Label Management for Multi-Object Estimates via Labelled Multi-Bernoulli Filters
Hoa Van Nguyen, Tran Thien Dat Nguyen, Changbeom Shim, Diluka Moratuwage, Seonggun Joe · 2024
This work addresses the challenging problem of online managing labels from a sequence of unlabelled multi-object estimates, which is a crucial task in multi-object tracking, particularly in noisy environments with unknown and time-varying numbers of mobile objects. The considered problem requires solving a multi-dimensional assignment problem, which is an NP-hard problem. In this paper, we propose reformulating the aforementioned label management problem within the recur-sive Bayesian framework and leveraging the effectiveness of the labelled multi-Bernoulli (LMB) filter to accurately and efficiently assign labels to unlabelled multi-object estimates in real-time. The proposed LMB labelling (LMBL) algorithm is agnostic to the filtering method and is capable of labelling any unlabelled multi-object estimates. Experimental results demonstrate significant labelling performance improvements of our proposed LMBL approach compared to other state-of-the-art methods. This work presents a robust and efficient solution for the critical problem of label management in multi-object tracking applications.