Sheaves are the canonical datastructure for sensor integration

This paper presents a framework for sensor integration that is built on top of sheaves. The use of this framework gives us the following:

  1. A sensor agnostic measure of data consistency (consistency radius)
  2. A sensor agnostic optimization method to fuse data (optimization problem)
  3. A sensor agnostic detection of "systemic blind spots" (sheaf cohomology)

The framework is general in that it allows data fusion from separate sensor modalities.

There are two distinct phases in this framework:

  1. Sensor integration: This is where we build a unified system from individual sensor models.
  2. Sensor data integration: This is where we actually use the data from sensors.

Sheaves

Axiom 1: Entities are quantities that we can observe or measure, for example, the last known aircraft location, or the bearing angle it makes with a radio station at some time, or a satellite image taken at some time.

Axiom 2: Attributes are quantities that are the actual values entities may have. For example, 3D position, which we parameterize as a vector in \(\mathbb{R}^3\) , or an instance in time ( \(\mathbb{R}\) ).

Sensors measure a collection of entities. And when two sensors measure the same entity, we expect them to assign similar attributes (for example, when two sensors measure the velocity of an aircraft, we expect them to both return a vector in \(\mathbb{R}^3\) ).

Axiom N: These sensors form a topology \(T\) on entities \(X\) .

Sensors assign a pseudometric attribute space to an open set \(U \in T\) via \(S(U) : T \rightarrow A\) .

There is a way to go from \(S(U)\) to \(S(V \subseteq U)\) , i.e., \(S(V \subseteq U) : S(U) \rightarrow S(V)\) . This is called a restriction, and it exists, is unique, and is an identity if \(U=V\) .

All of the above give us a presheaf.

If \(S(A)\) and \(S(B)\) agree on \(S(A \cap B)\) , \(S(A \cup B)\) is defined.

This gives us a sheaf .