Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

Persistent automatic tracking of multiple surface

vessels by fusing radar and lidar



Jungwook Han1 , Jinwhan Kim1 and Nam-sun Son2
1
Department of Mechanical Engineering, KAIST, Daejeon, Korea
2
Korea Research Institute of Ships and Ocean Engineering, Daejeon, Korea

e-mail: jinwhan@kaist.ac.kr

Abstract—This paper addresses the problem of automatic fusion. This method enhances the automatic radar plotting
target tracking of multiple surface vessels for unmanned surface aid (ARPA) capability and the surveillance performance of
vehicles (USVs). For safe USV operation, the detection of vessels a USV system. The proposed tracking approach consists of
in the surrounding environment is an important capability, and
marine radars have been used to detect and estimate their three parts. First, target objects are detected using both a pulse
motion. However, vessel detection at short-range using radars radar and a 3D lidar by applying automatic feature extraction
is challenging due to their inherent shadow zone. Therefore, we algorithms on their measurements. Second, the relative bearing
proposes a vessel tracking approach fusing a pulse radar and and range information of the targets is obtained from the
a 3D lidar. The relative bearing and range information between extracted features. Finally, the target motion is analyzed in
a USV and nearby vessels is obtained using radar and lidar
sensors, and their motion including the position, heading, and a dual filter structure using an extended Kalman filter (EKF).
speed is estimated based on a dual filter structure using an
extended Kalman filter (EKF). This approach enables persistent II. R ELATED S TUDIES
tracking of multiple surface vessels. To verify and demonstrate Estimating the motion of target traffic vessels is an impor-
the feasibility of the proposed method, a field experiment was tant capability for safe navigation in marine environments.
performed in an inland river environment and the results are
presented. A variety of tracking algorithms have been introduced, and
Index Terms— Multiple target tracking, unmanned surface exteroceptive sensors such as cameras, sonars and radars have
vehicle, sensor fusion been employed for target tracking [1–5]. Among them, passive
sensors (e.g., cameras or passive sonars) have been applied in
I. I NTRODUCTION the framework of bearing-only tracking (BOT) approach [2],
In recent years, unmanned surface vehicles (USVs) have [4] for target motion analysis. However, the performance of
attracted much interest for their potential applications to time- the BOT approach is highly dependent on the observability
consuming and dangerous missions such as environmental defined by the relative trajectory of an observing ship with
monitoring, reconnaissance, and surveillance. To perform these respect to a target ship, and thus the BOT performance may
missions, autonomous vehicle capabilities including precise degrade severely in some situation (e.g., head-on targets).
navigation, path planning, control, mapping, and obstacle Compared to the passive sensors, marine radars provide both
detection are required. Among them, this study focuses on bearing and explicit range information. Various types of radar
automatic detection and motion estimation of marine obstacles sensors such as S-band, X-band, and high-frequency (HF)
which are important capabilities for safe vehicle operation. radars have been utilized. Among them, X-band radars are
Marine radars are standard navigational sensors for target commonly used as the standard navigational sensor for surface
ship tracking and collision avoidance. The sensor provides rel- ships [6]. With the radars, a pixel-wise detector and a joint
ative bearing and range information of surrounding obstacles probabilistic data association (JPDA) tracker were imple-
over a wide range. However, their detection capability greatly mented to estimate kinematics and physical size of targets in
degrades when target objects are at close range due to their a tracking filter [7]. To improve estimation of target trajectory,
inherent sensing characteristic (e.g., close-range shadow zone). multiple motion models were considered in an alpha-beta-
For persistent automatic target tracking, additional sensors are tracking filter [8].
required to detect short range obstacles. To obtain more accurate target state estimates and broader
In fact, three-dimensional (3D) lidars have been successfully tracking coverage, multi-sensor tracking methods can be em-
applied to self-driving cars to collect traffic information in the ployed [9]. For example, camera measurements were com-
surrounding environment. Lidars can provide precise relative bined into a radar-based tracking algorithm to complement the
bearing and range information of close-range obstacles within radar’s bearing measurement and minimize the dead-zone [10].
the radar’s shadow zone. Therefore, persistent target tracking A multiple sensor module including radars, a lidar, and a stereo
with no dead-zone can be achieved by fusing radar and lidar camera mounted on a USV platform was applied for target
measurements. tracking of marine traffic ships [11]. For maritime surveillance,
In this paper, we address a multiple-target tracking method multi-radar tracking was implemented by incorporating the
using radar and lidar sensors in the framework of sensor weighted center of detected targets into a tracking filter [12].

