{
  "handle": "ConfidenceCascadeTracking",
  "ref": "ConfidenceCascadeTracking#1973",
  "stub": "1973",
  "semaId": "sema:ConfidenceCascadeTracking#mh:SHA-256:19736fba4c132b16ae36577b7d58b119b1437c16b1ceacb61a484c428fea01a3",
  "layer": "Mind",
  "category": "Inference",
  "gloss": "Match confident detections first, consult appearance only on ties, and let weak detections fill gaps, to keep identities unbroken",
  "mechanism": "Multi-object tracking for interaction analysis, where every lost or swapped identity breaks a chain in a {{second_order_network}}. Each track holds a Kalman state (x, y, s, r, dx, dy, ds): box centre, area, aspect ratio and velocities, with a constant-velocity model, large initial velocity uncertainty (1e4) and small process noise. Each frame: (1) match high-confidence detections (score > 0.7) to all active tracks by IoU distance 1 - IoU with the Hungarian algorithm; (2) when best and second-best IoU candidates are too close, use a re-identification network (ResNet-50 fine-tuned with triplet loss, cosine distance on embeddings) as a tie-breaker; (3) match remaining low-confidence detections (0.1-0.7) only to tracks still unassigned, rejecting matches above a distance threshold; treat scores below 0.1 as background. Unmatched tracks keep their last position (no Kalman drift) and are deleted after 60 frames (1 s), which bridges brief occlusions. Export per-frame centroids and keypoints per track.",
  "invariants": [
    "Low-confidence detections can never displace a track already matched by a high-confidence one",
    "Appearance is consulted only when geometry is ambiguous",
    "A track survives a bounded number of unmatched frames, then is deleted"
  ],
  "failureModes": [
    "Dense clusters (aggression) still produce identity swaps that need manual or rule-based repair downstream",
    "A re-ID model trained on few individuals from one viewpoint generalises poorly to new animals or lighting",
    "Holding position during occlusion mislocates fast movers when they reappear"
  ],
  "ring": 2,
  "tier": 1,
  "aliases": {
    "second_order_network": "SecondOrderInteractionNetwork"
  },
  "buildsOn": [
    {
      "handle": "SecondOrderInteractionNetwork",
      "type": "yields",
      "alias": "second_order_network",
      "url": "/knowledge-graph/SecondOrderInteractionNetwork/",
      "json": "/knowledge-graph/SecondOrderInteractionNetwork.json"
    }
  ],
  "usedBy": [],
  "sources": {
    "research": [],
    "works": [
      {
        "kind": "thesis",
        "label": "Kevin Do, PhD thesis on ant collective behaviour (NCSU, 2026)",
        "note": "Ch. 2.2.4-2.2.5, Eqs. 2.6-2.8; deepSORT/ByteTrack adaptation",
        "url": ""
      }
    ],
    "homePage": false
  },
  "url": "/knowledge-graph/ConfidenceCascadeTracking/",
  "graphUrl": "/knowledge-graph#ConfidenceCascadeTracking",
  "directory": "/knowledge-graph/patterns.json"
}