While various tracking algorithms have demonstrated effectiveness in terrestrial and aerial contexts, their performance in underwater settings remains unexplored. Object tracking in underwater videos presents unique challenges due to variable lighting, water turbidity, and unpredictable camera movement, all of which are likely to hinder the performance of traditional detection and tracking methods. Addressing this gap is crucial for applications such as marine biology research, underwater surveillance, and autonomous underwater vehicles.

This thesis first evaluates existing tracking algorithms, SORT, ByteTrack, and Bag-of-Tricks SORT (BoT-SORT), each incorporating motion estimation and linear data assignment methods on a novel underwater video dataset with a moving camera. The thesis then improves on the SORT algorithm by adopting velocity estimation techniques, a formula-based and an optical flow-based, giving rise to two new algorithms, SORT-V and SORT-OF. Furthermore, the thesis proposes a novel tracker that utilises an Interacting Multiple Model filter to estimate the location of the target object. The evaluation focuses on finding a balance between specific metrics, such as tracking accuracy, identity switches, and the number of tracked and lost trajectories. The results indicate that using velocity estimation techniques improves the tracking accuracy by 11% and tracks more objects by nearly halving the number of lost objects. Incorporating a constant acceleration model in the IMM filter gives the best result, with the highest tracking accuracy and with the least number of identity switches, all in real-time computational speed.

Marine pollution is a critical issue impacting the global community, with underwater waste a particularly daunting challenge. While autonomous detection and collection of underwater waste is highly desirable, these are extremely difficult tasks. This difficulty arises from the intrinsic complexities of the aquatic environment, including variable lighting conditions, reduced visibility, and the complex nature of water currents. This paper focuses on novel approaches for autonomous detection of underwater waste and proposes to incorporate domain knowledge to refine deep-learning-based underwater object detection techniques. More specifically, the domain knowledge is represented with models in the state space form that describe the motion of the target objects to assist the classification of objects. Moreover, optical flow is combined with a k-means clustering algorithm to extract the trajectory of the target objects from videos. These trajectories are subsequently fed into a neural network that is trained using knowledge distillation, enhanced with domain knowledge. For our experiments, a simulator is devised to facilitate the creation of a dataset for developing and testing the proposed architecture. The results of the experiments demonstrate that including domain knowledge within the object detection approach with neural networks provides numerous substantial advantages, including enhanced robustness against noisy and poorly labelled data, facilitation of semi-supervised learning, and consistent superiority in accuracy over the baseline scenario. Additionally, combining the domain knowledge with a neural network significantly increases the computational speed of object detection compared to using the standalone domain knowledge module.