978-1-5090-5278-3/17/$31.00 ©2017 IEEE

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on April 24,2022 at 12:50:43 UTC from IEEE Xplore. Restrictions apply.
In this paper, multiple target tracking by fusing radar and background of each mask are extracted. The extracted signals
lidar sensors is proposed for persistently tracking multiple sur- are classified into identical targets, and then the weighted
face vessels with no dead-zone. To detect short range obstacles center of each target is estimated considering the intensity of
within the radar’s dead-zone, a 3D lidar is employed and its the each signal.
measurements are incorporated in an EKF-based tracking filter 2) Lidar-based detection: Lidar measurements are provided
with radar measurements. The proposed method is expected to as point cloud data, and the point-distance-based segmen-
greatly enhance the surveillance capability of USV systems. tation (PDBS) algorithm is applied to extract targets from
the data [14]. A 3D lidar which consists of multiple lasers
III. P ROBLEM FORMULATION arranged in the vertical direction is used in this study. Each
The formulation of the proposed multiple target tracking laser provides a point cloud set reflected from the surfaces of
mainly consists of two parts. The first part is the automatic target vessels. The coordinates of the data are transformed
target detection from radar and lidar measurements. To extract into the stabilized body-fixed frame which is obtained by
the relative bearing and range information of target vessels, compensating the lidar’s roll and pitch angles, and then the
segmentation and classification methods are applied. The sec- point cloud data are projected onto the vehicle’s motion
ond part is the multiple target tracking to estimate the motion plane. The PDBS algorithm is applied on the projected point
of each target including its position, heading, and speed using cloud data to assign a set of classified objects (i.e., clusters).
an EKF. The tracking algorithm is designed to have a dual Then, the clusters driven from noises are removed considering
structure to improve the robustness and computational effi- the density-based statistical characteristics of clusters. To be
ciency of target tracking. The overall procedure is illustrated more specific, the lidar measurements reflected from a glint
in Fig.1. of sunlight may appear sporadically, and the cluster driven
from this has low spatial density compared to that of a real
target (see Fig. 2(b)). Therefore, clusters whose spatial density
are smaller than a predefined threshold value are ignored. The
overall procedure is summarized in Algorithm 1.

Algorithm 1 Pseudocode for lidar-based target detection


Input:
3D point cloud data : P
Roll and pitch angles : φ, θ
Cluster density threshold : td
Output:
The number of targets and their center coordinates
Fig. 1: The flowchart of the proposed multiple target tracking. Procedure:
A dual filter structure (reserved and confirmed tracks) is intro- 1. Coordinate transformation to the stabilized body-fixed
duced to enhance the robustness and computational efficiency frame : s P ←− CoordinateTransform(P, φ, θ)
of target motion estimation. 2. Point cloud data projection to the vehicle’s motion plane
: s P0 ←− PointCloudProjection(s P)
3. Point cloud clustering :
A. Automatic target detection s 0
Pc ←− PDBS algorithm(s P0 )
The relative bearing and range information between an s 0
Pc = {ci | i = 1, · · · , m}
observing ship and each target ship is required for target 4. Density-based cluster filtering :
motion estimation. Segmentation algorithms are applied on for i ← 1 to m do
radar and lidar measurements to assign the data into a set ρi ←− SpatialDensityEvaluation(ci )
of classified targets and extract their relative positions. if ρi < td then
1) Radar-based detection: To extract targets from radar EliminateCluster(ci )
measurements, a set of preprocessing steps is implemented. end if
Sensitivity time control (STC) and fast time constant (FTC) end for
modules are applied to minimize unwanted echoes returned
from rain, snow, and water surface. Radar signals reflected
from a land are separated by designating the region of in- B. Multiple target tracking
terest (ROI) on radar images. To extract potential targets on The EKF-based motion estimation of the detected targets
the preprocessed radar image, a dynamic thresholding method is implemented using a constant velocity model. A dual
using the cell averaging constant false alarm rate (CA-CFAR) filter structure is introduced to increase the robustness and
detector [13] is applied. For this, a window mask is defined computational efficiency of target tracking by reducing false
on the radar image considering the size of targets, and then measurements. The filter structures are composed of a reserved
the signals which are sufficiently stronger than the local track and a confirmed track. A detected target is initially

