2020 IEEE International Conference on

Multisensor Fusion and Integration for Intelligent Systems (MFI)

Virtual Conference, Sept. 14-16, 2020

Bayesian Extended Target Tracking with Automotive Radar using

Learned Spatial Distribution Models
Jens Honer? and Hauke Kaulbersch?,†

690
Abstract— We apply the concept of random set cluster
processes in combination with a learned measurement model
to extended target tracking. The spatial distribution of mea- 685
surements generated by a target vehicle is learned via a
variational Gaussian mixture (VGM) model. The VGM is then
680
interpreted as the measurement likelihood of a Multi-Bernoulli
(MB) distribution. We derive a closed-form Bayesian recursion

[m]
for tracking an extended target by the use of random set cluster 675
process. This formulation is particularly successful for sparse
and noisy measurements, and is applied to automotive Radio ground truth
670
Detection and Ranging (RADAR) detections. Last, we provide initial ground truth
a large-scale evaluation of our approach based on the data radar detections
published in the Nuscenes data set. 665 radar clutter detections
Robust performance of Advanced Driver Assistant Sys-
260 270 280 290
tems (ADAS) and Automated Driving (AD) depends to a
[m]
large extent on the robustness of the environment perception.
Modern high-resolution sensors, such as automotive Light (a) Data-starved scenario in automotive RADAR tracking. Data
Detection and Ranging (LIDAR) and RADAR, yield multiple from Nuscenes [1], token: 04500c91262e487a9f0a1de91b0f68b5
measurements per target and hence give rise to the so-called 10
extended target (ET) problem: Measurements generated by 8
percentile05
an extended target only provide indirect information for the percentile95
6
counts

median
target position and higher-order state parameters such as the average
4
velocity and acceleration. The performance of an extended
2
target tracking algorithm is thus strongly tied to its ability to
capture this indirect relation. 0
0 100 200 300 400 500 600
The use of Gaussian distributions was proposed in the trace index
seminal random matrix model [2] and is a straight forward (b) Statistics of detections per frame over the selected 650 traces.
but extremely efficient way to model the target extent. This
model has been and still is successfully applied to a variety of Fig. 1: Example data and statistics for data-starved tracking.
different problems, e.g. pedestrian tracking via RADAR or
LIDAR. Later extensions estimate the number of expected
measurements by use of a Gamma-distribution (Gamma- of LIDAR point clouds, tracking the same vehicle using
Gaussian Inverse Wishart (GGIW) [3]), or combine multiple 79 GHz RADAR will result in considerable ambiguities,
coupled random matrix model to represent more complex since measurements can be extremely sparse and have a
distributions tailored to specific targets [4]. However, the relatively high spatial uncertainty. See Fig. 1 for reference.
use of multiple scattering centers results in an ambiguity The problem of unknown assignments is canonically
for the source of each measurement: Different assignments rooted in multi-target tracking. While an exhaustive treatment
between measurement source and measurement can yield of the association problem is computationally intractable,
vastly different results for the posterior target pose. To the a multitude of approximation schemes have been proposed
authors’ knowledge, this problem was first investigated in [9]. Especially the introduction of Random Finite Set (RFS)
the Random Hypersurface Model (RHM)[5], [6], [7], [8]. methods [10] led to a deeper understanding of the problem
The severity of this effect is of course tied to (i) the sensor and paved the way for new approximation methods and
properties and (ii) the target in question. While tracking a related filters. A comprehensive overview of the relation
vehicle using a LIDAR results in relatively little ambiguity between heuristic and set-based approaches can be found
in most situations1 due to the high precision and density in [11] and [12].
In our previous work [13] we exploited this relation and
? Valeo Schalter und Sensoren GmbH, Bietigheim-Bissingen, Germany proposed the use of an probabilistic multi hypothesis tracker
[email protected]
† Data (PMHT) to account for the assignment ambiguities. For a
Fusion Group, Institute of Computer Science, University of Göttin-
gen, Germany. given measurement the PMHT approximates the complete
1 There are exceptions though. spectrum of associations per scattering center and reduces

