Real-time Collision Avoidance Considering Ship Position Uncertainty Based on AIS Data
Article information
Abstract
Autonomous ships require reliable situational awareness to ensure safe navigation, particularly during collision avoidance in congested maritime environments. Although an automatic identification system (AIS) is widely used for vessel traffic awareness, its update intervals, communication delays, and measurement noise introduce positional uncertainty, which limits its direct application to real-time collision avoidance. This study aims to develop a collision avoidance framework that incorporates AIS-derived position uncertainty from the perspective of an autonomous ship. An unscented Kalman filter is applied to estimate the nonlinear motion and positional uncertainty of surrounding vessels in real time to compensate for information gaps in AIS data. Using the predicted vessel states, collision risk is quantitatively evaluated through a collision risk index, and avoidance maneuvers are generated using the velocity obstacle method. The framework is validated through simulations in which positional uncertainty is introduced deliberately, and vessel maneuvering behavior is represented using a dynamic model derived from the maneuvering modeling group theory. Simulation results show that the proposed system improves the consistency of collision risk assessment and estimates target vessel trajectories with an average positional accuracy of ~3 m. This corresponds to an ~96% reduction in estimation error compared to that obtained using nonpredictive approaches.
1. Introduction
The International Maritime Organization (IMO) classifies autonomous ships into four categories: Level 4, which is the highest level, refers to fully autonomous navigation without human intervention. Accurately identifying the surrounding traffic situation, which leads to precise decision-making and accurate maneuvering control, is essential for operating a highly reliable navigation system. Autonomous ships use various sensors and navigation equipment for improving situational awareness. Among these, the most critical source of information is the automatic identification system (AIS) data, which has become mandatory under the SOLAS Convention established in 2000 for enhancing maritime navigation safety. However, data uncertainty persists because of the inherent characteristics of the equipment. Accordingly, this study focused on compensating for the position uncertainty in AIS data that occurs when an autonomous ship identifies a target ship requiring collision avoidance. Position uncertainty refers to inaccuracies arising from equipment errors, communication delays, and environmental noise mixed into data generated when the navigation equipment detects a target ship, thereby making it difficult to accurately recognize the actual position, speed, and other motion-related information of the ship. In addition, ships perform nonlinear maneuvers, which causes position uncertainty to accumulate in unpredictable directions over time. Therefore, the direct use of raw perception data for collision-avoidance path planning without preprocessing is not recommended.
Thus far, numerous studies have focused on collision avoidance for autonomous ships. Kim (2019) designed a safe collision avoidance system using a distributed probabilistic search algorithm, whereas Lee et al. (2025) developed a safe collision avoidance system by applying the velocity obstacle (VO) method, which is widely used in robotics. Seo et al. (2025) investigated the timing of collision-avoidance initiation using real-world AIS operational data. Decision tree models were trained using AIS data to identify operational rules for determining avoidance timing by incorporating navigator expertise. However, these studies focused on generating avoidance trajectories, and therefore, they did not consider the position uncertainty associated with the motion information of target ships used as input data.
The utility of AIS, which enables recording and tracking all aspects of ship operations, has been widely recognized. Accordingly, various studies have been utilized AIS data. For example, Lee et al. (2019) examined the applicability of current AIS systems to future applications and emphasized the importance of using AIS as a data source. Last et al. (2014) argued that the availability of large-scale accumulated historical AIS data remains insufficient and that reporting intervals should be improved. Further, Lee et al. (2010) investigated the systematic management and utilization of ship information by designing and implementing a database for maritime information applications.
