Association metric

Association metric is a matrix of likelihoods $A_{rt}$ between sets of detection reports (vectors) $\mathbf{z}$ and target (predicted) states $\tilde{\mathbf{x}}$. Number of detections and targets generally differ, i.e. the association metric is a rectangular matrix $A_{rt}$.

There are several definitions of the association metrics implemented in the kinematic tracker.

• Mahalanobis distance inside the gaussian function with sum of process and detection covariances.
• Generalized intersection over union (GIoU) for cuboids aligned with cartesian axes.
• GIoU for cuboids oriented with the yaw angle.
• Mahalanobis distance inside the gaussian function with size-modulated covariance.
• GIoU-calibrated diagonal size-modulated Mahalanobis distance inside the exponent function.

The association metrics are defined below. In order to select a metric, we provide a setter method .set_association_metric(...), e.g.

from kinematic_tracker import NdKkfTracker

tracker = NdKkfTracker([2, 1], [2, 2])
tracker.set_association_metric('giou-axes-aligned')

The list of accepted metric names can be obtained from the enumerable class AssociationMetric.

from kinematic_tracker.association.association_metric import AssociationMetric

print([m.value for m in AssociationMetric])

For a quick exploration of the association metric definition we recommend our related project Association studio

Mahalanobis association metric

Name for the setter mahalanobis.

Defined via a so-called Mahalanobis distance @ Wikipedia.

See our discussion of the Mahalanobis association metric

GIoU for cuboids aligned with Cartesian axes

Name for the setter giou-axes-aligned.

Defined via a ratio of intersection and union volumes of cuboids @ Medium.

See our discussion of the GIoU axes aligned metric.

GIoU for yaw-oriented cuboids

Name for the setter giou-yaw.

See our discussion of the GIoU yaw metric.

Mahalanobis with size-modulated covariance

Name for the setter mahalanobis-size-modulated.

This metric is experimental. See our discussion of the size-modulated Mahalanobis.

GIoU-calibrated diagonal size-modulated Mahalanobis distance inside the exponent function

Name for the setter iou-like-two-yaw.

This metric is experimental. See our discussion of the iou-like-two-yaw.