978-1-7281-6422-9/20/$31.00 ©2020 IEEE 316

Authorized licensed use limited to: Shanghai Library. Downloaded on February 20,2021 at 04:51:26 UTC from IEEE Xplore. Restrictions apply.
it to a single synthetic measurement per hypothesis. The To avoid any clutter in the notation we will drop the time
synthetic measurement is a weighted combination of the index t in the remainder of the paper. The main difficulty in
original measurements, in which the weights encode the ETT is that changes in measurement data can be either at-
multi-target association probabilities and are calculated via tributed to a change of the target pose, or to the measurement
an EM-procedure under the assumption that individual scat- originating from a different scattering center or, most likely,
tering centers generate measurements following a Poisson a combination of both. This ambiguity is more pronounced
Point Process (PPP). In contrast to previous works that use the more data-starved the scenario is. In turn, such situations
handcrafted target models[14], [15], [16] we applied the require detailed a priori knowledge in the target model. With
PMHT in combination with a VGM [17]. Although the ap- the availability of large-scale data sets, e.g. KITTI data set
proach performed quite well, especially compared to greedy [19] and Nuscenes [1], machine learning algorithms can be
association algorithms, the PMHT had difficulties recovering easily deployed in order to capture minute details in the
once it settled onto a wrong assignment of measurements sensor data. Here we follow the approach proposed in [17] to
to scattering centers. The combination of sparse data and use a learned VGM to model the predicted likelihood p(z|x)
changing scattering centers, e.g. due to maneuvering targets, for an extended target, namely
results in a relatively high frequency of such situations. X
The algorithmic root cause of this behavior seems to be a p(z|x) = πi N (z; µi (x), Σi (x)), (3)
combination of the uni-modality of the model and a general i∈I

tendency of the PMHT to be over-confident in its estimation. with I the index set and the Gaussian components defined
Scenarios as depicted in Fig. 1a can profit from a multi- by the corresponding prior weight πi and a Gaussian with
modal distribution of the target state, with different modes state with mean µi (x) and covariance Σi (x).
driven by different association histories. The natural way
to create such a model is the use of a random set cluster II. R ANDOM S ET C LUSTER P ROCESS FOR ETT
processes [18]: We model the doubly-hidden target state In the following section we are going to map the extended
as the parent processes that conditions the child process target tracking problem description to the RFS formulation.
encoding the distribution of the scattering centers and emits We will follow the general ideas of [18], and refer to the
measurements. This formulation allows for a vast amount of original work for explicit details regarding the derivations.
freedom in model design, and even supports the option for For details regarding the Probability Generating Functional
online adaptation of the child distribution. In the following (PGFL) formulation we refer to [20], [10] and references
we concentrate on the extremely data-starved case in which therein.
even multiple subsequent frames do does not provide suffi-
cient information to gauge the full target distribution. This A. Definition of parent and child process
situation often occurs in automotive RADAR at medium dis- The PGFL for tracking a single target with pdf p(x) is
tances. Therefor we follow [17] and our own work [13] and linear in the test function h̄(x), namely
use a data-driven approach to learn a stationary distribution
for the child process, in which we accumulate training data GETT [h] = h̄(·), p(·) . (4)
to learn a VGM to model a predictive likelihood.
In
R this work we use the bracket notation for the inner product,
I. P ROBLEM F ORMULATION dx h̄(x)p(x) = hh̄(·), p(·)i, and sometimes add the inte-
In the extended target tracking (ETT) problem a single gration variable as a subscript to the bracket. An extension to
target generates multiple measurements wl per time frame t, incorporate the existence probability of the target is straight-
forward2 , but omitted due to spatial constraints. The target
Wt = {wl }N
l=1 .
t
(1) state x cannot be measured directly, but is only accessible
Individual measurements are corrupted by noise, and the via measurements from scattering centers that interact with
number of measurements can change from frame to frame. the sensors. The scattering centers themselves are distributed
Moreover, measurements do not originate from the target over the extended target’s surface and volume and encode
alone, but also from unmodelled events, so that the full set the shape information. In the notation of random sets the
of measurements Zt = Wt ∪ Ct within a given time frame scattering centers are a child process that is conditioned on
consists of target generated measurements Wt and clutter the target state, i.e. its parent process. For our application
measurements Ct . we model the child process as a MB of the form
The general goal is to estimate a target state xt over time Y
1 − ri (x) + ri (x)hh, f i (: |x)i (6)


Gchild [h|x] =
given the measurements Zt . In ETT xt typically consists of
i∈I(x)
the kinematic parameters xkin that contain, position, velocity,
and higher order time derivatives, and extend parameters xext 2 The linear PGFL changes to
that describe the target shape. In the following we will drop
this distinction and only use the stacked state vector GETT [h] = 1 − r + r h̄(·), p(·) . (5)
T with r the existence probability of the target. See, [21] or [20] for a detailed
xt = xTkin,t , xText,t

(2) introduction.