Studies on ship position estimation have also conducted to compensate for the inherent limitations of the AIS equipment. These studies include research on ship position estimation and collision avoidance considering position uncertainty using the unscented kalman filter (UKF). Chen et al. (2018) utilized AIS data as input and applied the time-discrete nonlinear velocity obstacle method to account for the position uncertainty of target ships and evaluate potential collision scenarios between data points. Wang et al. (2023a) proposed a collision probability estimation method for multiple ships while considering ship position uncertainty. They predicted future ship trajectories by combining gated recurrent units with UKFs; further, they validated the early warning capability for collision risk among multiple ships using AIS data collected from Victoria Harbor. Jeong et al. (2023) compared position tracking accuracy using UKF integrated with the Lie group theory. After correcting the positions of target ships collected through AIS, they confirmed an error of 14.7 m compared to that of the actual trajectories. Wang et al. (2023b) predicted ship motion states by integrating two motion models, the curvilinear motion model (CMM) and constant turn rate and acceleration, with UKF and the extended Kalman Filter (EKF). Through computer simulations, the performance of the filters combining motion models and Kalman filters was validated, and consistent results were observed, except for the CMM/EKF combination. In studies related to position uncertainty estimation, research has focused on estimating ship positions within an accuracy of several meters and utilizing such estimates for target ship tracking or warning of accident-prone areas. However, many existing studies were conducted from the perspective of port operations, and therefore, research focusing on ship-to-ship collision avoidance and tracking remains limited.
Accordingly, this study aims to develop a collision avoidance system that recognizes and avoids target ships from the perspective of an autonomous ship while overcoming hardware limitations such as sensor constraints and operating safely and robustly even in environments with noise or unstable communications, including missing measurements. UKF is employed to compensate for uncertainty caused by information gaps in AIS data and estimate ship motion in real time. The estimates calculated using the filter incorporate the position uncertainty of the actual ships. Collision-avoidance paths are generated based on these estimated states and subsequently used to construct a collision-avoidance system that reflects a higher level of situational awareness. The proposed algorithm is validated through computer simulations. The input data for the target ships are configured in the same manner as AIS data, and information transmission delays are intentionally introduced to reproduce AIS update intervals. Ship motion is implemented using a maneuvering modeling group model to establish a realistic simulation environment. This approach increases the applicability of the proposed system to actual ship operations. Collision risk is calculated using the dynamic information of target ships estimated by UKF, and collision avoidance paths are generated accordingly. The collision risk index (CRI) is employed to quantify the collision risk, and the VO method is applied as the vector selection strategy for collision avoidance. A PID controller is used to ensure stable ship control.
2. Proposed System
This section describes the proposed collision avoidance system that considers position uncertainty in AIS data. This section follows the overall data flow of a collision avoidance system.
2.1 Automatic Identification System (AIS)
An AIS is installed on most ships navigating at sea. It is responsible for transmitting dynamic and static ship information to nearby vessels, and the shore-based vessel traffic service (VTS) is responsible for the waters in which the ship is operating. AIS messages are categorized into 27 message types and contain static information (such as ship name and specifications) as well as dynamic information (including heading, speed, and message timestamp) representing the current navigation status. Only Message Types 1, 2, and 3 are considered because only dynamic ship information is required in this study. Message Type 1 is used to automatically transmit ship information of the own ship at predefined intervals. Message Type 2 is transmitted at predefined intervals to the shore-based VTS. Finally, Message Type 3 is transmitted to a shore-based VTS in the operating region when an information transmission request is received. The primary information contained in these messages includes the speed over the ground, latitude, longitude, and heading. Detailed descriptions of the information used in this study are presented in Table 1.
In addition, AIS has different data update intervals depending on the operational status of the equipped vessel. This difference in update intervals can result in information gaps when autonomous ships receive data. This can generate unexpected system errors because the AIS update cycle differs from the route calculation cycle of autonomous navigation systems. Thus, the impact of information gaps can be minimized by identifying and accurately estimating the information update interval. The standard AIS information update intervals are listed in Table 2.
The applicability of autonomous navigation systems can be expanded by utilizing AIS data containing the information described above as input data. Consequently, the AIS can play a significant role in enabling ships to recognize target vessels.
2.2 Unscented Kalman Filter (UKF)
In the EKF developed to address nonlinear problems, nonlinear functions are approximated through linearization using a first-order Taylor series expansion. However, this approach introduces errors during the linearization process, making it difficult to represent nonlinear functions accurately, potentially leading to divergence issues in the algorithm. To overcome these limitations, Julier et al. (1995) proposed the UKF. For Gaussian distributions, UKF provides second-order approximation accuracy for process and measurement noise through a Taylor series expansion. The linearization of nonlinear equations is not required because it employs Unscented Transformation (UT), and therefore, the mean and covariance of multiple samples are estimated using sigma points.