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on April 24,2022 at 12:50:43 UTC from IEEE Xplore. Restrictions apply.
The two measurement datasets are employed to update the
motions of the observer and targets.
The first measurement dataset is for updating the motion of
the observer whose equation is expressed as
zo = [ zx zy zψ zV ]T = [ xo yo ψo Vo ]T + vo (5)
where zx , zy , zψ , and zV denote the position, heading and
speed measurements from onboard motion sensors. The vo is
the measurement noise, which is assumed to follow a zero-
mean Gaussian distribution.
(a) (b) The second measurement dataset is for updating the motion
of targets whose equation is expressed as
Fig. 2: Automatic target detection. Sensor measurements were    
yT −yo
collected in an inland river environment (in blue dots). (a)   arctan xT −xo − ψo
zβ  
radar measurements (100 m ring). (b) lidar measurements (10 zT = =  + vT (6)
zρ  q
2 2

m ring). (xT − xo ) + (yT − yo )

where zβ and zρ denote the relative bearing and range


registered into the reserved track to find whether it is a real measurements obtained from radar and lidar sensors. The
target or a clutter through the traditional M-of-N rule. The measurement noise vT is applied differently depending on the
targets which satisfy the M-of-N rule are registered into the measurement accuracy of radar and lidar sensors, and the two
confirmed track, and then their motion is estimated in the measurement noises are assumed to be mutually independent.
tracking filter. However, the targets whose position uncertainty For data association between the newly detected target
level exceeds a pre-defined value are removed. measurements and the existing targets in the filter state, the
In the filter structure, a three-degrees-of-freedom (3 DOF) global nearest neighbor method (GNN) is applied [10]. To
kinematic model is used to describe the motions of the eliminate unlikely measurement-to-track pairs, an ellipsoidal
observing vehicle and targets. The state vector representing gate is applied and the association is allowed if the following
the observer is described as condition is satisfied
xo = [ xo yo ψo Vo ]T (1) z0 S −1e
d2 = e z≤G (7)
where xo and yo are the observer’s position defined in the where d2 is the normalized statistical distance, S is the
global frame, ψo is the observer’s heading, and Vo is the speed measurement innovation covariance matrix, and e z is the mea-
in the longitudinal direction. The equations of motion for the surement innovation vector which is defined by the difference
filter’s system kinematics of the observer can be written as: between the actual sensor measurement and the expected
measurement estimated by the filter. The maximum likelihood
ẋo = [ Vo cos ψo Vo sin ψo 0 0 ]T (2) gate G [10] can be written as
 
In order to estimate the motions of the observing vehicle PD
G = 2 ln p (8)
and targets simultaneously, the state vector of the tracking (1 − PD (2π)M/2 β |S|)
filter structure is augmented by cascading the observer state
where PD is the detection probability, M is the measurement
vector xo and the target state vector xT = [ xT T
T1 xT2 · · · ]
T
T dimension and β is the new source density value which is
where xTi = [ xTi yTi ψTi VTi ] describes the state vector of
defined as the false alarm probability divided by the scan
the ith registered target. The equations of motion for the filter’s
area. The GNN method is a single hypothesis data association
system kinematics of the ith target can be written as:
approach, and thus association conflicts may occur for multiple
target tracking. For example, more than one measurement can
ẋTi = [ VTi cos ψTi VTi sin ψTi 0 0 ]T (3)
be allocated in a track gate or a single measurement can be
Note that multiple targets are considered in this study, and thus allocated in the gates of multiple tracks as shown in Fig. 3.
the total dimension of the target state vector increases with To solve the measurement-to-track assignment problem, an
the number of registered targets. The system dynamics can auction algorithm [10] is applied by defining a cost value c as
be described as the augmented vector equation which consists follows.
of the motion models of the observer and the targets. The
c = G − d2 − ln[|S|] (9)
equations of the system dynamics can be written as
In the cost evaluation, the determinant |S| is used to penalize
ẋ = [ ẋT T T
o ẋT ] + w (4)
the tracks with large prediction uncertainty. In an actual imple-
where w is the zero-mean Gaussian process noise which mentation, the rectangular gate condition [10] is additionally
reflects the uncertainty of the motion model. considered for computational efficiency, and the ellipsoidal

