraints and retrieval-based candidate generation are natural scaling improvements.