Naslov
Dynamic Collision Avoidance for Sea Surface Vehicles with a Hidden Markov Model

Autori
Sumner, Matthew

Vrsta, podvrsta i kategorija rada
Ocjenski radovi, doktorska disertacija

Fakultet
Pomorski fakultet

Mjesto
Rijeka

Datum
24.11

Godina
2021

Stranica
452

Mentor
Rudan, Igor

Ključne riječi
dynamic collision avoidance ; hidden Markov model ; partially observable Markov decision processes ; reinforcement learning ; ship motion control ; intent-aware navigation ; early detection of collision risks

Sažetak
In this thesis an integrated dynamic collision avoidance and hazard alerting system is proposed and identified as Marine Collision avoidance and Alerting System (MCAS). It is comprised of four integrated models that aid navigators in making appropriate decisions to prevent collisions at sea. Before autonomous sea surface vehicles would be allowed to navigate on commercial routes, a robust collision avoidance system has to be developed. Even though MCAS system is feasible for autonomous navigation, development of a decision support system that can be used within the current legal frameworks is in focus. Research problem is formulated within the Hidden Markov Model (HMM) framework and solutions that are based on various Partially Observable Markov Decision Processes (POMDP) and Reinforcement Learning (RL) solvers are proposed. This approach is based on offline development of robust look-up tables, rules, and protocols that aid online computation of conflict resolutions while preserving overall feasibility and reduce computational expense. To ensure feasibility of trajectories that are generated by collision avoidance algorithms, autopilot and auto-telegraph models are developed for the motion control of sea surface vehicles. Motion control algorithms thrive under dynamic environmental loads and are capable to control underactuated sea surface vehicles. Heading, course, and throttle algorithms are proposed to allow for larger action space when avoiding collision. Even though model-free approach is envisioned, Model-Predictive Control framework is exploited to propagate signals to motion control actuators. With the intention of reducing uncertainties and improve input data stability, a non-linear dynamic state estimator is proposed, named Foraging Particle Filter, that is based on swarm algorithmic approaches, and is utilized to filter input signals to motion control algorithms. With the purpose of developing a robust and feasible collision-avoidance system, it is important to ensure that it can be used within the legal framework of collision avoidance at sea. COLREGs classification algorithm that quantifies requirements of collision regulations by reducing vagueness and uncertainties is proposed. Quantification is based on empirical studies and case laws. COLREGs classification algorithm is used to produce input signals to collision avoidance algorithm and in that way decentralize computation. In order to generate evasive trajectories, predictor is developed that takes feasibility of turns into account and ensures trajectories are hazard free. Simulation results confirmed that the proposed system is capable to avoid complex close-quarter situations. Previous research has focused on egocentric resolutions where only own vehicle is equipped with collision avoidance systems, while in this research, a holistic collision risk resolution model for multiple targets in mixed equipage situations is developed. Communication protocols are utilized to share intent and other relevant information that is required to reduce uncertainties and computational complexities of trajectory generations. Simulation results demonstrate feasibility and have shown that intent-aware approach outperforms egocentric conflict resolutions, as well as leads to reduction of close-quarter situations as it is possible to foresee collision risk in early stages of passage exploitation. Proposed multi-objective optimization-based collision avoidance method allows conflict resolutions with higher CPAs, reduces distance travelled to avoid collision and shortens time required to go back to the original route. To further reduce computational complexity of the collision avoidance algorithm, benefit of having decentralized unit for hazard alerting is investigated. This research showed that nuisance alerts onboard commercial sea surface vehicles are a substantial problem that has to be confronted by exploiting design of trajectory generator and hazard alerting algorithm that managed to considerably reduce nuisance alerts and ensure that only relevant alerts are triggered in collision-avoidance situations.

Izvorni jezik
Engleski

Znanstvena područja
Računarstvo, Tehnologija prometa i transport

POVEZANOST RADA