Dolphin Detection and Tracking

Detecting and tracking social marine mammals, including bottlenose dolphins, can help to explain their social dynamics, predict their behavior, and measure the impact of human interference. Multi-camera setups provide an opportunity to record the behavior of captive dolphins and create a massive dataset from which long term statistics can be extracted, but the use of human experts to track dolphin positions over time demands a high time and financial investment cost. In this paper, we examine automated methods to analyze large video corpora. We use background subtraction to detect dolphins over time in the video data and to visualize the paths by which dolphins regularly traverse their environment. We demonstrate the use of these background subtraction methods as a way to initialize a preexisting real-time compressive tracking algorithm, which previously required human interaction. Detecting when dolphins transfer between pools both supplements 3D tracking efforts, helping identify moments when dolphins move between cameras, and augments human performance, providing a way for researchers to manually annotate dolphin pool positions 14 times as fast.