A Novel Ship Trajectory Clustering Analysis and Anomaly

Detection Method Based on AIS Data

Chuang Zhang1, Songtao Liu1, Muzhuang Guo1**, Yuanchang Liu2*
1. Navigation College, Dalian Maritime University, Dalian, China. 2. Department of Mechanical
Engineering, University College London, London, UK.
Abstract
The increasing volume of ship traffic has resulted in new challenges for the supervision of maritime
safety administration. The conventional manual monitoring approach for maritime traffic is
inefficient and lacks specifics, particularly for supervising ships with abnormal trajectories. To
address this issue, this study proposes utilizing the minimum description length criterion to extract
features from ship trajectory data provided by the automatic identification system (AIS). This
approach simplifies the compression of ship trajectories. Additionally, the dynamic time warping
trajectory similarity measurement algorithm is employed to optimize the density-based spatial
clustering of applications with noise algorithm. This optimization enables the clustering of ship
trajectories and the acquisition of normalized ship motion trajectories. Furthermore, a ship trajectory
prediction method based on a transformer model is proposed, and the normalized motion trajectory
is used as the training set for model training. The trained ship trajectory prediction model is
subsequently utilized to predict the target ship trajectory. The AIS ship trajectory data in the vicinity
of Yantai Port were used for experimental verification. The results demonstrate the effectiveness of
the proposed approach in identifying abnormal ship trajectories.

Keywords: Ship trajectory; Minimum description length; Dynamic Time Warping; Transformer

1. Introduction
The rapid development of the automatic identification system (AIS) has enabled the acquisition
of extensive ship motion trajectory data, providing a foundational basis for predicting ship behavior
and monitoring abnormal ship behavior. In the realm of maritime research, ship trajectory clustering
and anomaly detection algorithms have become focal points, particularly with the development of
Internet technology and big data analysis (Yang et al., 2019). The significance of ship trajectory
clustering analysis and anomaly detection lies in advancing shipping management intelligence,
enhancing navigation safety, and enhancing the efficiency of the shipping industry.
Research of anomaly detection was initially conducted using statistical methods (Kowalska
and Peel, 2012), the form of normal trajectory model was a probability model of trajectory point
information, which primarily include Kernel Density Estimation (KDE) (Ristic et al., 2008; Dai et
al., 2020; Wang et al., 2022), Gaussian Mixture Model (GMM) (Laxhammar et al., 2009), Gaussian
Process (GP) (Smith et al., 2012), Hidden Markov Model (HMM) (Shahir et al., 2014), and Bayesian
Network (BN) (Mascaro et al., 2014; Zhen et al. 2017; d’Afflisio et al. 2021). The statistical method
of anomaly detection uses statistical testing to determine whether the behavior of ship matches a
statistical model representing conventional ship behavior. When the matching probability is low, it
is considered as abnormal behavior. The disadvantage of statistical method is that the matching
accuracy is related to historical data and ignores real-time scene. And this method is difficult to
combine multi-source information, expert knowledge, etc., and is not suitable for ship trajectory
anomaly detection in the increasingly complex marine traffic circumstance.
In recent years, much effort has been done to improve the performance of the existing
algorithms to make them applicable for ship trajectory anomaly detection in the increasingly
complex marine traffic circumstance. Clustering, as a tool for big data analysis, is an unsupervised
technique that does not depend on any prior knowledge. A trajectory clustering framework based on
AIS data was designed to analyze routes, which considered the geographic spatial information and
contextual features of ship trajectories, and thereby the density-based clustering algorithm
automatically classified different routes (Sheng and Yin, 2018). A new features of local fast ship
trajectories method was proposed to search for global and local features of ship trajectories (Tang
et al., 2021). Soares et al., (2015) proposed an unsupervised method to segment trajectories without
predetermined clear criteria. The DBSCAN algorithm (Ester et al., 1996) is a pioneer technique in
the context of density-based clustering. Lei (2016) developed a framework called MT-MAD
(maritime trajectory modeling and anomaly detection) to explore frequent movement behaviors and
established a single index for combining anomaly scores to determine the suspicious level of each
ship's trajectory. The DBSCAN approach is used to cluster course over ground (COG) and speed
over ground (SOG) in AIS data, considering both density points and those with similar COGs and
SOGs. To handle large datasets, an improved density-based spatial clustering of applications with
noise (DBSCAN) clustering algorithm was introduced (Nooshin and Hamid, 2022; Li et al., 2020).
Kontopoulos et al. (2021) clustered the track points and extracted the ship’s steering points based
on the DBSCAN algorithm. DBSCAN algorithm was reused to cluster track lines, in which the
Lagrange interpolation algorithm was used to fill the gaps between the steering points. Han et al.
(2021) proposed an optimized DBSCAN algorithm for abnormal ship trajectories detection by
clustering AIS ship trajectories with the same features, and ultimately validated the performance of
this method using data from the Gulf of Mexico. To speed up the computational efficiency of the
clustering algorithm. Yang et al. (2022) proposed Density based Trajectory Clustering of
Applications with Noise (DBTCAN) algorithm. This method uses Hausdorff distance as a similarity
measure to cluster trajectories of different lengths. Therefore, the DBTCAN algorithm can not only
recognize noise trajectories but also adaptively select its optimal input parameters, which can be
widely applied in ocean research.
Rong et al. (2020) employed the DP algorithm to identify turning points based on ship type,
size, final destination, and other maritime traffic patterns. The DBSCAN method clusters turning
points, and then combines the density region of these points by KDE. DBSCAN and the Kernel
Density Estimation-based Outlier Factor processing algorithm was introduced to calculate the
abnormal probability distribution value of trajectory points, eliminating low-probability distribution
edge points (Jin et al., 2023). Bai et al. (2023) designed an adaptive threshold fast DBSCAN
algorithm for vessel trajectory clustering. The fast DTW algorithm is used to reduce the
computational complexity and ensure the accuracy of trajectory similarity, and the DBSCAN
parameters are adaptively determined by combining the similarity distribution of trajectories with
an improved K-adaptive nearest neighbor. However, the method is distance-based trajectory
clustering algorithm and does not consider the speed and acceleration of ship, and thereby attracting
the attention of more and more scholars in the field of artificial intelligence technology.
Detecting anomalies in ship trajectories is a complex and challenging task due to the influence
of the surrounding environment on the navigation statuses of both the own ship and other ships (Liu
et al., 2022). Thus, artificial intelligence technology is increasingly applied in ship trajectory
anomaly detection research. Traditional supervised learning techniques were employed to predict
arrival time and reorganize data based on spatial grids. Deep learning architectures based on long
short-term memory (LSTM) are also explored to address the next position prediction problem
(Gözde et al., 2021). Park et al. (2021) employed the spectral clustering method to cluster ship
trajectories and established a ship trajectory prediction model using bi-directional LSTM (Bi-
LSTM). The support vector machine is employed to predict the ship's next course at the exit of the
traffic route in Tokyo Bay based on dynamic historical AIS data. Nevertheless, greater emphasis
should be placed on enhancing prediction accuracy by increasing efforts (Nishizaki et al., 2018).
Venskus et al. (2017) presented a self-learning adaptive classification method based on the self-
organizing map and virtual pheromone for ship trajectory anomaly detection. By utilizing the
gradient descent algorithm, the verification dataset is used to calculate the pheromone intensity
threshold for trajectory anomaly detection. Venskus et al. (2019) extended their previous work and
studied a data batch processing strategy for neural network retraining to detect anomalies in
streaming maritime traffic data. The results demonstrate that it is possible to reduce the retraining
time while keeping the accuracy relatively unchanged. Mantecón et al. (2019) proposed a supervised
deep learning framework for ship anomaly detection, utilizing a convolutional neural network to
infer navigation states from the ship trajectory based on AIS information. Zhao and Shi (2019)
combined DBSCAN and recurrent neural networks (RNN) to obtain the clustering ship trajectories
and prediction of large-scale ship trajectories. Wen et al. (2020) clustered the ship trajectories based
on DBSCAN algorithm to identify key regions, and then applied artificial neural networks to learn
the relationships between key regions to generate reasonable routes for different ships. However,
these algorithms face challenges related to parameter setting, noise recognition, and sensitivity to
density distribution in datasets. With the development of deep learning technologies, Li et al. (2023)
divided multiple different subsequences by quantifying the similarity of time distribution and
matching the time distribution, and then constructed the adaptive transformer model based on
transfer learning to conduct the accurate prediction of the future trajectory. However, limited by the
long training and learning time of the mentioned models above, the time complexity of parameter
calculation is still relatively high and cannot be used for trajectory prediction of a large amount of
ship data.
Given this background, analyzing and studying ship trajectory clustering analysis and anomaly
detection based on machine learning becomes crucial. The main contributions of this paper are as
follows:
(1) The minimum description length (MDL) criterion is designed based on the model
description length and the data description length using original AIS ship trajectory data, which
ensures to extract ship trajectory features while well maintaining the trajectory features and shape.
(2) The DBSCAN clustering algorithm using the dynamic time warping (DTW) trajectory
similarity measurement method is proposed according to the position, speed and course of ship,
which essentially solves the problem of slow accuracy and trajectory similarity between ship
trajectories and clustering effectiveness.
(3) A ship trajectory prediction model based on the transformer model is constructed for the
anomaly detection of ship by using the transformer's self-attention mechanism to capture the long-
term dependency relationship between ship trajectory information. The results demonstrate that this
method is superior to other algorithms in terms of effectiveness and reliability.
This paper is organized as follows: Section 2 presents the proposed method for ship trajectory
clustering analysis and anomaly detection. Section 3 presents the experiments setup include the
datasets used in the experiments, evaluation results of the proposed method. At the end of this
section, the proposed method is compared with other previous recent studies. Finally, the paper is
concluded and some suggestions are given for future studies.
2. Methodology
To address the challenges related to long-term jumps and missing ship trajectories, a method is
proposed to divide sub-trajectories and ensure the availability of ship trajectory data. The MDL
criterion is utilized to extract ship trajectory feature points and derive ship trajectory features. In
water traffic scenarios, the lengths of AIS ship trajectories cannot be made consistent for all
trajectories. Therefore, in view of the lack of consideration for water traffic scenario factors when
measuring ship trajectory similarity, a ship trajectory feature extension method is introduced,
incorporating ship position, speed, and course. This extended feature set is used in conjunction with
the DTW trajectory similarity measurement distance and the DBSCAN algorithm to achieve ship
trajectory clustering. This approach enhances the accuracy of similarity measurement between ship
trajectories and enables the generation of normalized ship motion trajectories. To address ship
trajectory anomaly detection, a ship trajectory prediction method based on a transformer model is
proposed. The normalized ship motion trajectory obtained from clustering serves as the training data
for the network. This enables the transformer model to predict trajectories and serve as a tool for
anomaly detection, completing the detection of ship trajectory anomalies. The flowchart depicting
ship trajectory clustering and anomaly detection is illustrated in Fig. 1.
Dynamic data Static data
Time
MMSI
Extraction
Data Position
Source AIS dataset AIS data
SOG
COG NAME

