Depth-Enhanced Observation-Centric Momentum

• Depth-enhanced Observation-Centric Momentum is a tracking framework that integrates per-object depth with observation-centric detection for robust trajectory correction.
• It combines depth augmentation and momentum-like temporal smoothing to improve association, gating, and occlusion management in 3D tracking.
• Research shows that enhanced depth estimation significantly improves performance in camera-compensated and depth-informed tracking systems.

Depth-enhanced Observation-Centric Momentum denotes a family of observation-first tracking formulations in which depth is attached to the observation and used to modulate association, gating, and temporal state correction. In the cited literature, the exact term is not uniformly used. OC-SORT establishes the observation-centric foundation and defines Observation-Centric Momentum (OCM) without depth (Cao et al., 2022). DepthMOT adds end-to-end depth estimation and camera-pose compensation to an observation-centric tracker (Wu et al., 2024). DepTR-MOT turns instance-level depth into a per-query detector output for depth-informed trajectory refinement (Deng et al., 22 Sep 2025). Work on monocular 3D perception further shows that per-object depth estimation is a major factor bounding downstream 3D detection and tracking performance, and that replacing only the depth estimate with a fusion-enhanced alternative yields substantial gains (Jing et al., 2022).

1. Conceptual definition and research lineage

Depth-enhanced Observation-Centric Momentum is best understood as an overview of three ideas. The first is the observation-centric premise: when occlusion or non-linear motion makes long-horizon state extrapolation unreliable, recent detections should dominate corrective updates. The second is depth augmentation: each observation carries either a per-box depth proxy, a per-object center depth, or an instance-level scalar depth. The third is momentum-like temporal smoothing: observation updates are stabilized by exponential averaging, Kalman-style covariance adaptation, or virtual-trajectory re-updating.

The literature organizes these ideas along complementary axes. OC-SORT is a 2D MOT method with no depth term, but it supplies the canonical observation-centric mechanisms: observation-derived direction consistency and observation-centric re-update across occlusion (Cao et al., 2022). DepthMOT moves the observation from 2D image geometry to depth-augmented, camera-compensated detections, with SE(3) warping and depth-based cascaded matching (Wu et al., 2024). DepTR-MOT attaches a scalar depth to each DETR query and uses that quantity directly in association, gating, and pseudo-3D trajectory refinement (Deng et al., 22 Sep 2025). The monocular 3D detection-and-tracking study centered on PRT fusion shows that improving per-object depth is not merely auxiliary; it is a principal bottleneck, and better depth propagates immediately to 3D detection and tracking quality [2206.036