Classification

Classification adds a semantic layer to objects being detected and tracked by the Qb. The resulting object class can be used to only generate alarms for certain types of objects (e.g., in object-based security zones).

See configuration: API definition.

Classification can be performed in two modes:

  1. Size-based classification (enabled by default, using a object-size-based heuristic)

  2. Model-based classification (using a machine learning model)

1. Size-based Classification

If no model is specified, each cluster is assigned to an object size category based on its bounding box surface area.

  • The largest plane of the bounding box is used for size determination.

  • Clusters smaller than the minimum threshold for SMALL objects are discarded.

    Minimum object surface area (unit: \(m^2\))

    The minimum bounding box surface area required for an object to be classified into a size category. These labels are later used for zone state computation (e.g., in object-based security zones).

Category Type Default Unit

SMALL

float

0.01

\(m^2\)

Lower size threshold for small objects. Objects smaller than this threshold are discarded. Valid range: [0.01 … 0.05].

MEDIUM

float

0.16

\(m^2\)

Lower size threshold for medium objects. This value has to be larger than the minimum size for small objects. Valid range: [0.05 … 0.5].

LARGE

float

1.6

\(m^2\)

Lower size threshold for large objects. This value has to be larger than the minimum size for medium objects. Valid range: [0.5 … 5.0].

2. Model-based Classification

If a classification model is specified, size-based labeling is replaced by machine learning-based classification.

Limitations

  • Size-based classification

    • Only considers the largest plane of the bounding box; shape and orientation details are ignored.

    • Fixed thresholds may not generalize across different sensor setups or environments.

  • Model-based classification

    • Requires a pre-trained model; on-device training is not supported.

    • Classification quality strongly depends on the chosen features and training dataset.

    • Only predefined object classes are supported.