MDL method
Ship
Data Data
trajectory
preprocess clean
feature
ing ing
extraction

K nearest neighbors DTW DBSCAN clustering

Speed

Trajectory Course
clustering

Position

Transformer model
Trajectory Predicting Speed anomaly
prediction the Position
and trajectory anomaly
anomaly of the target detection
Detection ship Course anomaly

Fig. 1. Flowchart of ship trajectory clustering and anomaly detection.

2.1 Feature Extraction of AIS Data
The ship trajectory data consists of a large number of AIS data points. Meanwhile, due to the
large tonnage of the ship, slow sailing speed, and gradual changes in motion status, a large overlap
of ship attribute information exists between trajectory points in a short period of time. Certain
trajectory points even possess identical position information, resulting in more redundancy in AIS
data. In addition, there are a large number of outliers in the original AIS trajectory data. To enhance
data mining efficiency, it is essential to ensure the accuracy, completeness, and conciseness of ship
trajectory information. Therefore, preprocessing operations are conducted to extract ship trajectory
feature points from AIS data before performing clustering analysis. The MDL criterion is chosen as
the method for extracting trajectory feature points. This criterion allows for compressing ship
trajectories and reducing data storage and transmission costs, thus facilitating data visualization and
analysis. The fundamental idea of MDL is to encode and compress a given set of instance data using
a specific model to save storage space. The compressed data, along with the model used, can be
saved for future correct recovery of the instance data. The total data length to be saved corresponds
to the sum total of the length of the encoded and compressed instance data and the length required
to store the model. This combined length is known as the total description length.
The MDL principle consists of two components: (1) the model description length (L(H)), which
represents the length needed to describe the model H and can be interpreted as the complexity of
the model itself. Generally, models with lower complexity feature shorter description lengths. (2)
The data description length (L(D/H)), which denotes the length required to describe the data D under
the given model H. It can be interpreted as the amount of information necessary to encode the data
D using the model H. Typically, if the model H effectively explains the data D, the description length
of data D will be shorter under the given model conditions. As shown in Fig. 2, for a trajectory {P1,
P2, P3, P4, P5} in a string of trajectory data, in the MDL approach, each point in the trajectory is
traversed to calculate the MDL under a compression model (MDLpar) and the MDL under a non-
compression model (MDLnopar) for that point. If MDLpar>MDLnopar, the point is considered a feature
point. Assuming a trajectory sequence T = {P1, P2, ..., Pn} is divided into multiple segments, the set
of feature points is denoted as P={ Pc1, Pc2, Pc3, ..., Pck}. The specific calculation formula for MDL
is as follows:
k −1
L( H ) =  log 2 (len( Pji Pji+1 )) (1)
j −1

n −1 k −1
L( D / H ) =  log 2 (d ( Px Px +1 , Py Py +1 )) + log 2 (d ( Px Px +1 , Py Py +1 )) (2)
x −1 y −1

where Len(P) represents the length of sequence P, and dα and dθ are the vertical distance and angular
distance between sequences, respectively.