317

Authorized licensed use limited to: Shanghai Library. Downloaded on February 20,2021 at 04:51:26 UTC from IEEE Xplore. Restrictions apply.
with I the index set of the MB, and ri and f i the existence C. Posterior via Bayes
probability and pdf of an individual Bernoulli component, The posterior PGFL is then given by Bayes theorem [20],
respectively. The hierarchy between parent process Gparent
δZ F [g, h̄, h]

and child process Gchild h| · is given by g=0
G[h̄, h] = . (15)
δZ F [g 0 , h̄0 , h0 ]
   
Gparent h̄Gchild [h|·] = h̄Gchild h| · , pparent (·) x (7) g 0 =0,h̄0 =1,h0 =1

see e.g. [10] for details. In order to derive a closed form In order to evaluate eq. (15), we calculate the derivative of
expression for the Bayes correction we will assume the (14) with respect to the set of measurements Z by use of the
parent pdf to follow a multi-modal distribution, so that product rule [23]. This directly yields
X
Wj pjparent (·).
!
pparent (·) = (8) X
C
j∈J
δZ F [g, h̄, h] = δZ\W G [g]
W⊆Z
+ (16)
P
with mixture weights Wj , j∈J Wj = 1. It follows directly * !
that ×
 child 
h̄ δW Gchild hGZ [g| :, ·]| · , pparent (·) .
 
Gparent h̄Gchild [h|·]
D E
(9)
X
Wj h̄Gchild h| · , pjparent (·) . By use of eq. (6) we find the derivative of the child process
 
=
j∈J w.r.t. the measurement set to be
δW Gchild hGchild
 
In turn we will add the parent mode index j to the child Z [g| :, ·]| · g=0
and suppress the explicit conditioning on the random state X X Y
0 (17)
variable x to avoid notational clutter. Ultimately, this yields = Wj Fj,i [0, h|·]
U
Y j∈J W= i Wi i∈Ij
Gjchild [h|x] = 1 − rj,i + rj,i hf j,i , hi .


(10)   0
i∈Ij with Fj,i [g, h|·] = Gchild hGchild
Z [g| :, ·]|· and Fj,i [0, h|·] =
δWi Fj,i [g, h|·]|g=0 . In the case of a Bernoulli measurement
Note that (10) together with (9) formally resembles an Multi-
process the derivative of the child PGFL w.r.t. Z is
Bernoulli Mixture (MBM). 
B. Measurement Process for Parent and Child Fj,i [0, h|·]
 if Wij = ∅
0
We model the measurement PGFL for the parent as a prod- Fj,i [0, h|·] = rj,i hpD φ(Wij | :, ·), f j,i if |Wij | = 1
if |Wij | > 1

uct PGFL of the individual child measurement processes3 0

Y (18)
Gparent
Z [g|·, Ξ] = Gchild
Z [g|ξ, ·] (12)
ξ∈Ξ Note that apart from the initial modeling assumptions,
with Ξ the RFS of the children. The PGFLs of the child namely a multi-modal parent distribution, a MB child process
measurement process are defined as and an MB measurement process, no further approximations
have been introduced. We will change that now by assuming
Gchild
Z [g| :, ·] = qD,child (:) + pD,child (:)LZ [g, :, ·] (13) that the child process is known given the state of the parent.
with pD,child and qD,child = 1 − pD,child the detection proba- The intention is to learn the child process a priori given a
bility and missed detection probability for a child component. larger data set to avoid the need to gather the data online.
LZ [g, :, ·] is the conditional measurement set likelihood for Hence, there is also no need to correct the child process,
a given child. a Bernoulli process hg(z0 ), φ(z0 | :, ·)i with which in turn allows us to marginalize over the child PGFL
the measurement likelihood φ(z| :, ·) is a typical choice for in eq. (16). After some lengthy calculation that is omitted
LZ [g, :, ·], and implies that each child can only be assigned a due to spatial constraints we find the posterior of the parent
single measurement. With the clutter PGFL denoted by GC process
and modelled as a PPP the joint PGFL is given by
P
post j∈J Wj Φj [Z|x]pj (x)
p (x) = DP E (19)
F [g, h̄, h] W Φ [Z|·], p (·)
j∈J j j j
h i (14)
= GC [g]Gparent h̄Gchild hGchild

