Multi-Object Tracking

ROS2 Node for tracking 3D objects detected around the car.

Overview

This tracker uses a modified version of ByteTrack to track 3D bounding boxes. Core logic draws inspiration from ByteTrackV2 (Zhang et al., 2023). The implementation is built on top of ByteTrack-cpp with modifications made for 3D tracking.

Topics

Subscribed

Published

Parameters

Global Defaults

Prediction (Kalman Filter)

Per-Class Tracking Overrides

When configured, separate BYTETracker instances run per class with independent thresholds and distance metrics. Detections are split by class before tracking and merged afterward with unique track ID offsets to avoid collisions. Classes not listed use the global defaults.

Empty arrays disable per-class tracking (single global tracker for all classes).

Example (cars use strict IOU, persons use permissive CIOU):

class_tracking:
  classes:           ["car",  "person", "truck", "bus"]
  match_thresholds:  [0.95,   0.99,     0.95,    0.95]
  high_thresholds:   [0.40,   0.001,    0.10,    0.10]
  track_thresholds:  [0.10,   0.01,     0.10,    0.10]
  dist_metrics:      ["IOU",  "CIOU",   "IOU",   "IOU"]

Why per-class? Small objects (persons, bicycles) have tiny 3D bounding boxes where frame-to-frame BEV IoU is near zero even for the same physical object. CIOU with a high match threshold bridges this gap. Large objects (cars, trucks) have stable boxes where standard IOU works well and CIOU can cause ID switches.

Acknowledgements

This package uses a modified version of ByteTrack-cpp. The original repository is licensed under MIT License (see THIRD_PARTY_LICENSES/BYTETRACK_LICENSE for details).

ByteTrackV2 paper can be found at:
Y. Zhang et al., “ByteTrackV2: 2D and 3D Multi-Object Tracking by Associating Every Detection Box,” arXiv, Mar. 2023, https://doi.org/10.48550/arXiv.2303.15334