Pc1 Pc2

P1 P4

P5
P2
P3

Fig. 2 Instance diagram of MDL.

According to the above formula,
 L( H ) = log 2 (len( PP
1 4 )) (3)

L( D / H ) = log 2( d ⊥( PP 1 2 ) +d ⊥( PP
1 4 ,PP 1 4 ,P2 P3 ) +d ⊥( PP
1 4 ,P3 P4 ) ) +
(4)
log（
2 d( PP 1 2 ) +d( P
1 4 ,PP 1 P4 ,P2 P3 ) +d( PP
1 4 ,P3 P4 ) )

The specific process of MDL principle compression trajectory is as follows:

Data input

Calculate the MDLpar and

Read trajectory data according MDLnopar of each trajectory
to MMSI point

Determine
No MDLpar> MDLnopar

Yes
Calculate the encoding
length L(H) of trajectory T
Construct a new trajectory
description, relative to the
T ={P1,P2,...} by connecting in
compressed encoding
sequence
length L(D/H) of the
original trajectory

MDLpar=L(H)+L(D/H) Determine
MDLnopar=L(T ) MDLpar< MDLnopar

Yes No
Select MDLpar as the Select MDLnopar as
description the description
encoding length for encoding length for
the final trajectory the final trajectory

Output final
trajectory

Fig. 3 Flowchart of MDL compression trajectory.

2.2 Cluster analysis of ship trajectories based on improved DTW distance algorithm
Trajectory data obtained from the AIS is a type of spatiotemporal data that includes information
such as position and time. Trajectory pattern mining and trajectory classification are essential
components of trajectory data research (Wang et al., 2023). Trajectory pattern mining involves
grouping similar trajectories together to analyze group behavior or periodic behavior of moving
objects. On the other hand, trajectory classification entails establishing a model to analyze the
similarity between trajectories and partitioning them into different states. When assessing trajectory
similarity, it is crucial to select an appropriate trajectory similarity measurement algorithm. To
determine the similarity between different trajectories, an improved DTW distance algorithm is
proposed, and the DBSCAN method is employed for clustering.
Suppose we have two trajectory sequences, 𝑇raj1=[p1, p2,..., pn] and 𝑇raj2=[q1, q2,..., qm], where
m and n represent the number of trajectory points in 𝑇raj1 and 𝑇raj2, respectively. The DTW distance
between 𝑇raj1 and 𝑇raj2 can be calculated using Equation (5), where dist(p, q) refers to the
Euclidean distance between trajectory points p and q. The Rest(𝑇raj1) and Rest(𝑇raj2) terms indicate
the remaining trajectory segments of 𝑇raj1 and 𝑇raj2, respectively, after removing the first trajectory
point. As illustrated in Fig. 4, the minimum distance between points on the trajectory sequence is
calculated to determine the matching combination of trajectory points. During this process, certain
trajectory points (such as q2 and p3) may be reused. Fig. 5 depicts the combination of all matching
trajectory points in the form of a matrix. The shortest Euclidean distance matrix between each point
of the two trajectory sequences is computed to determine the shortest matching path from the upper-
left corner to the lower-right corner of the matrix. The sum of the weights of the connections on the
optimal path represents the DTW distance.

 0 m=n=0
  m = 0orn = 0


DTW (T1 , T2 ) =   DTW ( Rest (T1 ), Rest (T2 )) (5)
dist (a , b ) + min  DTW ( Rest ( A), B)
 1 1 
  DTW ( A, Rest ( B))


p11
p5
p4 p10
p6
p1 p3
p2 p7 p9
T1
p8

Mij

q10
q6
q2 q9
T2 q5 q7
q1 q3
q4 q8

Fig. 4 Schematic of DTW calculation.

p11

p10

p9

p8

p7

p6

p5

p4

p3

p2

p1
T1
T2 q1 q2 q3 q4 q5 q6 q7 q8 q9 q10

Fig. 5 Diagram of DTW time-warping distance.

The calculation process of the DTW algorithm is as follows:
1) For two trajectories with assumed lengths of n and m, they are represented as point
sequences of 𝑇raj1=[p1, p2,..., pn] and 𝑇raj2=[q1, q2,..., qm], respectively.
 2) An n × m distance matrix D is defined, where D(i, j) represents the distance between point
pi and point qj.
3) An n × m cumulative distance matrix C is initialized, where C(i, j) represents the minimum
distance between the partial trajectories from point p1 to point pi and from point q1 to point qj.
4) The first line of C and the first line D are initialized, such that C(1,1) = D(1,1), C(i, 1) = ∞,
and C(1, j) = ∞, where the range of i is [2, n] and the range of j is [2, m].
5) In i∈[2, n], j∈[2, m], the value of C(i, j) is calculated. Among them, D(i, j) represents the
distance between point pi and point qj, whereas min{C(i−1, j), C(i, j−1), C(i−1, j−1)} represents the
minimum value of three adjacent cells.
6) The final DTW distance is C(n, m).
Because ship AIS data contains dynamic trajectory information, including position, course, and
speed, and considering that the course and speed in the trajectory information will also have a certain
impact on ship trajectory similarity, the AIS data is expanded to include position, speed, and course.
This expansion is achieved by using a feature vector expansion method, which extends the feature
vectors of trajectory points from two-dimensional vectors (longitude and latitude) to four-
dimensional vectors (longitude, latitude, course, and speed). This comprehensive consideration of
similarity allows for a more accurate calculation of distances between trajectory points. Assuming
XT represents the ship's navigation characteristic data at time T, the expanded feature matrix can be
expressed as follows:

X T =[xt , yt , st , ct ] (6)