Z [g| :, ·]| · .
with the short-hand
See e.g. A. Swain [22] for a detailed derivation. X Y
0
Φj [Z|x] = F̄j,i [0, 1|x] (20)
3 Note that in (12) we omitted the option for a missed detection of the whole U
i Wi =W⊆Z i∈Ij
target and only allow for this option purely by missed detections of the
children. The underlying rational is that individual children are sufficiently and
independent from each other so that there is no common root cause for a 
missed detection of all children at once. However, this assumption might 
1 − ri + ri qD , f j,i (: |·) if Wij = ∅
be violated, e.g. due to occlusions or sensor blindness. On the other hand,  j,i
0
this generalization can be easily introduced by changing (12) to F̄j,i [0, 1|·] = ri pD φ(w|:,·),f (:|·) if Wij = {w}
 κ(w)
Gparent
Y
Gchild if |Wij | > 1

Z [g|·, Ξ] = qD,parent + pD,parent Z [g|ξ, ·]. (11)

0
ξ∈Ξ
(21)

318

Authorized licensed use limited to: Shanghai Library. Downloaded on February 20,2021 at 04:51:26 UTC from IEEE Xplore. Restrictions apply.
where we factored out the clutter factor in (21). Note that Moreover, note that x itself is a random variable, and in turn
while the summation over all associations encountered in (23) describes in fact a distribution over transformations.
(20) is conditioned on the parent state, the conditioning Ultimately, we find the joint measurement and child PGFL
is constant over a given parent mode j. The problem of from the learned model to be
finding all associations between different modes of the parent 
decouples, and allows for a parallel implementation. 1 − γi πi j

 if Wij = ∅
j
0
F̄j,i [0, 1|·] = γi πi N (w;µ̄κ(w)
i (x),Σ̄i (x))
if Wij = {w}
U weight of the j-th
From (19) we find the relative posterior
mode and a given association mapping i Wi = W ⊆ Z is

if |Wij | > 1

0
post
Wj| U (24)
i Wi =W⊆Z

with µ̄ji (x) and Σ̄ji (x) defined in (23).

* +
Y
0 (22)
∝ Wj F̄j,i [0, 1|·], pj (·)
i∈Ij
U
i Wi =W⊆Z III. C LOSED F ORM GM I MPLEMENTATION FOR THE
The summation over all associations in (20) can be efficiently PARENT P ROCESS
approximated within a Gibbs sampling procedure [24] or a In this section we will describe a Gaussian mixture im-
stochastic optimization approach [25]. plementation of the parent process. We write the prior of the
parent pdf as
D. Static Child Process and VGM X
Following [13], we learn the VGM on a unit rectangle, i.e. p(x) = Wj N (x; xj , cj ) (25)
j∈J
all data is rotated and scaled by the ground truth oriented
bounding box. However, the application of the VGM is Then, by (19) the posterior parent pdf is
slightly different to the PMHT implementation in [26]. X
First, we note that since we use sensor data to learn the ppost (x) ∝ Wj Φj [Z|x]N (x; xj , cj )
child process, we are only able to learn a combination of j∈J
 
both the measurement and the child process. Hence the VGM X X Y
0 (26)
will provide an approximation to hhpD φ(W| :, ·), f j,i (: |·)i = Wj  F̄j,i [0, 1|x]
rather than individual expressions for φ or f . Therefor, we j∈J
U
i Wi =W⊆Z i∈Ij
may as well choose a simple child pdf of the form f i (ξ|x) = × N (x; xj , cj )
δ(ξ − µi (x)) that allows for easy integration.
0
Second, the VGM learns a pdf rather than a MB distribu- with F̄j,i [0, 1|x] given by eq. (24). We use an Unscented
tion. The sum of weights πi is normalized to one, which Transform (UT) with sigma points χσj in order to handle the
leaves us with a missing factor that yields the expected non-linearity in (26). For the j-th parent component we thus
number of generated measurements4 . The correct method find the expected measurements and related covariances via
of resolving this problem would be to modify the learning σ
procedure, which is beyond the scope of this publication. Yj,i = µ̄i (χσj ) (27)
X
On the other hand, the maximum weight within the VGM ȳj,i = wσm Yjσ (28)
is 0.193, with a 95 percentile at 0.105, while the average σ
number of detections per frame and mode γi is 2.56. A crude
X
cj,i σ
wσc (Yj,i σ