The UT algorithm serves as the core mechanism of the UKF. If the state estimate
where n represents the system dimension, and α represents a scaling parameter that determines the spread of the sigma points assigned a small positive value (1 ≤ α ≤ 10−4)). The parameter β incorporates prior knowledge of the distribution of the random variable, and when a Gaussian distribution is assumed, β = 2 yields optimal results.
The parameter λ is defined as
The parameter k is a secondary scaling parameter and is set to zero in this study.
The obtained (2n+1) sigma points are propagated through the system state equation to estimate the state and covariance of x. This process can be described using the following equations.
where Zi, xi, n, z̄, and PZ represents the transformed sigma point obtained by propagating the i-th sigma point through the system state equation f(·), ith sigma point generated at time step k, dimension of the state vector, mean of the estimated measurements, and measurement–estimation covariance, respectively. The UKF is constructed by applying this UT algorithm within the Kalman filtering framework.
In this study, the UKF was employed to estimate the nonlinear motion of floating ships and synchronize the information update interval of the target ship with the collision avoidance path generation cycle of the own ship. The proposed approach determines whether the filter input values have been updated. If updated values are available, the measured values are used as filter inputs; otherwise, the estimated values calculated in the previous step are reused. The update verification process is performed at a system operating frequency of 10 Hz, and it determines whether the newly received values differ from those of the previous time step and are not empty. The flowchart of the proposed UKF framework is shown in Figure 1. In the flowchart, the input data referred to as measures, comprises the latitude, longitude, speed, and heading of the target ship, as estimated by the own ship. Through the UKF, estimates representing the estimated positions during information gaps are generated as latitude, longitude, speed, and heading values. This configuration corresponds to actual ship applications. However, this study is validated in a virtual simulation environment, and therefore, the input and output data are represented using Cartesian coordinates (x and y) instead of latitude and longitude in the spherical coordinate system.
The measurement function used in this study receives the ship position coordinates x and y, speed v, and heading angle ψ. This relationship is expressed in the matrix form as
where z, H, x, and wm represent the measurement vector, measurement matrix, state vector, and measurement noise, respectively.
Euler discretization was employed for the state transition function. The mathematical formulation is presented as
where
The parameters used in the simulation are as follows: system noise covariance (Q) = 0.1, measurement noise covariance (R) = 0.01, and initial covariance (P) = 100. These parameters are determined using heuristic tuning. Q and R were selected by varying their ratios, and the highest estimation accuracy was achieved when the ratio was set to 10:1.
2.3 Maneuvering Modeling Group
The second target vessel of the Korea autonomous surface ship (KASS) project, jointly organized by the Ministry of Oceans and Fisheries and the Ministry of Trade, Industry and Energy of the Republic of Korea, is used as the ship model for simulation. The vessel is a twin-screw ship developed by the Korea Research Institute of Ships and Ocean Engineering, and it features a center skeg located at the stern center section; a three-degree-of-freedom motion model is also applied. The ship-fixed coordinate system is illustrated in Fig. 2, and the overall ship geometry and principal parameters are presented in Fig. 3 and Table 3, respectively. In this study, the coefficients obtained from the model tests are utilized for validating the simulation. Accordingly, the simulation is conducted using the same scale ratio of 1:31.65 as in the model tests.
The three-degree-of-freedom motion model was constructed by considering planar motions in surge, sway, and yaw. They can be expressed as
where m, u, v, and r represent the ship mass, surge velocity, sway velocity, and yaw angular velocity, respectively; u̇, v̇, and ṙ represent the surge, sway, and yaw angular accelerations, respectively; ur represents the lateral acceleration component caused by the forward and rotational motions; and IZZ represents the moment of inertia about the z axis. The terms xG, xGṙ, and mxG(v̇+ur) represent the longitudinal position of the ship’s center of gravity along the x-axis, lateral acceleration term generated by the coupling between the yaw acceleration and center of gravity, and yaw rotational moment transformation term of the lateral acceleration caused by the center of gravity, respectively. Further, XH, XP, and XR represent the forces acting in the x-direction generated by the hull, propeller, and rudder, respectively; YH, YP, and YR represent the forces acting in the y-direction generated by the hull, propeller, and rudder, respectively; and NH, NP, and NR represent the yaw moments generated by the hull, propeller, and rudder, respectively.