Each row of the expanded feature matrix represents the feature vector of a trajectory point,
where xt and yt represent the longitude and latitude, respectively. st indicates the speed, and ct
indicates the course. DBSCAN is a density-based clustering algorithm that automatically discovers
clusters of any shape in data and identifies noise points. In this study, the improved DTW distance
algorithm is utilized as a replacement for the ε-field in DBSCAN clustering.
2.3. Ship trajectory prediction and anomaly detection
A ship trajectory prediction model based on the transformer architecture is developed using the
normal trajectory model obtained from DBSCAN clustering as the foundation for data.
Simultaneously, the criteria for trajectory anomaly detection are translated into feature deviation
values, encompassing position, speed, and course. Real-time ship trajectory anomalies are detected
by assessing the predictability of normal trajectories and the unpredictability of abnormal
trajectories.
2.3.1 Transformer model
The proposed transformer model in this study incorporates several key components, including
position coding, a multihead attention mechanism, residual connections, and normalization (Zhang
et al, 2023).
(1) Position coding
The transformer architecture, being a parallel input model that lacks the sequential iteration
advantage of RNN structures, necessitates the introduction of position coding to incorporate
positional information. The input position information is expressed as follows:

 PC( pos ,2i ) = sin( pos /100002i / dmod el )

 (7)
 PC( pos ,2i +1) = cos( pos /10000
2 i / dmod el
)
where pos is the position of sample, dmodel is the total latitude of the input feature, i is the latitude
label of the input feature, and the value range of i∈[0, dmodel/2].
(2) Multihead attention mechanism
The main component of the transformer model is built using a multihead attention mechanism,
which comprises the self-attention mechanism. The self-attention mechanism operates on input
sequences by attending to different positions within the same sequence. In Fig. 6, the self-attention
mechanism is exemplified using a2. In this mechanism, three parameters are generated: q, k, and v.
For each q, calculations are performed with k from a1, a3, and a4 to produce attention scores. These
scores are multiplied by their respective input v, resulting in four extracted vectors. These vectors
are then added together to obtain the output, b2, produced by the self-attention model for a2.

Fig. 6 Diagram of b2 calculation.

b2 =  a2, i vi (8)
i

 Q = XW Q
 (9)
 K = XW
K

 V = XW V

The attention score matrix (Fig. 7) composed of α is calculated using Q and K. Subsequently,
the final output matrix of attention layer comprising b is calculated using V and the attention score
matrix. The output result is recorded as Attention(Q, K, V):

QK T
Attention(Q, K , V ) = softmax( )V (10)
dk
where Q is the query matrix, K is the key matrix, V is the value matrix, and X is the input matrix; W,
Q, K, and V are linear transformation weight matrices, and softmax is used to calculate the weight.
dk is the dimension of input data.
 Fig. 7 Calculation process of attention score.
The calculation of the sub-attention mechanism involves applying linear transformations to the
K, Q, and V matrices. These linear transformations serve to capture different aspects of the input
and enable the model to focus on multiple levels of information. After each self-attention operation,
the outputs from multiple attention heads are spliced together to form the final output. This allows
the model to integrate information and enhance its overall performance. The multiple-head attention
mechanism, denoted as MultiHeads, can be represented as follows:

MultiHeads(Q, K ,V ) = Concat (head1 ,, headh )W o (11)

headi = Attention(QWi Q , KWi K ,VWiV ) (12)

where Wi Q , Wi K , and WiV are the linear transformation weight matrices of the i-th head of Q,

K, and V, respectively, and Wo is the linear transformation weight matrix after the multihead
attention matrix is concatenated, Concat is a feature concatenation function.
(3) Residual connection
The submodule of the transformer model primarily consists of a multihead attention
mechanism layer and a feedforward layer. Between these layers and the input layer, as well as
between the attention layer and the feedforward layer, there exist residual connections and data
normalization. Residual connections allow for the connection of input and output data, addressing
the issues of gradient vanishing and weight matrix degradation. The dimensions of the input and
output data from the attention layer are consistent, facilitating the residual connection. The residual
connection formula for the multihead attention mechanism (H) and the residual connection formula
of feedforward (H’) are as follows:

H  = X input + Attention(Q, K ,V ) (13)

H = H  + Feedforward ( H ) (14)

where Xinput is input sequence.

2.3.2 Ship trajectory prediction model based on Transformer model
In order to detect ship trajectory anomalies and construct a transformer model for predicting
ship trajectory data, a threshold-based detection method is proposed. The approach involves
defining trajectory clusters obtained through clustering as normalized ship motion trajectories
within a selected water area. These trajectories are then used as training data to train the constructed
transformer model, enabling the transformer to learn the normal ship motion trajectory model and
predict future trajectory points using the historical trajectory point data of the target ship. The
transformer network, composed of attention mechanisms, features excellent parallel capability and
omits the processing steps of previous historical experience. The specific process steps are
illustrated in Fig. 8 and can be summarized as follows:
(1) Dataset construction: The dataset consists of ship trajectory data following clustering
analysis. Trajectories that can participate in clustering are selected as normalized motion trajectories
within the specified area to establish the dataset. The dataset is divided into a training set and a test
set, with an 8:2 ratio based on the ship's MMSI number. Each ship's trajectory data includes
longitude, latitude, speed, and course, and a single trajectory's feature value at time t is denoted as
Xt = [xt, yt, st, ct].
(2) Data formatting: To simplify the data, the time interval between trajectory data points is
adjusted to 1 min and the data is normalized. The deviation standardization method is employed for
normalization, and the transformer model is used to normalize the data obtained from ship trajectory
feature prediction. The standardization formula for deviation is as follows:
X − Min (15)
X* =
Max − Min
where Max and Min are the maximum value and minimum value in the sample data, respectively. X
is the original training data, and X* is the normalized data.

Fig. 8 Trajectory prediction process based on Transformer model.

(3) Step size and output: In order to achieve multivariate prediction with the transformer model,
the original data is first transformed into a supervised learning dataset. The dataset is then converted
into trajectory feature data {Xt, Xt+1,..., Xt+n} to serve as the input for the transformer model. Herein,
the input consists of the trajectory feature values of 10 consecutive time points, whereas the output
corresponds to the trajectory feature values of the 11th time point. This setup allows the model to
be trained on the input–output pairs and predict future ship trajectories.
When constructing the transformer model, the original encoder and decoder of the transformer
is not used, and the decoder is replaced with a fully connected layer. The general process for
prediction is as follows:
1) Input encoding: The input sequence is passed through an encoder, generating a series of
encoding vectors. Each encoding vector encapsulates a portion of information from the input
sequence.
2) Initialization of the output: A special starting symbol is added to the output sequence, and it
is converted into a vector representation using an embedding layer. This vector serves as the output
for the first time step.
3) Prediction of the output sequence: Starting from the first time step, the output vector and the
encoding vector from the current time step are fed into the fully connected layer. The fully connected
layer processes these vectors and generates the output vector for the next time step. This process is
repeated until the length of the output sequence reaches the specified maximum value or the model
produces a special ending symbol.
4) Generation of final prediction: Each output vector in the output sequence is converted into
corresponding markers to obtain the predicted output sequence.
3. Experiment and discussion
The source of the data come provided by a professional data manufacturer of China (The time
coverage is from January to June in 2022), and the format of data is shown in Table 1. The focus of
this study is the sea area near the Port of Yantai in China. To facilitate the study, AIS data from a
rectangular water area within the Yantai sea area was extracted. The latitude and longitude ranges
selected for this area are as follows: (121.37°–121.53°E), (37.53°–37.67°N). The extracted AIS data
represents the original trajectory map within this specified water area. Fig. 9 displays the extracted
original trajectory map, providing a visual representation of the ship trajectories observed in the
study area.

