UCMCTrack: Multi-Object Tracking with Uniform Camera Motion Compensation

Introduction

• 65.8 HOTA and 81.1 IDF1 on the MOT17 test set are reported for UCMCTrack+, with significant competitiveness against state-of-the-art trackers.Figure 1 compares trackers by IDF1, HOTA, and AssA.Source: Figure 1 · p. 1
• Image-plane IoU can fail when camera movement leaves detection and tracking boxes without overlap, motivating ground-plane motion association.The introduction describes ground-plane association as more direct, convenient, and efficient than frame-by-frame camera motion compensation.Source: Introduction · p. 2
• UCMCTrack is a motion-only tracker that avoids IoU-based methods and applies the same camera compensation parameters throughout a video sequence.The paper reports that, when given detections, it exceeds 1000 FPS on a single CPU.Source: Introduction · p. 2

Distance Measures

• Image-plane IoU can fail when boxes do not overlap, while GIoU, DIoU, and BIoU variants still cannot fully capture ground-plane motion patterns under camera movement.Ground-plane methods add depth or bird’s-eye-view distances, but the cited approaches still rely on IoU and do not account for projection uncertainty.Source: Distance Measures · p. 2

Motion Models

• Constant Velocity motion models and other image-plane bounding-box approaches do not reflect targets’ actual ground-plane motion during camera movement.Learned motion methods include LSTM, RNN, transformers, query propagation, and graph-based association, but complex designs can require substantial data and compute and may not meet real-time requirements.Source: Motion Models · pp. 2–3

Camera Motion Compensation

• Camera motion compensation commonly aligns frames using image registration, but current techniques can impose substantial computational overhead on high-resolution video.Examples include ECC, feature matching, optical flow, and affine transformation estimation.Source: Camera Motion Compensation · p. 3

Method

• UCMCTrack augments tracking-by-detection across the motion model, distance measure, and process-noise compensation.The paper says these changes together are intended to provide adaptability and efficiency across tracking challenges.Source: Method · p. 3

Motion Modeling on Ground Plane

• The Kalman constant-velocity model uses ground-plane coordinates, with the bottom-edge midpoint of each image box projected as the observation.The state vector contains ground-plane position and velocity; the projection matrix combines camera intrinsic and extrinsic parameters.Source: Motion Modeling on Ground Plane · p. 3

Correlated Measurement Distribution

• Mapped detector errors form a correlated, non-diagonal ground-plane measurement distribution, which the figure associates with corrected matches.The figure contrasts this with IoU failing when boxes do not intersect and MMD without CMD producing incorrect associations.Source: Correlated Measurement Distribution · pp. 3–4, Figure 3 · p. 4

Mapped Mahalanobis Distance

• Normalized Mahalanobis distance on the ground plane addresses IoU’s failure when detection and track boxes do not overlap.Adding the log determinant of measurement covariance makes association account for measurement uncertainty as well as prediction-measurement discrepancy.Source: Mapped Mahalanobis Distance · p. 4

Process Noise Compensation

• The process-noise model treats camera-induced acceleration as noise in the constant-velocity model, with σx and σy compensating for tilt- and rotation-related motion noise.The process-noise covariance is specified for a two-dimensional constant-velocity Kalman filter.Source: Process Noise Compensation · p. 4

Setting

• The evaluation covers pedestrian MOT17 and MOT20, irregular-motion DanceTrack, and low-rate (10 FPS), high-motion KITTI, using MOTA, IDF1, AssA, and HOTA among its metrics.Source: Datasets · p. 5, Metrics · p. 5

Benchmark Evaluation

• On KITTI, adding CMC did not yield favorable results, possibly because its parameters were inaccurate; UCMCTrack addresses high-speed motion and low-frame-rate detections.Source: Table 3 · p. 6, KITTI · p. 6
• 0.9 HOTA, 0.5 IDF1, and 0.7 AssA on MOT17, and 2.3 HOTA, 3.4 IDF1, and 5.5 AssA on DanceTrack are reported gains for UCMCTrack+ over SOTA methods.Source: MOT17 and MOT20 · p. 6, DanceTrack · p. 6

