Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Immortal Tracker: Tracklet Never Dies

Qitai Wang, Yuntao Chen, Ziqi Pang, Naiyan Wang, Zhaoxiang Zhang

TL;DR

The paper identifies premature tracklet termination as the primary cause of identity switches in online 3D MOT. It proposes Immortal Tracker, a simple predict-to-track framework that uses a vanilla $3$D Kalman filter to predict and maintain tracklets when targets go dark. Across Waymo Open Dataset and nuScenes, it achieves state-of-the-art or competitive MOTA/AMOTA with dramatically lower mismatch and IDS, without learned parameters. The results suggest trajectory-based tracklet preservation is a powerful and practical approach for reducing identity switches in real-time 3D MOT.

Abstract

Previous online 3D Multi-Object Tracking(3DMOT) methods terminate a tracklet when it is not associated with new detections for a few frames. But if an object just goes dark, like being temporarily occluded by other objects or simply getting out of FOV, terminating a tracklet prematurely will result in an identity switch. We reveal that premature tracklet termination is the main cause of identity switches in modern 3DMOT systems. To address this, we propose Immortal Tracker, a simple tracking system that utilizes trajectory prediction to maintain tracklets for objects gone dark. We employ a simple Kalman filter for trajectory prediction and preserve the tracklet by prediction when the target is not visible. With this method, we can avoid 96% vehicle identity switches resulting from premature tracklet termination. Without any learned parameters, our method achieves a mismatch ratio at the 0.0001 level and competitive MOTA for the vehicle class on the Waymo Open Dataset test set. Our mismatch ratio is tens of times lower than any previously published method. Similar results are reported on nuScenes. We believe the proposed Immortal Tracker can offer a simple yet powerful solution for pushing the limit of 3DMOT. Our code is available at https://github.com/ImmortalTracker/ImmortalTracker.

Immortal Tracker: Tracklet Never Dies

TL;DR

The paper identifies premature tracklet termination as the primary cause of identity switches in online 3D MOT. It proposes Immortal Tracker, a simple predict-to-track framework that uses a vanilla D Kalman filter to predict and maintain tracklets when targets go dark. Across Waymo Open Dataset and nuScenes, it achieves state-of-the-art or competitive MOTA/AMOTA with dramatically lower mismatch and IDS, without learned parameters. The results suggest trajectory-based tracklet preservation is a powerful and practical approach for reducing identity switches in real-time 3D MOT.

Abstract

Previous online 3D Multi-Object Tracking(3DMOT) methods terminate a tracklet when it is not associated with new detections for a few frames. But if an object just goes dark, like being temporarily occluded by other objects or simply getting out of FOV, terminating a tracklet prematurely will result in an identity switch. We reveal that premature tracklet termination is the main cause of identity switches in modern 3DMOT systems. To address this, we propose Immortal Tracker, a simple tracking system that utilizes trajectory prediction to maintain tracklets for objects gone dark. We employ a simple Kalman filter for trajectory prediction and preserve the tracklet by prediction when the target is not visible. With this method, we can avoid 96% vehicle identity switches resulting from premature tracklet termination. Without any learned parameters, our method achieves a mismatch ratio at the 0.0001 level and competitive MOTA for the vehicle class on the Waymo Open Dataset test set. Our mismatch ratio is tens of times lower than any previously published method. Similar results are reported on nuScenes. We believe the proposed Immortal Tracker can offer a simple yet powerful solution for pushing the limit of 3DMOT. Our code is available at https://github.com/ImmortalTracker/ImmortalTracker.
Paper Structure (11 sections, 2 equations, 5 figures, 9 tables, 1 algorithm)

This paper contains 11 sections, 2 equations, 5 figures, 9 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related works
  3. Method
  4. Experiments
  5. Datasets and Metrics.
  6. Baselines
  7. Implementation Details
  8. Main Results
  9. Ablation Studies
  10. Qualitative Results.
  11. Conclusion and Future Direction

Figures (5)

  • Figure 1: Bird’s Eye View(BEV) tracking visualization of vehicles with their predicted trajectories. For each tracked vehicle, we plot their bounding boxes in the initial frame and predicted(marked in cyan) or ground-truth(marked in red) center locations for the next 30 frames. Small displacements are added to the overlapped trajectories for better visualization.
  • Figure 2: Algorithm Pipeline. (a)We employ an off-the-shelf 3D detector to obtain 3D detections $\mathcal{D}$ from input point clouds. (b)For data association, we compute 3D IoU or GIoU between detected 3D bounding boxes and boxes predicted by Immortal Tracker. Then we perform Hungarian matching based on 3D IoU/GIoU. (c)We use a vanilla 3D Kalman filter(3DKF) for prediction. Based on the outputs of Hungarian matching process, The states of matched tracklets $\mathcal{X}_\text{m}$ are updated by its corresponding detection $\mathcal{D}_\text{m}$. The unmatched tracklets $\mathcal{X}_\text{um}$ will be updated with their predicted object states and the unmatched detections $\mathcal{D}_\text{um}$ will be initialized as new tracklets $\mathcal{X}_\text{new}$. (d)Only the alive tracklets which have left their birth stage and successfully been matched in the current frame are adopted as outputs.
  • Figure 3: Details for mismatch ratio vs $A_\text{max}$ on WOD validation set.
  • Figure 4: Visualization of the remaining wrong association case. Outputs of the wrongly associated tracklet are drawn in cyan. The ground truth bounding boxes are drawn in black and the detection results are drawn in brown. The tracklet is misled by sequential inaccurate or false positive detections from its initial location to another. Such a case is extremely rare in our method.
  • Figure 5: Visualization of a prevented premature termination. In this case the vehicle is not detected in frame 100-112. We plot the estimated bounding boxes of tracklets every two frames. Tracklets with different identities are drawn in different colors. For a better view we do not plot the ground truth bounding boxes in frame 100-112. (a)In CenterPoint++, the tracklet(marked in cyan) is terminated when the object is temporarily not observed, causing an identity switch. (b)While in Immortal Tracker, the initial tracklet is preserved until the object is detected again.