Fig. 9 Schematic of the original trajectory in the region.

Table 1 Sample table of partial ship trajectory data
 Time MMSI longitude（°） latitude（°） speed（kn） course（°）

6/11 16:01 413556760 121.390627 37.566217 8.1 82.3

6/11 16:01 413301080 121.430653 37.654277 9.8 347.3

6/11 16:01 636017220 121.401017 37.591267 9.9 172

6/11 16:01 413020510 121.442738 37.587440 9.2 222.1

6/11 16:01 414760000 121.441682 37.587040 9.3 224.8

6/11 16:01 413556760 121.391107 37.566290 8.1 81.3

6/11 16:01 413020510 121.442338 37.587088 9.2 222.1

6/11 16:01 413020540 121.439147 37.581850 4.1 16.4

6/11 16:01 414760000 121.441253 37.586712 9.3 226.4

3.1 AIS data acquisition and preprocessing

To address information errors in AIS data, this research utilizes interpolation and sub-trajectory
partitioning methods to handle cases of missing speed, course, and trajectory points. Speed and
course errors are directly supplemented using interpolation. However, for cases of missing trajectory
points, it is necessary to distinguish between trajectory drift and trajectory point missing. Although
the reasons for these abnormalities are similar, the difference lies in the duration of missing
trajectory points. To differentiate between them, a clear threshold of 10 min is employed. If the time
interval between adjacent points in a ship's trajectory with the same MMSI is less than 10 min,
interpolation is used to supplement the missing points. However, if the interval exceeds 10 min, the
trajectory is divided into two different sub-trajectories. It should be noted that sub-trajectory division
may result in some sub-trajectories containing only single points or a small number of trajectory
points. As these sub-trajectories do not effectively reflect the motion characteristics of the ship, only
ship motion trajectories containing more than 100 data points are considered for further analysis in
this research. The flowchart illustrating the process of sub-trajectory division is presented in Fig. 10,
and Fig. 11 displays the trajectory image after the data cleaning procedure.

Start

Sort by MMSI
identification code Delete the error data of MMSI
and receiving time

Calculate the interval between

adjacent trajectory points

Save the previous trajectory of

Determine if it is greater
this trajectory point as a sub Yes
than 10 minutes
trajectory

Save all trajectories

and then output
 Fig. 10 Flowchart of sub-trajectory division.

Fig. 11 Trajectory after data cleaning.

To demonstrate the effect of MDL trajectory compression, two ships with MMSI 997760305
and 413556260 were selected for compression. A comparison between the uncompressed and
compressed trajectories of these ships is depicted in Fig. 12 and 13, respectively.

37.66 Pre-processing trajectory

Processed trajectory
37.64

37.62
Latitude(°)

37.60

37.58

37.56

121.38 121.40 121.42 121.44 121.46

Longitude(°)

Fig. 12 Comparison between uncompressed and compressed trajectory of MMSI 997760305.

37.66

37.64
Latitude(°)

37.62

37.60 Pre-processing trajectory

Processed trajectory
37.58

121.39 121.40 121.41 121.42 121.43 121.44

Longitude(°)

Fig. 13 Comparison between uncompressed and compressed trajectory of MMSI 413556260.

Furthermore, after applying the MDL criterion to extract feature points and compress all ship
trajectories in the sea area, the overall results are presented in Fig. 14. Both individual trajectory
data and the entire region's ship trajectory data exhibit a simplified and adjusted representation,
while preserving the original characteristics of the ship trajectories.
 Fig. 14 Schematic of trajectory after MDL criterion in Yantai Port.
3.2 Ship clustering analysis
In terms of measuring trajectory similarity, the Frechet distance, DTW distance, fast DTW
distance (Bai et al., 2023), and the improved DTW distance algorithm (after expanding trajectory
similarity) are employed. Using ship trajectory data from the Port of Yantai, the DBSCAN density
clustering algorithm is utilized for clustering.

Fig. 15 Diagram of the clustering effect based on Frechet distance.

Fig. 16 Diagram of clustering effect based on DTW distance.

Fig. 17 Schematic based on improved fast DTW distance.

 Fig. 18 Schematic based on improved DTW distance.

Table 2 Clustering number of different distance formulas

Distance similarity formula Clustering number
Frechet distance 11
DTW 12
Fast DTW 12
Improved DTW 13

Fig. 19 Main shipping trajectory in the region.

