Cross Camera Association

Cross-Camera Association

Cross-camera association combines several weak or heterogeneous signals rather than relying on one cue.

Inputs

For observations from different cameras, the reference implementation uses information derived from:

• appearance similarity;
• corrected time difference;
• spatial distance;
• observation quality;
• sensor modality.

Modality correction

The controlled benchmark applies a modality-dependent appearance-bias correction before cosine similarity. In a learned deployment, the correction can be replaced by a sensor-specific alignment network without changing the observation/graph interfaces.

Clock correction

Camera-local clock offsets are removed before temporal agreement is computed.

Reference association score

The controlled implementation uses a transparent weighted combination dominated by appearance, with temporal and spatial terms, a small heterogeneous-modality corroboration bonus, and a quality gate.

The current controlled reference weights and thresholds are diagnostic constants. They have not been established as generally optimal values for real camera networks.

Clustering

Accepted pairwise associations form an undirected graph. Connected components define cross-camera entity clusters.

This choice is intentionally auditable: a reviewer can inspect the exact pairwise links that caused tracks to merge.

Baselines

The controlled system diagnostic compares:

• time only;
• appearance only;
• naive fixed fusion;
• IncidentGraph reference fusion.

All methods share the graph/evaluation interface to separate association quality from representation semantics.