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on April 24,2022 at 12:50:43 UTC from IEEE Xplore. Restrictions apply.
gate is only applied to the pairs for which the rectangular
gate condition is met.

(a) (b)

Fig. 3: Example of data association conflict. The measuremet


m1 is satisfied by the ellipsodial gate of three targets.

IV. F IELD EXPERIMENT


A. Experimental setup
Field experiments were carried out in an inland river using
a USV (observing vehicle) developed by Korea Research
Institute of Ships and Ocean Engineering (KRISO). The USV
was equipped with a pulse radar (Furuno FAR-2117), a 3D (c)
lidar (Velodyne HDL-32E), and a suite of motion sensors in-
cluding an inertial measurement unit (IMU) and the Real-Time
Kinematic Global Positioning System (RTK-GPS). In addition,
an automatic identification system (AIS) was installed in the
target vehicles to obtain their ground-truth trajectories. The
proposed algorithm was applied to the USV and implemented
using SPx software library for radar signal processing [15] and
point cloud library (PCL) for lidar data processing [16]. For
multiple target tracking, the USV was commanded to approach (d) (e)
three-static targets (buoys) and one-moving target (ship) within
the radar shadow range of approximately 80 ∼ 100 m in the Fig. 4: Experimental setup. (a) ARAGON USV (7.5 m in
environment. The speeds of the USV and the moving target length). (b) sensor configuration. (c) experimental scenario in
were set to be ∼ 6 knots. The USV platform and experimental an inland river. (d)-(e) snap shots of field experiment.
scenario are described in Fig. 4.
B. Results degrades when the relative range falls within 100 m. Thus,
Figure 5 compares the resulting target trajectories. The the position uncertainties of the targets monotonically increase
magenta dots represent the estimated target trajectory by the as the USV approaches the targets at close range, as shown
proposed tracking filter; the blue dots represent the trajectory in Fig. 6(b). However, the large position uncertainties are
of the USV obtained by the onboard motion sensor. The AIS considerably reduced when lidar measurements are combined
data were collected for performance comparison, and these into the proposed tracking filter, which shows the feasibility
are represented as black circles. The estimated trajectory of the of the proposed method for persistent automatic tracking.
moving target satisfactorily follows the ground-truth trajectory.
Also, the estimated trajectories of the three stationary targets V. C ONCLUSION
are well confined within a moderate range. In addition to the This study proposed a method for automatic target tracking
trajectory comparison, the relative ranges of the targets with re- of multiple surface vessels with a USV. A pulse radar and
spect to the USV are shown in Fig. 6(a) with the measurements a 3D lidar were used to detect obstacles in the surrounding
that were used to update the estimated motion of the registered environment, and segmentation algorithms were applied to
targets in the filter structure. While all the targets could be extract targets for the sensor measurements.
detected from their initial positions using the radar sensor, For target motion analysis, the detection results were in-
the detection performance of the radar-based tracking method corporated into an EKF-based tracking filter. To increase

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on April 24,2022 at 12:50:43 UTC from IEEE Xplore. Restrictions apply.
(a)

(b)

