benchmarks.hota.tracker.build_kalman_tracker

The tracklet centroid state is 4-D ``[x, y, vx, vy]`` (predicted forward each frame); detections carry a 2-D ``[x, y]`` centroid. The association cost is the Mahalanobis chi-squared distance to the *predicted* centroid (projected into the 2-D measurement subspace), gated by ``MotionGate`` at ``max_chi2`` and ``ClassGate``. A plain ``CDist("centroid")`` cannot be used here: the 4-D tracklet mean and 2-D detection mean have mismatched feature dimensions; ``Mahalanobis`` is the primitive that projects between them. ``max_chi2`` bounds the ``MotionGate`` (pairs with a larger chi-squared distance are masked out before assignment). ``jonker_threshold`` bounds the Jonker assignment cost and defaults to ``max_chi2`` (so the gate and the assignment share one rejection radius, the standard SORT setting); passing a looser ``jonker_threshold`` isolates the ``MotionGate`` as the sole cause of a rejection, which the gate-isolation test relies on. ``height`` / ``width`` are accepted for a uniform factory signature but the Kalman state is resolution-independent (pixel centroids). ``q`` / ``r`` are sized for pixel-space centroids (process/measurement noise of order one pixel²), not the normalized-coordinate defaults of ``KalmanCentroid2D``.