The DBSCAN algorithm was employed to cluster ship trajectories based on the four trajectory
similarity measurement methods. Each trajectory segment was considered as a core, and
surrounding objects were traversed to obtain trajectory class clusters. The resulting clustering is
depicted in the graph, where each color represents a different type of trajectory cluster. The
clustering effect based on the four trajectory similarity measurement methods was evaluated, and
the optimal clustering effect is observed. An analysis of the ship trajectories within the selected
range revealed a total of 15 main ship navigation trajectories in the area (as shown in Fig. 19). Table
2 provides a summary of the clustering results using the DBSCAN algorithm based on different
similarity measurement methods. The proposed improved DTW distance yielded 13 clusters of ship
trajectories near the Port of Yantai, closely aligning with the primary ship movement patterns. The
DBSCAN algorithm based on the Frechet distance function resulted in 11 clusters. Both fast DTW
and DTW yielded 12 clusters of ship trajectories near Yantai Port, mainly because the purpose of
fast DTW was to reduce overfitting and improve computing speed. It can be observed that the DTW
distance-based clustering outperformed the Frechet distance-based clustering within the research
sea area. The DBSCAN algorithm utilizing the improved DTW distance effectively clustered ship
trajectories into different clusters based on the shape characteristics of the trajectory space and the
position characteristics of the start and end points. This improved accuracy in clustering algorithms
is particularly valuable in areas with cross routes and high traffic intensity. The experiment
demonstrated the accuracy and effectiveness of the ship normal trajectory model established using
this method.
3.3 Abnormal detection of ship trajectory
The primary ship trajectory features include position, speed, and course, play a crucial role in
detecting abnormal ship trajectories. The evaluation of deviation values in these features allows for
the detection of trajectories that deviate from the normal trajectory model.
The position of a ship serves as the most intuitive indicator of its trajectory status and is the
most important evaluation factor in detecting abnormal ship behavior. The detection process
involves ascertaining abnormalities in the behavior of the ship’s position; in the presence of such
abnormalities, the ship’s course and speed are further examined for abnormalities. The abnormality
with respect to the position of a ship is determined based on the AIS data transmitted by the ship by
comparing the ship’s position with the predicted position and evaluating the distance deviation.
Considering various factors such as meteorological and sea conditions, ship maneuverability, and
collision avoidance behavior, it is challenging for ships to precisely adhere to their intended route
while sailing at sea. Consequently, two deviation warning thresholds have been established during
the detection of positional abnormalities: a low threshold of 150 m and a high threshold of 250 m.
Abnormal situations surpassing the high threshold trigger an alert. As depicted in Fig. 20, the target
ship’s trajectory is predicted using its current and historical trajectory points. Assuming that the next
trajectory point after X3 is X4, two warning circles are defined based on the predicted point X4.
Point X5 lies outside the warning circle due to abnormal ship course, while the abnormality of point
X6 can be attributed to abnormal speed. According to this study, if we set the abnormal threshold
for speed and course at 0.03 times of the predicted speed of 10 kn, then a normal trajectory would
fall within the range of 9.7–10.3 kn for the actual speed.
 Fig. 20 Schematic of abnormal trajectory.
The flowchart illustrating the process of ship trajectory anomaly detection is presented in Fig.
21. By combining the historical trajectory of the ship with the transform model trained based on
historical ship trajectories within the region to obtain the predicted trajectory of the target ship, the
position of a ship trajectory status is detected for abnormal ship behavior, in the presence of such
abnormalities, the ship’s course and speed are further examined for abnormalities. To evaluate the
effectiveness of the experiment, two abnormal trajectories featuring abnormal steering and
acceleration are selected. Fig. 22 and 23 depict the instances of abnormal steering and abnormal
acceleration during ship sailing.

Fig. 21 Anomaly detection process of ship trajectory.

 Fig. 22 Schematic of abnormal steering trajectory (trajectory 1#).

Fig. 23 Schematic of abnormal acceleration trajectory (trajectory 2#).

Fig. 24 illustrates the detected abnormal trajectory points during abnormal ship turning, which
are marked with yellow circles as warning indicators. To assess the effectiveness of the detection,
the differences between the real trajectory and the predicted trajectory are displayed in Fig. 25, 26,
and 27. In Table 3, it can be observed that four abnormal trajectory points are successfully detected
during the abnormal ship turning. However, according to the set course anomaly threshold,
trajectory number 15 should also be detected as abnormal (assuming the starting point of the input
trajectory is identified as 0 in the anomaly detection); however, it is not detected.
 37.66

37.64
Latitude(°/N)

37.62

37.60

37.58

37.56

121.38 121.40 121.42 121.44 121.46

Longitude(°/E)

Fig. 24 Abnormal detection diagram of ship steering.

450
Distance difference from predicted

400 373.54

350 315.8

300
255.25
point(m)

250 230.06

200
145.51 141.28 139.54
150 133.54 128.71
118.04 117.1
107.12 112.42 113.46 111.01
101.24
91.06 89.5 96.37
100 82.32 79.5 83.55
66.35 72.09 69.42
64.55 68.92 72.57
54.32
52.46 61.05
48.66 50.88 46.51 45.58
50
0
1 2 3 4 5 6 7 8 9 1011121314151617181920212223242526272829303132333435
Sequence of points

Fig. 25 Schematic of the distance difference between the real trajectory and the predicted
trajectory of ship steering.
Course difference from predicted

14 12.34
12 10.96

10
point(°)

8.13
8 6.81

6 4.684.564.89 4.58
4.29
4 3.13
2.34
2 0.75
1.280.96 0.861.010.79 1.060.96 1.241.49
0.10.110.470.120.190.4 0.320.190.220.38 0.20.320.06 0.38
0
1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930313233343536
-2
Sequence of points

Fig. 26 Schematic of the course difference between the real trajectory and the predicted
trajectory of ship steering.
 4
3.5
Speed difference from predicted 3 3.2

2.5
point(kn)

2 2.27 2.232.32
2
1.5
1
0.5
0.46
0
0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.10.050.050.050.050.050.050.160.150.150.150.150.140.140.140.140.140.140.130.130.13 0.12
1 2 3 4 5 6 7 8 9 1011121314151617181920212223242526272829303132333435
-0.5
Sequence of points

Fig. 27 Schematic of the speed difference between the real trajectory and the predicted
trajectory of ship steering.

Table 3 Detection results when the ship turns abnormally.

The serial number of the Distance difference from
Course Speed Detection result
trajectory predicted position
12 373.54 214.96 16.9 Warning
13 230.06 202.91 16.1 Warning
14 315.80 189.32 15.2 Warning
15 133.54 190.83 14.8 No warning
17 255.25 202.09 14.4 Warning
To analyze the second trajectory with the acceleration anomaly, the proposed anomaly
detection method from this research is employed, resulting in the outcomes displayed in Fig. 28 to
31. Fig. 28 demonstrates the detection of abnormal trajectory points—marked with yellow circles—
as a warning for abnormal ship acceleration. In order to assess the effectiveness of the detection,
three discrepancies between the real trajectory and the predicted trajectory are presented in Fig. 29,
30, and 31. Table 4 illustrates that five abnormal trajectory points are successfully detected during
the abnormal ship turning. However, the speed anomaly with a serial number of 34 is not detected,
even though it exhibits abnormal behavior.
37.64

37.63

37.62
Latitude(°/N)

37.61

37.60

37.59

37.58

121.38 121.40 121.42 121.44 121.46 121.48 121.50 121.52

Longitude(°/E)
Fig. 28 Abnormal detection diagram of ship acceleration.
 Distance difference from predicted
300 264.08
250 207.42
200 164.1
157.05 157
point(m)

150 113.55
110.46 110.06
95.98 95.98 95.56 91.91 86.84 87.38 85.85 99.17 99.75
94.16
84.56 80.73
75.04 75.14 81.82 81.03 85.34
83.32 77.74
100 66.6
64.32 68.26
58.11
74.37 72.72
54.26 46.62 54.16
38.16 31.32 33.42 38.67
38.07
50
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41
sequence of points

Fig. 29 Schematic of the distance difference between the real trajectory and the predicted
trajectory of ship acceleration.
3
2.47
Course difference from predicted

2.5
1.91
2 1.71 1.73 1.66 1.7
point(°)

1.5
1.16
1.04
0.89
1 0.75
0.71
0.65 0.67
0.5 0.57 0.52
0.47
0.5 0.31
0.24 0.27 0.29
0.19 0.2 0.26 0.28
0.15
0.05 0.13 0.15
0.04
0.04 0.11 0.16 0.09 0.09
0.01 0 0.03 0.01
0.01
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41
-0.5
Sequence of points

Fig. 30 Schematic of the course difference between the real trajectory and the predicted
trajectory of ship acceleration.
3.5 3.2
Speed difference from predicted

3
2.5 2.32
2.27 2.23
2
2
point(kn)

1.5
1
0.46
0.5 0.16 0.15
0.10.10.10.10.10.10.10.10.05
0.05
0.05
0.05 0.05 0.15
0.05 0.15
0.15
0.14
0.14
0.14
0.14
0.14
0.14
0.13
0.13
0.13 0.120.10.10.10.140.1
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
-0.5
Sequence of points

Fig. 31 Schematic of the speed difference between the real trajectory and the predicted
trajectory of ship acceleration.

