Cai
Levi
Cai
Levi
No Thumbnail Available
Search Results
Now showing
1 - 2 of 2
-
ArticleSemi-supervised visual tracking of marine animals using autonomous underwater vehicles(Springer, 2023-03-01) Cai, Levi ; McGuire, Nathan E. ; Hanlon, Roger ; T Aran Mooney ; Girdhar, YogeshIn-situ visual observations of marine organisms is crucial to developing behavioural understandings and their relations to their surrounding ecosystem. Typically, these observations are collected via divers, tags, and remotely-operated or human-piloted vehicles. Recently, however, autonomous underwater vehicles equipped with cameras and embedded computers with GPU capabilities are being developed for a variety of applications, and in particular, can be used to supplement these existing data collection mechanisms where human operation or tags are more difficult. Existing approaches have focused on using fully-supervised tracking methods, but labelled data for many underwater species are severely lacking. Semi-supervised trackers may offer alternative tracking solutions because they require less data than fully-supervised counterparts. However, because there are not existing realistic underwater tracking datasets, the performance of semi-supervised tracking algorithms in the marine domain is not well understood. To better evaluate their performance and utility, in this paper we provide (1) a novel dataset specific to marine animals located at http://warp.whoi.edu/vmat/, (2) an evaluation of state-of-the-art semi-supervised algorithms in the context of underwater animal tracking, and (3) an evaluation of real-world performance through demonstrations using a semi-supervised algorithm on-board an autonomous underwater vehicle to track marine animals in the wild.
-
ThesisTracking and monitoring marine animals with vision-guided underwater vehicles(Massachusetts Institute of Technology and Woods Hole Oceanographic Institution, 2026-02) Cai, Levi ; Girdhar, YogeshMarine animals constitute an estimated 80% of all animal biomass on Earth, with nearly one- to two-thirds of marine species still undescribed to science and much of the ocean at depth remains underexplored. Even among the species we know, for many, we have few observations and limited understanding of their natural behaviors. The scientific community needs strategies for broader and more scalable coverage at all ocean depths and benthic environments to map, monitor, and study these organisms in-situ. Currently, scientists rely on tags, divers, human-operated and remotely-operated vehicles to conduct in-situ studies of marine animals. Tags provide high-resolution and long-term data about particular animals, but each tag is designed for a specific species at a time and cannot be deployed ad-hoc. For new species discovery and studies, scientists use divers for shallow depths and human-operated and remotely-operated vehicles at greater depths, because they can be repurposed and make in-situ decisions to track and sample novel organisms or environments. However, these approaches are costly and operationally challenging, making them difficult to scale for sufficient coverage of the overall ocean. This thesis proposes the use of vision-guided autonomous underwater vehicles (AUVs) as a scalable and opportunistic approach to tracking and monitoring marine animals in-situ, both to study populations as well as individual animal behaviors. The goal is to enable AUVs to track and monitor any visible marine animal(s), including those that are rare or elusive, across any habitat or environmental conditions, while minimizing invasiveness or observation bias of animal behavior caused by the platform itself. In particular, it addresses fundamental challenges in the perception, controls, and design of AUVs with a focus on strategies that minimize reliance on a-priori knowledge of marine animals and their habitats, as well as measuring disturbance of the animals. First, this thesis explores how to leverage generalized data and machine learning approaches, such as semi-supervised learning, to enable visual tracking and detection of any marine animal. Next, using highly parallelized simulations combined with reinforcement learning methods to train agile and robust controllers while minimizing tuning and re-ballasting in the field. This encourages AUVs to be more opportunistically deployed with new payloads and enables them to follow fast-moving organisms or operate within complex seafloor environments. Finally, we investigate whether AUVs themselves impact animal behavior, potentially biasing estimates that are important to scientists. We propose a novel model of relating animal behavior to AUV presence, and also explore strategies to measure and mitigate these biases through changes to both AUV behavior and physical design while maintaining their agility or maneuverability.