Tracking animal identities with machine learning to analyze cichlid aggression

Abstract

Aggression is a central characteristic of social animal behavior that can drive early life development. We can use the social cichlid fish Astatotilapia burtoni as a model to study aggression by analyzing stereotyped chase behaviors defined by simple metrics such as changes in velocity. This enables a computer based data collection process, streamlined through the use of the SLEAP machine learning program. This method seeks to translate large quantities of cichlid video into positional coordinate data of multiple points on the animal’s skeletal limbs. Refining how SLEAP tracks animal identity through video improves the quality of positional data by labeling positional tracks with unique animal identities. SLEAP applies three primary parameters to match identities across frames: tracker methods create an initial positional estimate based on movement prediction, similarity methods connect two instances of identified fish from one frame to the next, and matching methods score each identity connection and optimize those scores. A variety of test cases examined the three parameters, Kalman filter application, and a diversified training data set to reduce the rate of identity swap based errors. Performance of these test cases were measured by the rate of identity errors relative to the other test cases, and tested across a variety of video examples differing by inclusion in the training data set and known machine learning model performance. The analysis of these relative error rates demonstrated a reduction of up to 79% of identity swap occurrences, which illustrates a significant improvement in the ability of the machine learning model to maintain separate animal identities.