2017 IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI 2017)

November 16 - 18, 2017, Daegu, Korea

Extended Object Tracking using IMM Approach for a Real-World Vehicle

Sensor Fusion System
Ting Yuan, Krishanth Krishnan, Bharanidhar Duraisamy
Michael Maile and Tilo Schwarz
Mercedes-Benz Research and Development, North America, USA
Group Research and Advanced Engineering, Daimler AG, Germany
fi[email protected]

Abstract— Autonomous driving poses unique challenges for the consideration of a trade-off between system performance
vehicle environment perception due to the complicated driving and implementation complexity.
environment where the autonomous vehicle connects itself with In vehicle applications, it is necessary for vehicle trackers
surrounding objects. Precise tracking of the relevant dynamic
traffic participants (e.g., vehicle/byciclist/pedestrian) becomes a to treat the object as extended (rigid) target. The extended
key component for the task of comprehensive environmental object tracking is very different from traditional point target
perception and reliable scene understanding. It is necessary tracking approaches (say, in aerospace applications) in many
for vehicle trackers to treat the objects as extended (rigid) important aspects [9][3]. A shape model is first required to
target, as opposed to traditional point target tracking (say, present spatial extent of the object. A good shape model
in aerospace applications). The extended object tracking is
an extremely challenging problem in real world, due to high can largely alleviate many complicated issues, such as mea-
requirements of the object estimation on accuracy of kine- surement origin uncertainty, target (representative) behavior
matic/shape information, association robustness, model match description and measurement-to-track association. In this
on various target motion behaviors, and statistical property study, we focus on on-road vehicle object and present a
amicability (e.g., estimation consistency/covariance reliability). rectangular model with a specified tracking reference point
We present an extended object tracker — based on an in-
teracting multiple model with unbiased mixing estimator for (i.e., the most observable corner over time).
kinematic information at a specified tracking reference point, a In complicated vehicle driving scenarios, a set of repre-
truncated Gaussian scheme for shape (width/length/orientation) sentative dynamic models that follow kinematic physics of
estimation, and a hierarchical association method according the traffic objects is required to comprehensively describe
to both kinematic and shape information — to tackle all of various behaviors of the sounding objects. The powerful
the major challenges. Our special effort is put on handling
an intriguing conflict between theory and practice: the so- interacting multiple model (IMM) approach is a natural
called likelihood credibility issue. That is, the likelihood is choice for such situations [2]. In this study, we choose a
expected to credibly reflect the data statistical probability but is model set consisting of a white noise stop model, a nearly
actually distorted/drifting in real world systems, due to mainly constant velocity model, a nearly constant acceleration model
artificial physics introduced in multiple-stage data processing. and a coordinate turn model. The adaptive switching in these
In this study, from systematic point of view, we design an
interacting multiple model based extended object tracker with selected models (with a proper process noise design) can
proper likelihood compensation in the statistically-distorted real reasonably handle the common, say, vehicle cut-in and stop-
world. It can be shown that the presented tracker can deliver go scenarios.
an effective estimation performance in real road traffic of the Our ultimate goal is to design a practically-effective track-
imperfect world. ing system. In order to achieve this, we must be able to
handle unavoidable conflicts between theory and practice.
I. I NTRODUCTION
One of the most concerning issues is the so-called likelihood
Autonomous driving poses unique challenges for vehicle credibility issue; that is, the data likelihood cannot credibly
environment perception due to the complex driving envi- reflect the corresponding statistical probability, due to mainly
ronment where the autonomous vehicle finds itself in and artificial physics introduced in earlier stage data processing1 .
interacts with surrounding objects. Precise tracking of the rel- This could lead to a likelihood drift — an input measurement
evant traffic participants (e.g., vehicles/byciclists/pedestrians) of the seemingly maximum likelihood is not necessarily
becomes a key component for the task of comprehensive en- the most probable one. Failure to compensate for the drift
vironmental perception and scene understanding. The desire would cause big problems in correct data association and
for preciseness requires a good shape model for the object mode switching of an IMM estimator. A proper strategy is
spatial extent, a set of representative kinematic models that required to deal with the intriguing data statistical likelihood
can comprehensively describe the object dynamic behaviors,
1 Some other aspects, say, data drop-outs and data asynchronization, could
and a proper practical strategy to handle data statistical
add up to the issue but will not be discussed in the study; the corresponding
“distortion” (between theory and practice) in real world. We impact is expected to be reasonably handled to an acceptable level in earlier
aim to deal with all the issues in our tracker design with stage.

