Module ocsort 

OC-SORT: Observation-Centric SORT

This module implements OC-SORT (Observation-Centric SORT), a robust extension of SORT that addresses tracker drift during occlusions by anchoring motion estimation to raw detections rather than Kalman filter predictions.

Algorithm Overview

SORT accumulates prediction errors during occlusions because its Kalman filter integrates noisy velocity estimates when no observation is available. OC-SORT corrects this with two complementary mechanisms:

Observation-Centric Velocity (OCV) Velocity is computed directly from consecutive detections, not from the Kalman filter state. This produces a more reliable momentum signal because it is grounded in actual observations.

Observation-Centric Momentum (OCM) Before IoU matching, a direction-consistency bonus is added to each IoU score based on how well the observed velocity direction aligns with the vector from the track’s last observation to each candidate detection. Matches that are consistent with the track’s momentum receive a higher effective IoU, improving association after missed frames.

Observation-Centric Re-Update (ORU) When a lost track is re-matched after a gap, the Kalman filter is “re-wound” by replaying linearly interpolated observations between the last seen position and the current detection. This corrects the accumulated drift so that future predictions start from an accurate state.

Parameters

Parameter	Default	Description
max_age	30	Frames a lost track is kept alive before deletion
min_hits	3	Consecutive matched frames required to confirm a track
iou_threshold	0.3	IoU threshold for detection-to-track association
delta_t	3	Observation window (frames) used to compute velocity
inertia	0.2	Weight applied to the direction-consistency cost bonus (OCM)

Tuning Tips

• Increase max_age (e.g. 60) when objects undergo long occlusions.
• Increase delta_t for smoother velocity at the cost of responsiveness to rapid direction changes.
• Increase inertia (max 1.0) if objects move at near-constant velocity; lower it for erratic motion.
• min_hits=1 gives immediate track output — useful when detections are already filtered upstream.

References

Observation-Centric SORT: Rethinking SORT for Robust Multi-Object Tracking Jinkun Cao, Xinshuo Weng, Rawal Khirodkar, Jiangmiao Pang, Kris Kitani CVPR 2023 arXiv:2203.14360

Structs

OcSort

OC-SORT tracker.

OcSortTrack

A single tracked object managed by OC-SORT.

Enums

OcSortTrackState

Track lifecycle state for OC-SORT.