yy = − ȳj,i ) ⊗ (Yj,i − ȳj,i ) + Σj,i (29)
approximation of an actual MB density may be obtained σ
by using the VGM weights times the average number of
detections for a given mode due to γi πi < 1. Following [24], we use a Gibbs sampling U algorithm to
Third, the conditioning of individual child components on approximate the sum over all partitions i Wi = W ⊆ Z
the parent mode pj (x) is performed via the transformations given a parent mixture component j. The sampling weights
according to eq. (22) are easily obtained by replacing µ̄i (x)
µ̄ji (x) = Rφ(x) S(x)µji + t(x) and Σ̄ji (x) by ȳj,i and cj,i
yy , respectively, and pj (x) =
(23) N (x; xj , cj ). For notational reasons it is convenient to
Σ̄ji (x) = Rφ(x) S(x)Σji S(x)Rφ(x)
−1
introduce the association mapping U function θ : Ij → Z
with t the translation relative to the parent position, Rφ that is equivalent to the partition i Wi = W. We further
the rotation matrix by the orientation of the parent φ and define the stacked measurement vector for all measurements
S = diag[l, w] a scale matrix defined by the bounding box that have been assigned to any of the children under the
length l and with w of the target. Note that we keep the association mapping θ
discrete parent index j for µji and Σji to allow for the use M
different VGMs for different target poses w.r.t. the sensor. wj|θ := w (30)
w∈Wij 6=∅
4 Since we only have access to the combination of child process and L
measurement process the equivalent to the cardinality in the multi-target The operation denotes vertical vectorial concatenation.
state is not accessible. Equivalently, we define the stacked predicted measurement

319

Authorized licensed use limited to: Shanghai Library. Downloaded on February 20,2021 at 04:51:26 UTC from IEEE Xplore. Restrictions apply.
vector via also depend on the aspect ratio, and are learned by a simple
σ
M counting approach.
Yj|θ := µ̄i (χσj ) (31)
i:Wij 6=∅
C. Pruning and Merging
X The number of modes in the parent process will grow ex-
ȳj|θ = wσm Yj|θ
σ
(32)
ponentially. This behavior is typical for random set methods
σ
and can be handled efficiently with pruning and merging
and corresponding correlations methods. In this work we first merge modes that are similar
and second prune mixture components that fall below a
X
j|θ
Cxy = wσc (χσj − x̄j ) ⊗ (Yj|θ
σ
− ȳj|θ ) (33)
σ
defined threshold.
X  
j|θ
Cyy = wσc (Yj|θ
σ σ
− ȳj|θ ) ⊗ (Yj|θ − ȳj|θ ) + Σ̄σj|θ D. Estimator
σ The state estimation is performed by taking the component
(34) of (26) with maximum posterior weight.
m σ
P P
with x̄j = and Σ̄j|θ = σ Σj,i . The corrected
σ w σ χj V. E VALUATION
parameters of the j-th Gaussian mixture component of the A. Data Set
parent process under the association mapping θ is then given
We evaluated our algorithm with RADAR data from
by
the Nuscenes [29], [1] data set. The Nuscenes data set is
xpost
j = xj + Gj|θ εj|θ (35) comprised of 1000 annotated scenes of 20 seconds. The
recording vehicles are equipped with a sensor suit of 1
cpost
j = cj − j|θ T
Gj|θ Cyy Gj|θ (36)
LIDAR sensor, 5 RADAR sensors, 6 cameras, an IMU/GNSS
j|θ j|θ −1 system. In this work we focus on RADAR data as an
with residual εj|θ = w̄j|θ − ȳj|θ and gain Gj|θ = Cxy Cyy .
example of an extremely data-starved scenario. The radar
IV. I MPLEMENTATION DETAILS sensors are emitting at 77 GHz with a 13 Hz update rate.
We apply the proposed model to vehicle tracking prob- In order to evaluate the tracking performance of our model
lem with automotive RADAR. To this end, the vehicle is we extracted single-target tracks for individual vehicles by
abstracted via an oriented bounding box with length l, width using the ground truth data within the data set. Note that
w and orientation φ as well as the velocity in the direction of we applied a gate around the ground truth, namely 1.2 in
the orientation v and the yaw rate ω. In turn, the dynamic part x and y in scaled dimensions. Thus the data provided to
of parent state vector is defined as x = [x, y, φ, v, ω, l, w]T the tracking algorithm may contain clutter measurements and
with x, y the Cartesian position. Note that the position [x, y] emulates imperfect segmentation results. Tracks were filtered
refers to the center of the rear axle and not the center to remove pathological cases, e.g. tracks with only three
of the bounding box. This choice is of course due to the frames that contain RADAR data or tracks without any data
application of vehicle tracking. The number of parent mixture at all. Ultimately, this resulted in a selection of 650 tracks and
components is set to J = 50, and we use up to 100 Gibbs a total of 170.887 frames that have been evaluated. Details
samples per parent mixture component. can be found in the supplementary material.
B. Baseline and Metrics
A. Motion Model
We compare our algorithm with our PMHT-based ap-
The target follows a standard coordinated turn model proach [13] and our spline-based extended target tracker [26].
[27] and uses a simple Unscented Kalman Filter (UKF) All algorithms were run in python [30] and share parameters
implementation to handle the non-linearity. for the motion and measurement model. In particular, we
B. Measurement Model and Model Training use σa = 1.5 m/s2 , σω = 5.0 · 10−2 rad/s2 for the
motion model, σl,w = 10−2 m/s for the size, and detection
Following our previous work [13] we define the learned
probability pD = 0.99 with a constant clutter density of
model in a unit square. Given a set of measurements in sensor
κ = 1 · 10−5 m−2 .
coordinates, we scale, translate and rotate the unit square and
We use the altered WSD proposed in [31], with a discrete,
the associated VGM to the size and pose defined by each
equally weighted set of eight points located at the corners and
target mode. The vector from the sensor to the target defines
face centers of the ground truth and estimation rectangles.
which of the discrete learned models is used. Note that this
Thus, the reported values refer to
vector may differ between modes. The explicit measurement
! p1
equation is given by (23). 8
1X i σ(i) p
The VGM itself is trained via a variational estimation dWSD = min kp − px̂ k , (37)
σ n i=1 x
model (sklearn.mixture.BayesianGaussianMixture) from the
scikit-learn Gaussian Mixture Library [28], version 0.22.2. with p = 1, σ(i) the ith permutation from the set of all
The number of components is fixed to 20, and we use a permutations. For the spline model we use the enclosing
hard binning to create eight conditionals depending on the rectangle to define the points rather than the points on the
aspect angle. See [13] for details. The scattering rates γi boundary defined by the spline.