978-1-5090-6064-1/17/$31.00 ©2017 IEEE 638

“distortion” problem2 . In the present study, we assume that the dynamic state and
In this study, we design the tracker from a systematic point shape state can be estimated in a decoupled manner
of view for a vehicle sensor fusion system of multiple-stage
p (x(k), s(k)|z1:k ) = p(x(k)|s(k), z1:k )p(s(k)|z1:k )
estimation procedure [14]. The tracking module using the
straightforward IMM approach with the naive rectangular ≈ p(x(k)|z1:k )p(s(k)|z1:k )
vehicle shape model is shown to be a safe choice at current ≈ p(x(k)|zd1:k )p(s(k)|zs1:k ) (2)
stage: we only have limited knowledge on the statistical
property of data from real traffic scenarios, due to the lack where zd1:k and zs1:k are the measurement histories for
of reliable ground truth3 ; we are still in the process of dynamic state and shape state, respectively.
understanding the multiple stage estimation, which has some Fig. 1 illustrates the shape model for a vehicle object,
known issues [16]. More advanced statistical approaches which is reasonably modelled with a rectangular shape, con-
(e.g., random hyper-surface model [3]) are surely worth try- sisting of width, length, and orientation (direction) compo-
Δ
ing but if not designed properly, might lead to a deteriorated nents (i.e., s(k) = [ w l o ] ). The representative kinematic
system. The present study — with a performance evaluation behaviors of the object are described by the state of a
based on some recordings with well-surveyed ground truth reference point, which is chosen to be the nearest observable
— is actually one step towards the understanding of data corner to the ego vehicle4 .
statistical property. As we will show later, the proposed
IMM-based extended object tracker is an effective design
well balanced between sensor property and implementation
complexity.
Admittedly, the rectangular vehicle shape model seems
decade old. A rigorous classic approach, no matter how
old-fashioned it might be, can always provide meaningful
insights about many universal issues and serves very well as
a starting point for system understanding. it is worth noting Fig. 1: The shape model
that our tracker design must be compatible with what we
have and aligned with the bigger fusion system: the input The shape model is a naive but well-thought choice
data are the edges (i.e., contour) of clusters generated from in consideration of the data property from current sensor
earlier stage naive dynamic gridmap implementation [12] system5 . Fig. 2 shows the Lidar measurement point clouds
and, actually due to clean corner property of the dense Lidar of clean corner observations of surrounding vehicles. Mul-
point cloud as shown later, the choice is rather reasonable; tiple benefits can be obtained from the amicable sensor
our approach here is more an engineering safe choice for property for the shape model: the nearest-corner reference
prototype running purpose. Our special effort is put on deal- point is essentially the most observable/visible corner; the
ing with the likelihood distortion issue, which always exists measurement origin uncertainty issue (that commonly needs
in the real world mainly due to many heuristics/unavoidable to be addressed in a Radar system) is largely alleviated
data bias in early stage data processing. with the clean corner illumination; the measurement-to-track
The paper is organized in the following manner. In association can largely ignore the possible miss detections
Section II, we present the shape model for the extended due to the dense reflections. Note that the simple refer-
object tracker. In Section III, we describe the IMM-based ence point strategy, due to the decoupled-state design, can
tracker design for the dynamic and shape state estimation be readily extended by many anchor-point approaches: the
in consideration of statistical likelihood “distortion”. The instant center of rotation (ICR) method as that in a Radar
scenario study is carried out in Section IV. Conclusions and only system [8]; the mean shift kernel-based approach as
summaries are presented in Section V. that in computer vision applications [4]; even some multiple
II. E XTENDED O BJECT S HAPE M ODEL reference point approaches [3].
Admittedly, the shape model has its own known prob-
An extended object tracking system requires a proper
lems: i) the reference point could be jumpy over time as
shape model. The posterior probability density of an ex-
object-sensor geometry changes, leading to degraded velocity
tended object state (at time k) can be shown as
estimation (sudden jumps could be observed as velocity
p (x(k), s(k)|z1:k ) (1) variations); ii) the reference point cannot fully represent the
where z1:k is the accumulative measurements and the object object behaviors (e.g., a smoothly turning object might be
state consists of two parts: the dynamic (kinematic) state observed as static if the reference point is the rotation pivot
x(k) and the shape (spatial) state s(k). point). However, these issues can be mitigated by track post-
processing, depending on specific use cases.
2 The terms “likelihood credibility”/“likelihood drift”/“likelihood distor-
tion” will be used interchangeably in our discussion; all indicate the same 4 For a stopped object, the reference is chosen to be the rectangular center;
issue. this is an internal agreement for consistent static object representation.
3 The ground truth generation is a very time-consuming job with high 5 A 360◦ Lidar coverage with sufficient redundance is setup in the sensor
requirements on hardware support. configuration.