Ablation Studies on UCMC

• MMD’s position-only distance initially trails IoU, but CMD’s use of distribution information surpasses the IoU baseline, while PNC mitigates camera-induced motion noise.Source: Table 4 · p. 6, Component Ablation · p. 6
• On MOT17, UCMC outperforms the baseline combined with CMC; on DanceTrack, CMC degrades both the baseline and UCMC, while UCMC boosts performance amid minor camera jitter.Source: Table 5 · p. 6, CMC Ablation · pp. 6–7
• Dynamic scenes benefit from a larger process compensation factor, while static scenes call for a smaller one.Source: Figure 4 · p. 7, Influence of Process Compensation Factor · p. 7
• HOTA and IDF1 peak at σm=0.05 and remain relatively stable from 0.04 to 0.1; Y-axis camera-parameter errors matter less than X- and Z-axis errors.Source: Influence of Detector Noise Factor · p. 7, Robustness to Camera Parameters Error · p. 7

• UCMCTrack demonstrates superior performance across a variety of datasets.Source: Conclusion · p. 7
• It uses normalized Mahalanobis distance on the mapped ground plane rather than relying on IoU, with consistent compensation parameters throughout each video sequence.Source: Conclusion · p. 7
• UCMCTrack assumes targets lie on a single ground plane; the paper proposes integrating IoU and ReID as possible extensions.Source: Conclusion · p. 7

Appendix A Pseudo-code of UCMCTrack

• UCMCTrack uses a compact set of parameters rather than frame-by-frame camera compensation and achieves state-of-the-art performance without IoU.The pseudo-code takes a video, detector results, and camera parameters as input, and outputs tracks.Source: Pseudo-code of UCMCTrack · p. 10
• For each frame, a constant-velocity model predicts retained trajectories and their ground-plane distributions.These predictions provide the trajectories used in subsequent association.Source: Pseudo-code of UCMCTrack · p. 10
• Detections are mapped to ground-plane positions and distributions, then compared with predicted trajectories using normalized Mahalanobis distance to form the association cost matrix.The mapped distances populate the matrix used for matching.Source: Pseudo-code of UCMCTrack · p. 10
• The Hungarian algorithm assigns detections and trajectories; unmatched tracks age until the lost-time threshold, while unmatched detections above the confidence threshold initialize tracks.Unmatched tracks are deleted after exceeding the lost-time threshold, and lower-confidence unmatched detections are discarded.Source: Pseudo-code of UCMCTrack · p. 10

Appendix B Mapped Coordinates and Distribution

• A linear camera model relates three-dimensional world coordinates to two-dimensional pixels, with the coordinate definitions identifying world and image axes and a scale factor.The model uses intrinsic and extrinsic camera matrices; the extrinsic matrix incorporates camera extrinsic errors considered in axis-wise ablations.Source: Appendix B · p. 10
• Pixel coordinates are converted to ground coordinates using the inverse camera mapping, with the scale factor defined as the reciprocal of b3.The expressions give x and y as scaled linear functions of pixel coordinates u and v.Source: Appendix B · p. 11
• The method propagates pixel-coordinate errors into ground-plane errors by taking the total derivative of the coordinate mapping.The resulting differential expressions describe how changes in image coordinates affect mapped coordinates.Source: Appendix B · p. 11
• Assuming independent normally distributed detector errors, the method maps measurement uncertainty to the ground plane to obtain Rk.The image-plane covariance uses detection noise factor σm and detected width and height; the resulting covariance is the ground-plane measurement-noise matrix.Source: Appendix B · p. 11

Kalman Filter Model

• The discrete-time filter uses transition and observation matrices, with independent identically distributed Gaussian process and measurement noise.The transition matrix maps the state from time k−1 to k.Source: Kalman Filter Model · p. 11
• The Kalman filter predicts a prior state and covariance, then updates the posterior state and covariance from the observation using the Kalman gain.The section presents the filter as recursive prediction and update equations.Source: Kalman Filter Model · pp. 11–12
• UCMCTrack's observation is ground-plane position [x, y] and its state is [x, ẋ, y, ẏ] under a constant-velocity model.The section references Qk and Rk as the process- and measurement-noise matrices for this model.Source: Kalman Filter Model · p. 12