2.4 Collision Risk Index
The CRI is used to quantitatively evaluate ship collision risk and ranges from 0 to 1. A value of 0 indicates no collision risk, whereas a value of 1 represents the maximum collision risk. The CRI comprises five membership functions, namely, distance to closest point of approach (DCPA), time to closest point of approach (TCPA), relative distance (RD), relative bearing (RB), and relative speed (K). The final index is calculated by multiplying the output of each membership function by its corresponding weight. This formulation is expressed as
where W and U represent the weight assigned to each membership function and calculated value of each membership function, respectively.
The weighting factors used in this study are determined according to the method reported by Hu et al. (2020) and presented in Table 4.
2.5 Velocity Obstacle
The VO method is a concept proposed to generate avoidance paths for moving obstacles in dynamic environments. This method assumes that the velocities of the own ship and the target ship remain constant under the current state. Candidate velocity vectors available to the own ship are generated. Then, a collision cone is constructed based on the velocity and acceleration of the own ship and target ship to identify regions where collisions may occur. Subsequently, collision-avoidance paths are generated by selecting velocity vectors outside the collision cone. At this stage, avoidance paths appropriate for encounter situations are selected in accordance with the International Regulations for Preventing Collisions at Sea (COLREGs), ensuring compliance with international maritime regulations during collision avoidance maneuvers. A conceptual illustration of the VO method is shown in Fig. 4.
3. Simulation Results
This study was validated using computer simulations. The simulation environment consisted of Ubuntu 20.04, Python 3.8, and robot operating system (ROS) Noetic Ninjemys. The ship speed in the simulation was set to 14 kn, and a deliberate data delay of 10 s was introduced according to the AIS information update interval standard corresponding to this speed. All physical quantities, including the ship dimensions, were scaled at a ratio of 1:31.65, which were identical to the model ship scale ratio of the second target vessel in the KASS project. All collision avoidance maneuvers were performed in accordance with the COLREGs rules. The results were visualized using ROS Plotjuggler and Python Matplotlib.
Four scenarios were designed for validation: two simple encounter situations, namely, Head-on and Starboard Crossing scenarios, and two scenarios involving target ship course alteration. The course alteration scenarios included one case in which the target ship approached from a bearing angle of 195° and altered its course to 180°, assuming north as 0°, and another case in which the target ship approached at 180° and altered its course to 135°. These scenarios violate the COLREGs rules, and therefore, hazardous scenarios with low occurrence probabilities in real operations were intentionally included to provide broader validation conditions. Each scenario was evaluated under three cases: a normal case without delay or estimation, a delay case with only information delay, and the proposed system case employing real-time estimation using UKF. The normal case served as the baseline for comparison with the other cases.
Cases for all scenarios were compared in terms of overall avoidance trajectories, heading command variations of the own ship, and CRI values. In each figure, the upper panel presents the overall navigation trajectories of the blue autonomous ship and red target ship. The lower graphs show the heading command calculated by the autonomous ship and variation in CRI values. The green solid line represents the heading command, while the red dashed line indicates the CRI value. In addition, timestamps were marked every 100 s of simulation time to examine the ship trajectories, heading commands, and CRI values at the corresponding moments. Finally, the estimation accuracy was compared to determine the magnitude of estimation errors and identify their underlying causes for evaluating the position-estimation performance of the filter.
3.1 Scenario 1
Scenario 1 represents a simple head-on encounter situation. As shown in Fig. 5, the own ship performs an avoidance maneuver to the starboard in accordance with the COLREGs rules to avoid the target ship approaching from the forward direction. The collision risk increases as the two ships approach each other, and it reaches a peak value of ~0.9 at the closest point of approach. Subsequently, it continuously decreases after the encounter situation is resolved.
The overall trajectories of the three cases are nearly identical because both the own ship and target ship are traveling along the same line of approach. In the delay case, the own ship initiates the avoidance maneuver later than that in the normal case. Further, the CRI graph exhibits a stepwise pattern because the own ship perceives the target ship as remaining stationary at a distant location. This perception makes accurate situational awareness difficult for operators or autonomous systems and can create risks by disrupting avoidance commands.
