MULTI-TARGET TRACKING WITH TRANSFERABLE CONVOLUTIONAL NEURAL NETWORKS Damian OwerkoCharilaos I. KanatsoulisJennifer Bondarchuk

2025-05-02 0 0 1.53MB 5 页 10玖币

侵权投诉

MULTI-TARGET TRACKING WITH TRANSFERABLE CONVOLUTIONAL NEURAL

NETWORKS

Damian Owerko⋆Charilaos I. Kanatsoulis⋆Jennifer Bondarchuk†

Donald J. Bucci Jr†Alejandro Ribeiro⋆

⋆Department of Electrical and Systems Engineering, University of Pennsylvania, PA, USA

†Advanced Technology Labs, Lockheed Martin, Cherry Hill, USA

ABSTRACT

Multi-target tracking (MTT) is a classical signal processing task,

where the goal is to estimate the states of an unknown number of

moving targets from noisy sensor measurements. In this paper, we

revisit MTT from a deep learning perspective and propose a convo-

lutional neural network (CNN) architecture to tackle it. We represent

the target states and sensor measurements as images and recast the

problem as an image-to-image prediction task. Then we train a fully

convolutional model at small tracking areas and transfer it to much

larger areas with numerous targets and sensors. This transfer learn-

ing approach enables MTT at a large scale and is also theoretically

supported by our novel analysis that bounds the generalization error.

In practice, the proposed transferable CNN architecture outperforms

random ﬁnite set ﬁlters on the MTT task with 10 targets and trans-

fers without re-training to a larger MTT task with 250 targets with a

29% performance improvement.

Index Terms—Multi-target tracking, convolutional neural net-

works, transfer learning, stationarity, shift-equivariance

1. INTRODUCTION

Multi-target tracking (MTT) involves estimating the state of multiple

moving targets from noisy sensor data [1]. MTT is an important

task and was originally introduced for aerospace surveillance, e.g.,

to track airplanes using radar for air trafﬁc control [2, 3]. Recently

MTT approaches have been used in various other ﬁelds including but

not limited to robotics [4, 5], health analytics [6, 7], and autonomous

driving [8].

A core challenge of the MTT problem deals with properly han-

dling measurement origin uncertainty (MOU). In particular, it is

unknown whether measurements come from targets or clutter and

whether or not existing targets are miss-detected or have disap-

peared. Joint probabilistic data association (JPDA) ﬁlters handle

MOU by creating posterior mixture distributions per target, based

on marginal association probabilities. Multiple hypothesis tracking

(MHT) approaches, on the other hand, maintain multiple data asso-

ciation decisions per target over a short history. Other methods use

global nearest neighbor tracking, which is a memoryless version of

MHT, and is used in many practical systems.

Random ﬁnite sets (RFS) [9, 10] is a large class of MTT ap-

proaches that provide a Bayesian analysis of the MTT problem and

have shown remarkable MTT performance. They were ﬁrst pop-

ularized for unlabeled multi-object state estimation via the Proba-

bility Hypothesis Density (PHD) [11, 12] and Cardinalized Prob-

ability Hypothesis density (CPHD) ﬁlters [13]. Labeled RFS ﬁl-

ters [14] were introduced to better approximate the optimal Bayesian

ﬁlter. The Generalized Labeled Multi-Bernoulli (GLMB) ﬁlter [15]

models the target states as a mixture of hypotheses, each represent-

ing a possible combination of target labels. Each hypothesis con-

tains mutliple single-target probability densities. The Labeled Multi-

Bernoulli (LMB) ﬁlter [16] reduces the complexity of the GLMB. It

uses a single mixture where each component represents a possible

measurment-target association. The LMB and GLMB ﬁlters pro-

vide excellent target tracking performance, but widespread adoption

is constrained by computational scalability, especially with multiple

sensors. Efﬁcient computing approaches for these ﬁlters are an area

of active reserach [17–19].

Deep learning approaches promise to alleviate computational

complexity problems. Transformer-based models have comparable

performance to RFS ﬁtlers on simple problems and provide state-of-

the-art results on more complex tasks [20, 21]. However, their scal-

ability is limited by quadratic runtime with respect to the number of

targets and measurements [22].

In this paper, we propose a novel transfer learning approach to

perform large-scale MTT with multiple sensors. In particular, we

recast the MTT problem as an image-to-image prediction task and

train a convolutional neural network (CNN) model to perform MTT.

To overcome the scaling limitation, the CNN model is trained on

small windows of the tracking area, but executed on much larger ar-

eas. The runtime of the model is proportional to the tracking area,

enabling MTT at previously intractable scales. The approach is sup-

ported by theoretical analysis that provides a bound on the gener-

alization performance of CNNs to large signals. Numerical exper-

iments on several MTT scenarios showcase the effectiveness of the

proposed approach. In particular, we show that training can be per-

formed in 1km2window and the performance does not degrade as

we scale up. Instead it improves by 29% on a 25km2window.

2. PROBLEM OVERVIEW

Let xn,i ∈Xbe a vector that represents the state of the i-th target

at time nin some state space X. In numerous applications, the state

of each target is a vector xn,i ∈R4of positions and velocities given

by equation (1).

xn,i =pn,i vn,i(1)

where pn,i,vn,i ∈R2are the two-dimensional position and ve-

locity respectively. Then, the multi-target state at time kis a set

Xn={xn,1, ..., xn,|Xn|}, where |Xn|is the possibly time vary-

ing cardinality of the multi-target state. Between each time step the

multi-target state can change in three ways. First, the states of indi-

vidual targets evolve according to an unknown model. Second, ex-

isting targets may die with probability pdeath. Third, new targets may

arXiv:2210.15539v4 [eess.SP] 25 Jul 2023

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

MULTI-TARGETTRACKINGWITHTRANSFERABLECONVOLUTIONALNEURALNETWORKSDamianOwerko⋆CharilaosI.Kanatsoulis⋆JenniferBondarchuk†DonaldJ.BucciJr†AlejandroRibeiro⋆⋆DepartmentofElectricalandSystemsEngineering,UniversityofPennsylvania,PA,USA†AdvancedTechnologyLabs,LockheedMartin,CherryHill,USAABSTRACTMulti-target...

展开>> 收起<<

MULTI-TARGET TRACKING WITH TRANSFERABLE CONVOLUTIONAL NEURAL NETWORKS Damian OwerkoCharilaos I. KanatsoulisJennifer Bondarchuk.pdf

共5页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

MULTI-TARGET TRACKING WITH TRANSFERABLE CONVOLUTIONAL NEURAL NETWORKS Damian OwerkoCharilaos I. KanatsoulisJennifer Bondarchuk

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: