Important note: On September 7, 2026, the KITTI and KITTI-360 servers had to be restored to the backed-up state of August 7, 2026. Results within this period have to be resubmitted.

Method

Graph-enhanced PMB [la] [on] [GePMB]
https://github.com/PeterXu0124/GePMB

Submitted on 13 Sep. 2026 13:18 by
cao cao (cao)

Running time:0.20 s
Environment:GPU @ 2.5 Ghz (Python)

Method Description:
Graph-enhanced PMB is a LiDAR-based online 3D
multi-object tracking method built upon the
Poisson Multi-Bernoulli filtering framework. A
graph neural network is introduced to learn
discriminative affinities between predicted tracks
and current detections, improving data association
in crowded and ambiguous scenes. In addition, a
neural-enhanced motion model combines a CTRA
physical prior with recurrent residual prediction
to better handle nonlinear target motion. A
lifecycle-aware management strategy is further
employed to improve track initialization, missed-
detection handling, and termination. The method
preserves probabilistic PMB inference while
enhancing association, motion prediction, and
track management with learned components.
Parameters:
Detector: PointRCNN
Object class: Car
Frame interval: 0.1 s
Tracking model: CTRA

Detection threshold: 0.8
NMS IoU threshold: 0.1
NMS maximum distance: 10.0 m

Survival probability: 1.0
Detection probability: 0.9
Minimum detection-probability scaling: 0.7
Detection-probability scaling point: 10

Acceleration deviation: 1.0
Turn-rate noise deviation: 0.15
Jerk deviation: 0.5
Measurement deviation: [0.5, 0.5, 1.0, 1.0, 0.5]

MBM gating distance: 4.0 m
PPP gating distance: 4.0 m
Used-measurement distance threshold: 1.0 m
Used-measurement likelihood threshold: 0.5
PPP maximum survival step: 1

Clutter rate: 10
Detected birth intensity: 0.2
Undetected birth intensity: 0.5
Maximum adaptive birth intensity: 0.2

Confidence threshold: 1.2
Track extraction probability threshold: 0.85
Track termination probability threshold: 0.9
Maximum unmatched extraction frames: 5

ID refinement maximum age: 2 frames
ID refinement base distance: 2.5 m
ID refinement distance increment: 1.5 m
Latex Bibtex:
@unpublished{xu2026graphpmb,
title = {Graph-enhanced PMB for LiDAR-based 3D
Multi-Object Tracking},
author = {JingLiang,Xu},
year = {2026},
note = {Manuscript under review}
}

Detailed Results

From all 29 test sequences, our benchmark computes the HOTA tracking metrics (HOTA, DetA, AssA, DetRe, DetPr, AssRe, AssPr, LocA) [1] as well as the CLEARMOT, MT/PT/ML, identity switches, and fragmentation [2,3] metrics. The tables below show all of these metrics.


Benchmark HOTA DetA AssA DetRe DetPr AssRe AssPr LocA
CAR 62.69 % 59.96 % 66.04 % 68.78 % 75.82 % 69.10 % 86.58 % 86.41 %

Benchmark TP FP FN
CAR 27305 7087 3891

Benchmark MOTA MOTP MODA IDSW sMOTA
CAR 67.13 % 84.70 % 68.08 % 327 54.98 %

Benchmark MT rate PT rate ML rate FRAG
CAR 62.15 % 19.08 % 18.77 % 810

Benchmark # Dets # Tracks
CAR 31196 2665

This table as LaTeX


This figure as: png pdf

[1] J. Luiten, A. Os̆ep, P. Dendorfer, P. Torr, A. Geiger, L. Leal-Taixé, B. Leibe: HOTA: A Higher Order Metric for Evaluating Multi-object Tracking. IJCV 2020.
[2] K. Bernardin, R. Stiefelhagen: Evaluating Multiple Object Tracking Performance: The CLEAR MOT Metrics. JIVP 2008.
[3] Y. Li, C. Huang, R. Nevatia: Learning to associate: HybridBoosted multi-target tracker for crowded scene. CVPR 2009.


eXTReMe Tracker