Table 4 Detection results when the ship accelerates abnormally.

The serial Distance difference
number of the from predicted Course Speed Detection result
trajectory position
 29 207.42 35.5 11.1 Warning
30 264.08 41.5 13.3 Warning
31 157.08 44.1 14.6 Warning
32 164.09 43.8 15.6 Warning
33 157.00 51.1 14.9 Warning
34 110.06 50.9 13.3 No warning
The detection accuracy of the ship trajectory anomaly detection model based on transformer
prediction can be calculated using the formula: DR = 1 − (N/M × 100%), where M represents the
total number of abnormal trajectory points and N represents the number of undetected abnormal
trajectory points in the anomaly detection results. For ship steering anomaly trajectories, the
detection accuracy is determined to be 80%, indicating that 80% of the abnormal trajectory points
were correctly identified by the model. Similarly, for ship acceleration anomaly trajectories, the
detection accuracy is 83.3%, indicating that 83.3% of the abnormal trajectory points were accurately
detected by the model. These high accuracy rates within the water area demonstrate the effectiveness
of the model in assisting maritime safety administrations in identifying abnormal ship behavior.
For the above two test trajectories, the proposed method is used to predict the ship trajectory
compared with the traditional LSTM network, and the detailed results of trajectory prediction are
shown in Table 5 and Table 6 respectively. It can be seen that the proposed method has the better
predictive performance (i.e., the minor error) in two test trajectories on three evaluation indexes
(position, course and speed). Among the 6 results of trajectory 1# adopting proposed method, the
Mean error and Maximum error value has increased by 14.05%, 24.84%, 51.04%, 13.1%, 52.78%
and 5.9% in position, course and speed, respectively. Similarity, for trajectory 2#, the Mean error
and Maximum error value has increased by 15.62%, 32.2%, 30.38%, 18.21%, 61.3% and 4.48% in
position, course and speed, respectively.
Table 5 Prediction error based on transformer model
Trajectory trajectory 1# trajectory 2#
features Mean error Maximum error Mean error Maximum error
Position 112.56 373.54 89.7 264.08
Course 2.11 12.34 0.55 2.47
Speed 0.34 3.2 0.41 3.2

Table 6 Prediction error based on LSTM model

Trajectory trajectory 1# trajectory 2#
features Mean error Maximum error Mean error Maximum error
Position 130.96 497.0 106.3 389.45
Course 4.31 14.2 0.79 3.02
Speed 0.72 3.4 1.06 3.35

