Modeling Movement of Fish

Abstract

Multi-animal pose detection algorithms such as Social LEAP Estimates Animal Poses (SLEAP) uses a generalized pose prediction algorithm to differentiate among individual animals in the same frame. However, such a method is resource intensive and wasteful when used with animals with simpler movement patterns compared to humans, such as dolphins. My thesis is that the movement of fish such as dolphins can be predicted more efficiently when developing non-generalized object tracking methods. This project uses Kalman Filters of size proportional to the size of the "computational skeleton" of a dolphin and is able to accurately classify between different individuals of fish while reducing the size of Kalman Filters by multiple hundred times compared to the method used in SLEAP.

A key feature of my work is the development of non-generalized techniques for animal pose prediction that yield a significant improvement in the speed of pose prediction for fish while preserving the high accuracy of large Kalman Filters. The implications of this result are important as reducing the computational resource load for algorithms such as SLEAP makes them more accessible to a wider set of devices that must operate with fewer resources (e.g., processing cores, memory, energy).