639
 i.e., the most observable corner of an extended target.
B. Shape State Estimation
There is a shape-visibility issue, due to limited sensor
detection/illuminating capability. The cluster pattern of point
clouds belonging to a same object could change dramatically
over time (especially for a partially occluded object). We
propose the following truncated Gaussian procedure for
shape estimation, to overcome the pattern sudden change
issue and yield a “stability”,
Without loss of generality, assume at time tk−1 we have
a shape estimate
s(k − 1) ∼ p(s(k − 1)|zs1:k−1 ) (3)
= N (ŝ(k − 1), Ps (k − 1)) (4)
Fig. 2: Lidar point clouds of clean corner observations of
surrounding vehicle objects With a selected shape propagation model, say, a Wiener
process model with process noise vs (k),
s(k) = s(k − 1) + vs (k) (5)
III. DYNAMIC AND S HAPE S TATE E STIMATION WITH
S TATISTICAL L IKELIHOOD “D ISTORTION ” we could have a predicted shape state estimate as
The dynamic state x(k) and shape state s(k) estimation s(k|k − 1) ∼ p(s(k)|zs1:k−1 )
procedures are carried out in a decoupled manner; an IMM = N (ŝ(k|k − 1), Ps (k|k − 1))
with unbiased mixing is used for the reference point dynamic Δ  
state estimation and a truncated Gaussian model is employed = N s̄(k), P̄s (k) (6)
for the shape estimation. A hierarchical association using It is desirable that the shape estimate could be reasonably
both kinematic and shape information is carried out. Our insensitive to pattern variations of observation clusters under
special effort is put on handling the likelihood credibility the rigid shape assumption. We modify the predicted shape
issue; the statistical “distortion” between the theory and estimate in (6) as a truncated Gaussian with support [as , bs ]
practice is compensated by a linear strategy with empirical
compensation factors learned from preliminary studies. s̄(k) ∼ pt (s(k)|z1:k−1 )
Δ  
A. Dynamic State Estimation = N s̄(k), P̄s (k), as , bs

p(s(k)|z1:k−1 )
There are two reasons making the IMM estimator a natural for as ≤ s(k) ≤ bs
= C(as ,bs ) (7)
choice for the reference point dynamic state estimation: 0 for others
i)the reference point dynamic state is the representative
kinematic information of the extend object, whose motion where C (as , bs ) is a normalization constant.
behaviors subject to change over time from one to another; Then, we can have the following update
ii) the reference point of the extended object is only a s(k) ∼ p(s(k)|z1:k ) (8)
partial representation of the extended object’s motions — for
= p(s(k)|z1:k−1 )p(z(k)|s(k)) (9)
example, when a vehicle makes a smooth turn, the nearest
corner might seem static. This could introduce significant ≈ p (s(k)|z1:k−1 )p(z(k)|s(k))
t
(10)
difficulty in data association and the IMM could alleviate Δ
= p (s(k)|z1:k )
t
(11)
the issues to avoid tracking divergence.
The IMM estimator for the reference point state estimation for any given new shape measurement, which is a bounding
consists of three major steps [2]: 1) mixing/interacting of box generated from Lidar clusters [12] based on the popular
the mode-conditioned estimates; 2) mode-conditioned base- Density-based spatial clustering of applications with noise
state estimation (prediction and update) with selected modes; (DBSCAN) [7].
3) mode probability prediction, update and retrodiction. For It can be seen that starting from a truncated Gaussian
typical vehicle driving scenarios, an IMM estimator should prior, we can also arrive a truncated Gaussian posterior6 .
have at least a stop mode (to handle stop-go situation) The shape estimation stability (i.e., overcoming possible
and a constant velocity mode — if needed, nearly constant sudden changes in measurement pattern) can be achieved
acceleration mode and nearly coordinate turn mode can be by carefully selecting the truncated support, which shall
added. The challenge lies in that the “distorted” likelihood consider noise level and point spread of the observation
information could lead to inaccurate mode interacting. clusters.
The measurement for the dynamic state estimation is the 6 The conjugate relationship can be straightforwardly reached by consid-
position of the nearest corner of each moving remote vehicle, ering truncated Gaussian as a scaled Gaussian.