Fig. 6: Result analysis. (a) Estimated relative range of targets


with sensor measurements. (b) Comparison of the position
uncertainty level (1σ).
Fig. 5: Comparison of the estimated target trajectories with
AIS data.
[5] J. Park, J. Kim, and N.-s. Son, “Passive target tracking of marine traffic
ships using onboard monocular camera for unmanned surface vessel,”
computational efficiency, a dual filter structure was introduced Electronics Letters, vol. 51, no. 13, pp. 987–989, 2015.
and implemented, and radar and lidar measurements were [6] S. S. Blackman, “Multiple-target tracking with radar applications,”
Dedham, MA, Artech House, Inc., 1986, 463 p., 1986.
combined in the dual filter structure. This approach enables [7] G. Vivone, P. Braca, and B. Errasti-Alcala, “Extended target tracking
persistent and robust tracking of multiple surface vessels. applied to x-band marine radar data,” in OCEANS 2015-Genova. IEEE,
2015, pp. 1–6.
To validate the proposed method, the proposed tracking [8] W. Kazimierski, “Determining of marine radar target movement models
algorithm was applied to a newly-developed USV system and for the needs of multiple model neural tracking filter,” in Radar
a field experiment was conducted. The results demonstrated Symposium (IRS), 2011 Proceedings International. IEEE, 2011, pp.
611–616.
the feasibility of the proposed approach. [9] M. Liggins II, D. Hall, and J. Llinas, Handbook of multisensor data
fusion: theory and practice. CRC press, 2017.
ACKNOWLEDGMENT [10] S. Blackman and R. Popoli, “Design and analysis of modern tracking
systems(book),” Norwood, MA: Artech House, 1999.
This research was supported by the projects titled ‘Devel- [11] L. Elkins, D. Sellers, and W. R. Monach, “The autonomous maritime
opment of autonomous unmanned surface vehicle for ocean navigation (amn) project: Field tests, autonomous and cooperative be-
haviors, data fusion, sensors, and vehicles,” Journal of Field Robotics,
observation and sea surveillance (PMS3490)’ funded by the vol. 27, no. 6, pp. 790–818, 2010.
Ministry of Oceans and Fisheries, Korea. [12] D. Nikolió, Z. Popovic, M. Borenovió, N. Stojkovió, V. Orlić,
A. Dzvonkovskaya, and B. M. Todorovic, “Multi-radar multi-target
tracking algorithm for maritime surveillance at oth distances,” in Radar
R EFERENCES Symposium (IRS), 2016 17th International. IEEE, 2016, pp. 1–6.
[13] H. Rohling, “Radar CFAR thresholding in clutter and multiple target
[1] J. Kim, S. Vaddi, P. Menon, and E. Ohlmeyer, “Comparison between
situations,” IEEE transactions on aerospace and electronic systems,
three spiraling ballistic missile state estimators,” in AIAA Guidance,
no. 4, pp. 608–621, 1983.
Navigation and Control Conference and Exhibit, 2008, p. 7459.
[14] C. Premebida and U. Nunes, “Segmentation and geometric primitives
[2] E. W. Frew and S. M. Rock, “Trajectory generation for constant velocity
extraction from 2d laser range data for mobile robot applications,”
target motion estimation using monocular vision,” in Robotics and Au-
Robotica, vol. 2005, pp. 17–25, 2005.
tomation, 2003. Proceedings. ICRA’03. IEEE International Conference
[15] Cambridge Pixel Ltd. SPx software library. [Online]. Available:
on, vol. 3. IEEE, 2003, pp. 3479–3484.
www.cambridgepixel.com
[3] Y. Bar-Shalom, X. R. Li, and T. Kirubarajan, Estimation with applica- [16] R. B. Rusu and S. Cousins, “3D is here: Point Cloud Library (PCL),”
tions to tracking and navigation: theory algorithms and software. John in IEEE International Conference on Robotics and Automation (ICRA),
Wiley & Sons, 2004. Shanghai, China, May 9-13 2011.
[4] A. Farina, “Target tracking with bearings–only measurements,” Signal
processing, vol. 78, no. 1, pp. 61–78, 1999.

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on April 24,2022 at 12:50:43 UTC from IEEE Xplore. Restrictions apply.

You might also like