4. Conclusion
In this research, a ship trajectory anomaly detection model is developed, taking into account
the characteristics of AIS trajectory data in maritime traffic research. The proposed model includes
an improved DTW algorithm that considers local trend characteristics and ship motion information,
enhancing the accuracy of trajectory similarity measurement. Additionally, a transformer trajectory
prediction model is constructed, incorporating a threshold-based anomaly detection method. The
transformer model is trained using normalized motion trajectories obtained through DBSCAN
cluster analysis, enabling it to predict ship trajectories based on historical data. The effectiveness of
the proposed method is demonstrated through AIS data from the Port of Yantai, highlighting its high
accuracy in detecting abnormal ship trajectories. Future research endeavors may involve considering
the impact of meteorological environments, as such information can influence ship clustering and
anomaly analysis. In addition, enhancing the adaptability of method parameters is one of the
important research interests in the future.
Disclosure statement
No potential conflict of interest was reported by the authors.
Acknowledgements
This work was supported by the Applied Basic Research Plan of Liaoning Province in 2022
(2022JH2/101300265); the Fundamental Research Funds for the Dalian maritime university
(017231034); 2023 DMU navigation college first-class interdisciplinary research project
(2023JXB14).
References
Bai, Xiangen, Xie, Zhexin, Xu Xiaofeng, Xiao, Yingjie, 2023. An adaptive threshold fast DBSCAN
algorithm with preserved trajectory feature points for vessel trajectory clustering. Ocean
Engineering. 280, 114930.
Dahlbom, Anders, and Lars, Niklasson, 2007. Trajectory clustering for coastal surveillance. 2007
10th international Conference on Information Fusion. IEEE, 1-8.
Dai, Jianhua, Liu, Ye, Chen, Jiaolong, Liu, Xiaofeng, 2020. Fast feature selection for interval-valued
data through kernel density estimation entropy. International Journal of Machine Learning and
Cybernetics. 11(12), 2607-2624.
D’Afflisio, Enrica, Braca, Paolo, Willett, Peter, 2021. Malicious AIS spoofing and abnormal stealth
deviations: a comprehensive statistical framework for maritime anomaly detection. IEEE
Transactions on Aerospace Electronic Systems. 57(4), 2093-2108.
Ester, Martin, Kriegel, Hans-Peter, Sander, Jorg, Xu, Xiaowei, 1996. A density-based algorithm for
discovering clusters in large spatial databases with noise. In: Kdd. Vol. 96. No. 34. pp. 226-231.
Gözde, Boztepe Karataş, Pinar, Karagoz, Orhan, Ayran, 2021. Trajectory pattern extraction and
anomaly detection for maritime vessels. Internet of Things. 16, 100436.
Han, Xuyang, Armenakis, Costas, Jadidi, Mojgan, 2021. Modeling vessel behaviours by clustering
AIS data using optimized DBSCAN. Sustainability. 13(15), 8162.
Jin, Chen, Hao, Chen, Quan, Chen, Xin, Song, Wang, Hongdong, 2023. Vessel sailing route
extraction and analysis from satellite-based AIS data using density clustering and probability
algorithms. Ocean Engineering. 280, 114627.
Kontopoulos, Ioannis., Varlamis, Iraklis, Tserpes, Konstantinos, 2021. A distributed framework for
extracting maritime traffic patterns. International Journal of Geographical Information Science.
35(4), 767-792.
Kowalska, Kira, Peel, Leto, 2012. Maritime anomaly detection using Gaussian process active
learning. In: 2012 15th International Conference on Information Fusion, pp. 1164-1171.
Laxhammar, Rikard, Falkman, Gӧran, Sviestins, Egils, 2009. Anomaly detection in sea traffic-a
comparison of the Gaussian mixture model and the kernel density estimator. In: 2009 12th
International Conference on Information Fusion, pp. 756-763.
Lei, Po Ruey, 2016. A framework for anomaly detection in maritime trajectory behavior.
Knowledge and Information Systems. 47(1), 189-214.
Li, Kexin, Guo, Jian, Li, Ranchong, Wang, Yujun, Li, Zongming, Miu, Kun, Chen, Hui. 2023. The
Abnormal Detection Method of Ship Trajectory with Adaptive Transformer Model Based on
Migration Learning. In: 2023 4th International Conference on Spatial Data and Intelligence, pp.
204-220.
Li, Huanhuan, Liu, Liu, Jingxian, Yang, Zaili, Liu, Ryanwen, Wu, Kefeng, Wan, Yuan, 2020.
Adaptively constrained dynamic time warping for time series classification and clustering.
Information Sciences. 534, 97-116.
Liu, Ryanwen, Liang, Maohan, Nie, Jiangtian, Yuan, Yanli, Xiong, Zehui, Yu, Han, Guizani, Nadra,
2022. STMGCN: Mobile edge computing-oempowered vessel trajectory prediction using spatio-
temporal multigraph convolutional network. IEEE Transactions on Industrial Informatics. 18(11),
7977-7987.
Mantecón, Tomas, Casals, David, Navarro-Corcuera, Juanjosé, del-Blanco, Carlos. R., Jaureguizar,
Fernando. 2019. Deep learning to enhance maritime situation awareness. In 2019 20th International
radar symposium. IEEE, 1-8.
Mascaro, Steven, Nicholso, Ann, Korb, Kevin, 2014. Anomaly detection in vessel tracks using
Bayesian networks. International Journal of Approximate Reasoning. 55(1), 84-98.
Nooshin, Hanafi, Hamid, Saadatfar, 2022. A fast DBSCAN algorithm for big data based on efficient
density calculation. Expert Systems with Applications. 203, 117501.
Park, Jinwan, Jeong, Ju-kwon, Park, Youngsoo, 2021. Ship trajectory prediction based on Bi-LSTM
using spectral-clustered AIS data. Journal of Marine Science and Engineering. 9(9), 1037.
Ristic, Bronko, La Scala, B., Morelande, M., Gordon, N., 2008. Statistical analysis of motion
patterns in AIS data: anomaly detection and motion prediction. In: 2008 11th International
Conference on Information Fusion, pp. 1-7.
Rong, Hao, Teixeira, A. P., Soares, Carlos Guedes, 2020. Data mining approach to shipping route
characterization and anomaly detection based on AIS data. Ocean Engineering. 198, 106936.
Shahir, Hamed Yaghoubi, Glässer, Uwe, Nalbandyan, Narek, 2014. Maritime situation analysis: a
multi-vessel interaction and anomaly detection framework. In: 2014 IEEE Joint Intelligence and
Security Informatics Conference, pp. 192-199.
Sheng, Pan, Yin, Jingbo, 2018. Extracting shipping route patterns by trajectory clustering model
based on automatic identification system data. Sustainability 10(7), 2327.
Smith, Mark, Reece, Steven, Roberts, Stphen, Rezek, Iead., 2012. Online maritime abnormality
detection using Gaussian processes and extreme value theory. In: 2012 IEEE 12th International
Conference on Data Mining, pp. 645-654.
Soares Júnior, A., Moreno, Moreno, B. N., Times, V. C., Matwin, S., Formiga Cabral, L. D. A.,
2015. GRASP-UTS: an algorithm for unsupervised trajectory segmentation. International Journal
of Geographical Information Science, 29(1-2), 46-68.
Tang, Chunhua, Chen, Meiyue, Zhao, Jiahuan, Liu, Tao, Liu, Kang, Yan, Huaran, Xiao, Yingjie,
2021. A novel ship trajectory clustering method for finding overall and local features of ship
trajectories. Ocean Engineering. 241, 110108.
Venskus, Julius, Treigys, Povilas, Bernatavičienė, Jolita, Medvedev, Viktor, Voznak, Miroslav,
Kurmis, Mindaugas, Bulbenkiene, Violeta, 2017. Integration of a self-organizing map and a virtual
pheromone for real-time abnormal movement detection in marine traffic. Informatica. 28(2), 359-
374.
Venskus, Julius, Treigys, Povilas, Bernatavičienė, Jolita, Tamulevičius, Gintautas, Medvedev, Viktor,
2019. Real-time maritime traffic anomaly detection based on sensors and history data embedding.
Sensors. 19(17), 3782.
Wang, Jiecheng, Liu, Yantong, Chang, Jincai, 2022. An improved model for kernel density
estimation based on quadtree and quasi-interpolation. Mathematics. 10(14), 2402.
Wang, Siwen, Li, Ying, Hu, Xing, 2023. A novel method for ship trajectory prediction in complex
scenarios based on spatio-temporal features extraction of AIS data. Ocean Engineering. 281, 114846.
Wen, Yuanqiao, Sui, Zhongyi, Zhou, Chunhui, Xiao, Changshi, Chen, Qianqian, Dong, Han, Zhang,
Yimeng, 2020. Automatic ship route design between two ports: a data-driven method. Applied
Ocean Research. 96, 102049.
Yang, Dong, Wu, Lingxiao, Wang, Shuaian, Jia, Haiying, Li, Kevin X, 2019. How big data enriches
maritime research–a critical review of Automatic Identification System (AIS) data applications.
Transport Reviews. 39(6), 755-773.
Yang, Jiaxuan, Liu, Yuan, Ma, Lingqi, Ji, Chengtao, 2022. Maritime traffic flow clustering analysis
by density based trajectory clustering with noise. Ocean Engineering. 249, 111001.
Zhang, Yuqi, Li, Chaofeng, Shang, Shaopeng, Chen, Xinqiang, 2023. SwinSeg: Swin transformer
and MLP hybrid network for ship segmentation in maritime surveillance system. Ocean Engineering.
281, 114885.
Zhao, Liangbin, Shi, Guoyou, 2019. Maritime anomaly detection using density-based clustering and
recurrent neural network. Journal of Navigation. 72(4), 894-916.
Zhen, Rong, Jin, Yongxing, Hu, Qinyou, Shao, Zheping, & Nikitakos, Nikitas, 2017. Maritime
anomaly detection within coastal waters based on vessel trajectory clustering and naїve bayes
classifier. Journal of Navigation. 70(3), 1-23.