In contrast, in the proposed system case, the trajectory of the target ship is accurately predicted, enabling the own ship to perform safe avoidance maneuvers. Trends in the heading command variation and CRI variation identical to those in the normal case were observed.
3.2 Scenario 2
Scenario 2 represents a simple starboard crossing encounter situation, as shown in Fig. 6. In accordance with the COLREGs rules, the own ship performs an avoidance maneuver toward the starboard. The CRI increases as the two ships approach each other and reaches the peak value when the relative distance is the smallest. Then, it continuously decreases after the encounter situation is resolved. The CRI values are generally higher than those in Scenario 1 because the relative distance is smaller in Scenario 2.
In the delay case, significant oscillations are observed in the heading command of the own ship throughout the avoidance process, and the resulting navigation trajectory also exhibits side-to-side fluctuations. This phenomenon occurs because the target ship moves continuously in reality, whereas the received data remain unchanged during the update interval. Consequently, the autonomous ship incorrectly recognizes the target ship as stationary at a distant location while retaining its speed information. Consequently, the system becomes unable to determine if it should avoid being behind the target ship or maintain its course based on the assumption that no collisions will occur because the target ship appears stationary. This demonstrates the criticality of information delay to autonomous navigation systems. Such situations not only increase collision risk but can also threaten control stability caused by abrupt steering actions from the perspective of ship maneuvering safety.
In contrast, the proposed system case accurately estimates trajectory gaps caused by delayed data updates, producing a continuous target ship trajectory comparable to that of the normal case instead of the point-like trajectory observed in the delay case. Accordingly, the own ship can recognize the target ship from a sufficient distance and safely execute collision avoidance maneuvers. Furthermore, the CRI exhibits trends nearly identical to those observed in the normal case.
3.3 Scenario 3
Scenario 3 represents the target ship-course alteration scenario. As shown in Fig. 7, the target ship initially approaches from a bearing angle of 195° and then alters its course to 180°, 10 s after the own ship initiates collision avoidance. This scenario represents a target ship that does not comply with the COLREGs rules, resulting in a high risk of collision. The own ship cannot anticipate the course alteration of the target ship, and therefore, it initially performs a starboard avoidance maneuver to comply with the COLREGs rules, subsequently executing an additional avoidance maneuver in response to the altered encounter situation caused by the course change of the target ship. Accordingly, the heading command graph exhibits two stepwise changes in the heading commands. Further, the CRI continuously increases, reaches a maximum value of 0.9, and then decreases.
However, in the delay case, the heading command variations become significantly larger because the own ship cannot accurately determine the position of the target ship. In addition, large oscillations can be observed when the encounter situation is resolved and the return to the global route occurs simultaneously at the point of the closest relative distance. Moreover, the CRI values exhibit a lower peak than that in the normal case and show a stepwise pattern. This indicates a highly dangerous situation from both collision risk and control perspectives.
In contrast, the proposed system case accurately estimates both the movement and course alteration of the target ship, which can be confirmed through the trajectories shown in the upper panel. The own ship performs avoidance maneuvers along trajectories identical to those in the normal case. Furthermore, the own ship exhibits increasingly conservative avoidance behavior with a decrease in the relative distance between the two ships. Further, it can be confirmed that the two stepwise heading command changes and CRI variations are consistent with those observed in the normal case.
3.4 Scenario 4
Scenario 4 is illustrated in Fig. 8. The target ship initially approaches at 180° and then alters its course to 135°, 10 s after the own ship initiates collision avoidance. In the initial head-on encounter situation, the own ship performs a starboard avoidance maneuver in accordance with the COLREGs rules, subsequently executing a larger starboard maneuver to avoid the target ship approaching toward its side. In this scenario, the course alteration of the target ship causes two stepwise changes in the heading command of the own ship, whereas the CRI increases and then decreases after reaching its maximum value of 0.9.
In the delay case, the overall avoidance process of the own ship is nearly similar; however, the initiation of the avoidance maneuver occurs later. Furthermore, when the relative distance is the smallest, the information delay introduces substantial disturbances, resulting in large fluctuations in the global route recovery command of the own ship. The CRI values also exhibit a stepwise pattern, which not only disrupts the judgment of operators and autonomous systems but also imposes a considerable burden on control stability.