320

Authorized licensed use limited to: Shanghai Library. Downloaded on February 20,2021 at 04:51:26 UTC from IEEE Xplore. Restrictions apply.
 8
std
percentile WSD is lower by a factor of 0.674. Additionally
percentile95 we selected two example scenarios to highlight the behavior
6
median of the different algorithms. To avoid clutter in the scenario
average
plot we only plot the data every 20-th frame. More scenarios
WSD

4
can be found in the supplementary material.
2

0
0 100 200 300 400 500 600
radar detections
trace index 2,060 spline model
(a) Statistics over our spline algorithm with a hand-crafted mea- PMHT model
surement model [26]. proposed
2,040

[m]
8 ground truth
std initial ground truth
percentile95 2,020
6
median
average
WSD

4 2,000
400 420 440 460 480 500 520
[m]
2
(a) Scenario 1, token: 8f96c3ad29fb48bcbf96258d453390e6
0 In this scenario the failure of the greedy association in the spline
0 100 200 300 400 500 600
model is clearly visible. In contrast, both the PMHT model and the
trace index
proposed model provide good good performance.
(b) Statistics over our PMHT algorithm with learned measurement
model [13].
8
std
2,000
6 percentile95 radar detections
median
[m]

spline model
average
WSD

PMHT model
4
1,980 proposed
ground truth
2 initial ground truth

380 400 420 440 460 480

0
0 100 200 300 400 500 600 [m]
trace index
(b) Scenario 2, token: 03d1b8e6b40d4b41a836a88a28eaa19e
(c) Statistics over our proposed algorithm with learned measurement Again the greedy association of the spline model results in a poor
model. estimation. The over-confidence of the PMHT model results in
diverging estimates, especially at the end of the trace. This behavior
Fig. 2: Statistics of the Wasserstein distance (WSD) over is not visible in the proposed method.
a selection of 650 traces from the Nusenes data set. The
proposed algorithm shows a clear performance improvement radar detections
over our previous proposals. Note that for better readability spline model
the traces are ordered according to ascending average WSD 940 PMHT model
for each algorithm individually. proposed
ground truth
initial ground truth
[m]

920
C. Results
An overview of the results can be found in table I and
Fig. 2. In summary we find that the average WSD of our 900
proposed algorithm compared to our previously proposed 1,680 1,700 1,720 1,740
PMHT approach is lower by a factor of 0.632, while the 95 [m]

