Extending IOU Based Multi-Object Tracking by Visual Information (VIOU-Tracker)英文解读

takeaways

Detectors/IOU Trackers are sufficient for simple scenarios.

For false negatives(missing detection), use VOT to fill the gaps

• KCF is enough

---

Motivation

• accurate detection?
  • 
  • 😅not that accurate
  • Missing Detections
  • deep learning and benchmark
• simple trackers
  • SORT=kalman + hungrian
  • IOU-Tracker merely iou, 1kfps
  • drawback: requirement for high recall
• missing detections
  • False Negatives
  • tracking termination/ID preservation

VIOU-Tracker = DETECTOR + SOT(Visual/Single Object Tracking)

---

Method

IOU-Tracker

In real-world ap plications, false positive/negative detections occur and will interfere with the tracking process. Therefore, the resulting tracks are filtered by requiring each track to contain at least one detection satisfying a high detection confidence \(\geq \sigma\) and to have a minimum length of at least \(t_{min}\) frames tmin

given \(\mathbf{b}_{t-1}^{i}\) (track at frame t)

• if no \(\mathbf{d}_{t}^{\cdot}\)(detection, dot superscript means number from 1 to n) has IoU \(\geq \sigma_{active}\)
  • start a SOT Tracker for at most \(\text{ttl}\) frames
• forward and backward
  • NOTE: VOT results outside the 「gaps」 are abandoned

TODO 『2022-03-17 [many SOT tracker exist in the frame can result in ID shifting]?』

Experiments

The MOTA metric is influenced by the number of false positive and negative bounding boxes of the tracks and the number of ID switches.

All three components are weighted equally. This means that the number of ID switches has only a minor impact on the overall MOTA scores.

The better the input detections for the tracker, the less visually tracked bounding boxes need to be inserted by the V-IOU approach.

NOTE: a video is excluded