In contrast, the proposed system case exhibits two stepwise heading changes identical to those of the normal case, and the calculated CRI values are consistent with those of the normal case. Although the results are not identical to those of the normal case, the proposed system demonstrates the ability to respond appropriately to target ship course alterations while accurately calculating CRI values, further confirming the necessity of prediction.
3.5 Comparison of Estimation Accuracy
The effectiveness of autonomous ship collision avoidance was evaluated by estimating the target ship positions. Although information delay causes critical errors in collision avoidance path generation for autonomous ships, such issues can be improved through state estimation. However, identifying and reducing the sources of estimation error is essential to ensure the accuracy of the estimation.
In this study, the position error of the ship was evaluated by calculating the Euclidean distance between the position of the target ship in the normal case and that in the proposed system case. Fig. 9 shows the position errors recorded at each time step for the four scenarios. In the graph, the red, blue, green, and orange solid lines represent the head-on scenario, starboard crossing scenario, scenario in which the target ship approaches 195° and alters its course to 180°, and scenario in which the target ship approaches 180° and alters its course to 135°, respectively.
As shown in Fig. 9, a maximum error of ~18 m occurs during the initial stage of the simulation in all scenarios. This error can be attributed to the characteristics of the Kalman filter-based methods, which require the convergence of the Kalman gain that determines the weighting between the measured and estimated values. These results suggest that an initial convergence period is necessary for achieving reliable filter performance. The estimation error converges within approximately 10 s and stabilizes at ~3 m. In Scenarios 1 and 2, only consistent estimation errors are observed because the target ship follows a simple encounter situation without course alteration. In contrast, Scenarios 3 and 4, which include target ship course alteration, exhibit nonuniform error variations because the information update interval changes when the target ship alters its course.
These position estimation errors are independent of ship size and demonstrate the feasibility of applying the proposed approach to full-scale ship operations. However, estimation performance can be improved further using various methods. For example, appropriate initialization and parameter selection can help reduce convergence time; however, they cannot completely eliminate estimation errors. Further, more sophisticated modeling of the state transition function can improve estimation accuracy. As demonstrated in previous studies, further reductions in estimation error may be achieved through sensor fusion with additional sensing systems such as light detection and ranging or cameras.
4. Conclusions
This study was conducted to compensate for data uncertainty in AIS, which is one of the most widely used navigation systems utilized by autonomous ships for target ship recognition. Real-time target ship position estimation was performed using UKF to address the information gaps caused by the AIS information update interval, which is an inherent characteristic of AIS equipment mandated by the IMO. In addition, a robust collision-avoidance algorithm was developed by synchronizing the AIS information update interval with the computation cycle of the autonomous navigation system. The simulation results demonstrated that the delay case consistently exhibited delayed avoidance initiation, significant oscillations in heading commands, and inaccuracies in collision risk assessment across all scenarios. These issues can interfere with situational awareness for both operators and autonomous systems and can negatively affect control stability and onboard comfort. In contrast, the proposed system generated avoidance trajectories similar to those of the normal case and accurately evaluated collision risk according to encounter situations through CRI values. After UKF convergence, the position estimation error remained within ~3 m, demonstrating a performance level applicable to full-scale ship operations. These findings indicate that information delay can significantly affect collision avoidance path generation in autonomous ships and confirm the necessity of applying ship position prediction using the UKF to mitigate such risks. Further, the proposed approach demonstrates safe and robust collision avoidance capability even under unstable communication conditions, including sensing noise, missing data, and delayed information.
One limitation of this study is that ship motion is predicted deterministically. Although this approach can effectively compensate AIS information gaps and synchronize computation cycles with input information updates, it cannot fully account for irregular maneuvers or speed changes in target ships, which potentially results in significant estimation errors under certain situations. To address this limitation, future work will extend beyond the target ship motion state estimation toward probabilistic range estimation methods for collision risk assessment and collision avoidance.
Notes
Kwang-Jun Paik serves as a journal publication committee member for the Journal of Ocean Engineering and Technology, but he did not have a role in the decision to publish this article. There are no potential conflicts of interest relevant to this article.
This research was supported by the Korea Institute of Marine Science and Technology Promotion (KIMST), funded by the Ministry of Oceans and Fisheries (Grant No. RS-2022-KS221571).