640
C. Hierarchical Association
50

Data association is one of the most complicated issues σvx [m/s ]

in traditional point target tracking and becomes even more

40

so in the extended object tracking: the association not only

should consider the uncertainty in kinematic space but also
the uncertainty in shape space. In the study, we carry out a 30

X Position (m)
hierarchical data association scheme based on first the object
dynamic information and then the shape information. 20

First, the kinematic information is used to calculate the

likelihood for each measurement that falls in a gate of 10

Gaussian volume (the time index is dropped for conciseness)

   
p zj |h(x̂i ) = N zj ; h(x̂i ), S ij (12) 0
−10 −5
Y Position (m)

0 5 10 15

The measurement (zj ) to track (x̂i ) association highly relies Fig. 3: Highly non-stationary data (in velocity component)
on the covariance information S ij , which could be contami-
nated. An integrated (joint) probability data association [11]
or global nearest neighbour approach can be carried out the vehicle is accelerating. The following linear compensa-
based on the likelihood information. tion strategy is used as a temporal empirical solution in our
Additionally, the object shape information can be used for scenario study
data association in a deterministic way. A simple geometric
intersection of the prediction shape and observed shape is λ̄ = aλ + b (13)
used. The extended sharp intersection is evaluated by two where the modified likelihood λ̄ is compensated from a scale
criteria: i) absolute area overlapping (say, 2 m2 ) and ii) factor a and a possible shifting factor b. The factors can be
area overlapping ratio (say, 40% of already observed area learned from the scenario data [1].
— calculated from estimated length and width). As we
will show later in the scenario study, the shape intersection IV. S CENARIO S TUDY
association is an effective add-on in some situations where
In the scenario study, a multiple stage estimation architec-
the kinematic statistical association fails.
ture is used (as shown in Fig. 4). The multiple-stage Bayesian
A more general association approach based on sequential
estimation procedure could easily cause likelihood drifting
multiple hypothesis test (particularly effective for two vehi-
issue.
cles driving in parallel, where association ambiguity could
happen) is introduced in [15]. DOGMA Inference Object Tracking
(Hidden Markov Clustering (Hidden Markov
D. Likelihood Drifting Raw Model) “Super- Model)
Measurements Measurements”
In many practical situations, multiple-stage data process-
ing/estimation procedure are generally employed to achieve Fig. 4: The environment perception architecture
desired performance with a tradeoff between system per-
formance and implementation complexity. This introduces
issues like artificially inflating covariance matrices [1], auto-
correlated measurement sequence [14][17], and in particular
for the current in-vehicle grid-based sensor fusion system
a quantization effect in the measurement grid generation/a
heuristic mass evidence being put on occupancy estima-
tion [13]. All the issues could cause the likelihood distor-
tion/drifting issue.
Here is an example that always happens in practical ap-
plications. Fig. 3 shows our non-stationary input data, which
is known to be unreliable for tracking. For such data, a fixed
covariance is generally used, due to a black-box agreement
among different function modules. Apparently, this is wrong,
leading to a likelihood that cannot credibly reflect the correct
data statistical property.
The likelihood drifting issue could lead to a disaster result
in data association and IMM mode switching. Wrong data Fig. 5: Tracker input data from an occupancy grid system
could be associated to an irrelevant object, especially for (rainbow ring indicates estimated target motion direction),
nearest neighbour based association; IMM could select a from which a measurement box is generated based on a
nearly constant velocity model as a best match but, in truth, clustering approach.