(c) Scenario 3, token: 2b37a4a1b27849dcba88f84a19c39848

Algorithm Spline [26] PMHT [13] proposed The low ambiguity of the RADAR data in this particular scenario
Average WSD 0.763 0.551 0.348 results in good estimation performance for the spline model. Inter-
Median WSD 0.541 0.431 0.256 estingly, the PMHT model has difficulties in this scenario.
95 percentile WSD 2.29 1.28 0.863
Fig. 3: Detailed view on the performance of the different
TABLE I: Summary average WSD statistics over all 650 algorithms on two traces within the Nuscenes data set. Note
traces. that we only plot every 20-th frame for better readability.

321

Authorized licensed use limited to: Shanghai Library. Downloaded on February 20,2021 at 04:51:26 UTC from IEEE Xplore. Restrictions apply.
 VI. C ONCLUSION [18] A. Swain and D. E. Clark, “Extended Object Filtering using Spatial
Independent Cluster Processes,” 2010 13th International Conference
We presented a closed form Bayesian algorithm for ex- on Information Fusion, pp. 1–8, 2010.
tended target tracking and demonstrated it in combination [19] A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vision meets Robotics:
with a learned measurement model on automotive RADAR The KITTI Dataset,” International Journal of Robotics Research
(IJRR), 2013.
data in extremely data-starved scenarios. We performed a [20] R. Mahler, Statistical Multisource-Multitarget Information Fusion, ser.
large-scale evaluation of the algorithm via our extended tar- Artech House information warfare library. Artech House, 2007.
get tracking benchmark based on the Nuscenes data set. The [21] B. Ristic, B. Vo, B. Vo, and A. Farina, “A tutorial on bernoulli
filters: Theory, implementation and applications,” IEEE Transactions
results show a significant increase in estimation performance on Signal Processing, vol. 61, no. 13, pp. 3406–3430, 2013.
compared to our previous publications. [22] A. J. Swain, “Group and Extended Target Tracking with the Probability
Hypothesis Density Filter,” Ph.D. dissertation, Heriot-Watt University,
VII. ACKNOWLEDGEMENTS 2013.
[23] P. Bernhard, “Chain differentials with an application to the mathemat-
We thank Marcus Baum for the fruitful discussions and ical fear operator,” Nonlinear Analysis: Theory, Methods & Applica-
continuous support. tions, vol. 62, no. 7, pp. 1225–1233, 2005.
[24] B.-N. Vo, B.-T. Vo, and H. G. Hoang, “An Efficient Implementation
R EFERENCES of the Generalized Labeled Multi-Bernoulli Filter,” IEEE Transactions
on Signal Processing, vol. 65, no. 8, pp. 1975–1987, apr 2017.
[1] H. Caesar, V. Bankiti, A. H. Lang, S. Vora, V. E. Liong, Q. Xu, [Online]. Available: http://arxiv.org/abs/1606.08350http://ieeexplore.
A. Krishnan, Y. Pan, G. Baldan, and O. Beijbom, “nuscenes: A ieee.org/document/7790845/
Multimodal Dataset for Autonomous Driving,” in Proceedings of the [25] K. Granström, S. Reuter, M. Fatemi, and L. Svensson, “Pedestrian
IEEE/CVF Conference on Computer Vision and Pattern Recognition, tracking using velodyne data — stochastic optimization for extended
2020, pp. 11 621–11 631. object tracking,” in 2017 IEEE Intelligent Vehicles Symposium (IV),
[2] M. Feldmann and D. Franken, “Advances on Tracking of Extended 2017, pp. 39–46.
Objects and Group Targets using Random Matrices,” in 2009 12th [26] H. Kaulbersch, J. Honer, and M. Baum, “A Cartesian B-Spline Vehicle
International Conference on Information Fusion. IEEE, 2009, pp. Model for Extended Object Tracking,” in IEEE 21st International
1029–1036. Conference on Information Fusion (FUSION), 2018, pp. 1–5.
[3] K. Granström and U. Orguner, “On the Reduction of Gaussian [27] X. Rong Li and V. P. Jilkov, “Survey of Maneuvering Target Tracking.