641
 We focus on tracking on-road dynamic objects. We carry
out our performance evaluation with best available informa-
tion we could have in different recordings from different
aspects: position estimation accuracy, shape estimation ac-
curacy and data association correct rate.
Fig. 5 illustrates the tracker input from earlier stage dy-
namic grid system. There are two known issues for current
implementation [12][16] that can affect our tracker perfor-
mance: i) it provides occupancy estimates with spatial extent
larger than an object should be (i.e., the non-white area is
persistently larger than truth); ii) it cannot handle high-speed Fig. 7: Shape estimation scenario (visible/partially visible
object (due to a failure on the dynamic estimation in the objects)
occupancy grid system) — e.g., as shown in the figure, the
Shape estimation
high-speed target (labelled 1) has inconsistent colors for its 5.5

front, middle and end. We will not fix the problems in current

5
study but instead highlight the issues for future refinement;
Length of visible vehicle
this is another important task for the performance analysis.

Geometric measurement (m)

4.5
Width of visible vehicle
Length of partially visible vehicle
4 Width of partially visible vehicle
A. Estimation accuracy True width of visible vehicle
3.5 True width of partially visible vehicle
We evaluate the kinematic state estimation accuracy based
3
on the so-called SimCity data, which is collected at the
SimCity testing ground located in the main Daimler Re- 2.5

search/Production/Testing Plant in Sindelfingen, Germany. 2

The plant is illustrated in Fig. 6. It is a well-surveyed testing

1.5
ground and claimed that “precision and reproducibility en- 765 770 775 780
Time (s)
785 790 795 800

sured by robotic vehicle that follow predefined trajectories”.

Fig. 8: Shape estimates over time (in width and length)

B. Extended object shape estimation accuracy

Fig. 7 illustrates a scenario from a 40-second real traffic
recording on California roads. The scenario has two same-
side remote vehicles: one is following and the other is on the
left lane. The remote vehicle that follows is categorized as
Fig. 6: SimCity testing ground (picture). partially visible as the length information cannot be extracted
from the measurement clusters. The other remote vehicle
has overtaken the ego vehicle on the left side during the
Table I shows the corresponding root mean square errors
40 seconds. Hence, the length and width information are
(RMSEs) over time in position/velocity space for selected
available during the course of the recording and categorized
SimCity data sets with object vehicle driving in different
as ’visible’ vehicle.
velocities at different motion behaviors The big velocity
Fig. 8 shows the width and length of visible vehicle and
estimation errors (SF02E0R50 and SF03E0R50) confirms
width of the partially visible vehicle compared over time.
the known estimation failure in earlier stage for high speed
It can be seen that the shape estimates change smoothly
objects (as shown in Fig. 5).
during the period as we expected from the truncated Gaussian
model. It is worth noting that an estimation bias around 20-
TABLE I: RMSE in positon/velocity spaces for SimCity
30 cm persistently exists, even though the shape estimation
data
converges. This estimation error confirms earlier failure
Sequence RMSE (pos. in m) RMSE (vel. in m/s) made by the grid occupancy estimation7 . The persistent bias
SF02E0R10 0.3879 2.044
should be considered when compensating for the shape-based
association as a likelihood shifting factor, as in (13).
SF02E0R30 0.0952 2.8857
SF02E0R50 0.192 10.2682 C. Data association correct rate
SF03E0R10 0.1705 1.1423 A real traffic scenario on California roads was selected
SF03E0R30 0.0269 1.7895 to evaluate the data association correct rate. The recording
SF03E0R50 0.0734 10.4482
7 It is considered non-trivial problem; the details about the mistake will
not be discussed here.

642
 ground truth can be obtained) and real traffic on California
roads. It is shown the presented extended object tracker is an
effective real-world solution for the current in-vehicle sensor
fusion system.
R EFERENCES
[1] P. Abbeel, et al., “Discriminative Training of Kalman Filters”, In
Robotics: Science and systems, vol. 2(1), 2005.
[2] Y. Bar-Shalom, X. R. Li and T. Kirubarajan, Estimation with Ap-
plications to Tracking and Navigation: Algorithms and Software for
Information Extraction, Wiley, 2001.
(a) Frame at t = 894 s [3] M. Baum and U. D. Hanebeck, “Extended Object Tracking with
Random Hypersurface Models”, IEEE Trans. on Aerosp. Electronic
Systems, vol. 50(1): 149-159, 2014.
[4] D. Comaniciu, et al., “Kernel-Based Object Tracking”, IEEE Trans. on
Pattern Analysis and Machine Intelligence, vol. 25(5): 564-577, 2003.
[5] C. Coue, C. Pradalier , et al., “Bayesian Occupancy Filtering for Mul-
titarget Tracking: An Automotive Application”, International Journal
of Robotics Research, vol. 25(1): 19-30, 2006.
[6] R. Danescu, et al., “Modeling and Tracking the Driving Environment
with a Particle-Based Occupancy Grid”, IEEE Trans. on Intelligent
Transportation Systems, vol. 12(4), December 2011.
[7] M. Ester, H. P. Kriegel, J. Sander, and X. Xu, “A Density-based
Algorithm for Discovering Clusters in Large Spatial Databases with
Noise”, Kdd, vol. 96(34): 226-231, 1996.
[8] D. Kellner, et al.“Tracking of Extended Objects with High-Resolution
(b) Frame at t = 967 s Doppler Radar”, IEEE Trans. on Intelligent Transportation Systems,
vol. 17(5): 1341-1353, 2016.
Fig. 9: Object tracking on road [9] W. Koch and R. Saul, “A Bayesian Approach to Extended Object
Tracking and Tracking of Loosely Structured Target Groups”, 8th
International Conference on Information Fusion (FUSION 2005), 2005.
[10] C. Laugier, Igor E. Paromtchik, et al., “Probabilistic Analysis of
consists of a drive along a straight road, stopping at a traffic Dynamic Scenes and Collision Risk Assessment to Improve Driving
signal with multiple surrounding vehicles and a left turn at an Safety, IEEE Intelligent Transportation Systems Magazine, pp: 4-19,
October 2011.
intersection. The number of remote vehicles in the scenario [11] D. Musicki, E. Robin and S. Stankovic, “Integrated Probabilistic Data
was changing from 1 to 8. Fig. 9 shows some typical scenes. Association”, IEEE Trans. on Automatic Control, vol. 39(6): 1237-1241,
The data association correct rate is evaluated in terms of 1994.
[12] J. Peukert, “Object Generation from a Dynamic Occupancy Gridmap
track loss. It is observed that the tracks were lost about 10 Considering Autocorrelation”, Master Thesis, Institute of Systems Op-
(out of 1800, for the section we selected) instances during the timization, KIT, Karlsruhe, Germany, Feb. 2016
test. Compared to the number of correct data associations, the [13] S. Thrun, et al.,, Probabilistic Robotics, MIT Press Campridge, 2005.
[14] T. Yuan, B. Duraisamy, J, Peukert, et al., “De-autocorrelation Tracking
track loss is insignificant. It is worth noting that, if the data Scheme for Multiple-Stage Bayesian Estimation in Vehicle Sensor
association is only carried out on kinematic information only, Fusion Systems”, in revision, IEEE Trans. on Industrial Informatics
track loss would increase when the object vehicle makes Special Section, Oct. 2017.
[15] T. Yuan, K. Krishnan, Q. Chen, et al., “Object Matching for Inter-
turns; the shape intersection helps to largely alleviate such Vehicle Communication Systems — an IMM-based track association
track loss (however, wrong association occasionally happens approach with sequential multiple hypothesis test”, IEEE Trans. on
for nearby objects driving in parallel; this can be improved Intelligent Transportation Systems, vol. 99: 1-12, Sep. 2017.
[16] T. Yuan, D. Nuss, et al., “Fundamental Properties of Dynamic Occu-
by a sequential hypothesis test approach [15]). pancy Grid Systems for Vehicle Environment Perception”, 2015 IEEE
International Workshop on CAMSAP, Cancun, Mexico, Dec. 2015
V. C ONCLUSIONS [17] T. Yuan, J, Peukert, et al., “Object Tracking with De-Autocorrelation
In the paper, we present an extended object tracking mod- Scheme for a Dynamic Occupancy Gridmap System”, 2016 IEEE
International Conf. on MFI, Baden Baden, Germany, Sep. 2016.
ule for a grid-based vehicle sensor fusion system. We design
the tracker from a systematic point of view, in combination
with other modules (e.g., occupancy inference/clustering) to
achieve a desired overall system performance. We resort our
solution to a balance between theory and practice, as opposed
to pursuit of more advanced statistical approaches (this is the
next step). An IMM estimator with a hierarchical association
using both kinematic and shape information is presented.
A special effort is put on handling a likelihood credibility
issue caused by mainly artificial physics in earlier stage
data processing. A linear likelihood compensation strategy
is used to account for the observed statistical “distortion”. A
comprehensive scenario study is carried out on the recordings
from the SimCity (a Daimler facility where well-surveyed

643