Inverse Wishart Mixtures,” in 2012 15th International Conference on part i. Dynamic Models,” IEEE Transactions on Aerospace and
Information Fusion. IEEE, 2012, pp. 2162–2169. Electronic Systems, vol. 39, no. 4, pp. 1333–1364, 2003.
[4] K. Granström, P. Willett, and Y. Bar-Shalom, “An Extended Target [28] L. Buitinck, G. Louppe, M. Blondel, F. Pedregosa, A. Mueller,
Tracking Model with Multiple Random Matrices and Unified Kine- O. Grisel, V. Niculae, P. Prettenhofer, A. Gramfort, J. Grobler, R. Lay-
matics,” in IEEE 18th International Conference on Information Fusion ton, J. VanderPlas, A. Joly, B. Holt, and G. Varoquaux, “API design for
(Fusion), 2015, pp. 1007–1014. machine learning software: experiences from the scikit-learn project,”
[5] M. Baum and U. D. Hanebeck, “Random Hypersurface Models for in ECML PKDD Workshop: Languages for Data Mining and Machine
Extended Object Tracking,” in IEEE International Symposium on Learning, 2013, pp. 108–122.
Signal Processing and Information Technology (ISSPIT), 2009, pp. [29] NuTonomy, “Nuscenes data set,” 2019. [Online]. Available: https:
178–183. //www.nuscenes.org/
[6] ——, “Shape Tracking of Extended Objects and Group Targets with [30] G. Van Rossum and F. L. Drake, Python 3 Reference Manual. Scotts
Star-Convex RHMs,” in IEEE Proceedings of the 14th International Valley, CA: CreateSpace, 2009.
Conference onInformation Fusion (FUSION), 2011, pp. 1–8. [31] S. Yang, M. Baum, and K. Granström, “Metrics for Performance
[7] F. Faion, A. Zea, and U. D. Hanebeck, “Reducing Bias in Bayesian Evaluation of Elliptic Extended Object Tracking Methods,” in IEEE
Shape Estimation,” in 17th International Conference on Information International Conference on Multisensor Fusion and Integration for
Fusion (FUSION), 2014, pp. 1–8. Intelligent Systems (MFI). IEEE, 2016, pp. 523–528.
[8] F. Faion, A. Zea, M. Baum, and U. D. Hanebeck, “Partial Likelihood
for Unbiased Extended Object tracking,” in 2015 18th International
Conference on Information Fusion (Fusion), 2015, pp. 1022–1029.
[9] R. L. Streit and T. E. Luginbuhl, “Probabilistic Multi-Hypothesis
Tracking,” Naval Underwater Systems Center Newport RI, Tech. Rep.,
1995.
[10] R. Mahler, Advances in Statistical Multisource-Multitarget Information
Fusion, ser. Electronic Warfare. Artech House, 2014. [Online].
Available: https://books.google.de/books?id=zPBaBAAAQBAJ
[11] B.-n. Vo, M. Mallick, Y. Bar-shalom, S. Coraluppi, R. Osborne III,
R. Mahler, and B.-t. Vo, “Multitarget Tracking,” Wiley Encyclopedia
of Electrical and Electronics Engineering, pp. 1–15, 2015.
[12] R. Streit, C. Degen, and W. Koch, “The Pointillist Family of Multi-
target Tracking Filters,” arXiv: Applications, 05 2015.
[13] H. Kaulbersch, J. Honer, and M. Baum, “EM-based Extended Target
Tracking with Automotive Radar using Learned Spatial Distribution
Models,” in Information Fusion (FUSION), 2019 22th International
Conference on, 2019, pp. 1–8.
[14] S. Bordonaro, P. Willett, Y. Shalom, T. Luginbuhl, and M. Baum,
“Extended Object Tracking with Exploitation of Range Rate Mea-
surements,” Journal of Advances in Information Fusion, 2017.
[15] M. Wieneke and W. Koch, “A PMHT Approach for Extended Objects
and Object Groups,” IEEE Transactions on Aerospace and Electronic
Systems, vol. 48, no. 3, pp. 2349–2370, Jul. 2012.
[16] H. Kaulbersch, M. Baum, and P. Willett, “EM Approach for Tracking
Star-Convex Extended Objects,” in IEEE 20th International Confer-
ence on Information Fusion (Fusion), 2017, pp. 1–7.
[17] A. Scheel and K. Dietmayer, “Tracking Multiple Vehicles using a Vari-
ational Radar Model,” IEEE Transactions on Intelligent Transportation
Systems, 2018.

322

Authorized licensed use limited to: Shanghai Library. Downloaded on February 20,2021 at 04:51:26 UTC from IEEE Xplore. Restrictions apply.