• Infrared depth imaging enables non-intrusive analysis of lobster behavior. • Vision algorithm extracts real-time behavioral features from depth imaging. • Feature sets from single lobster trials were tested for aggressiveness prediction. • The framework offers a feasible basis for high-throughput behavioral phenotyping. The European lobster ( Homarus gammarus ) is a highly valued seafood, but its aggressive behavior impedes large-scale aquaculture production in efficient communal rearing systems. This paper introduces a novel computer vision system designed for high-resolution analysis of lobster behavior and a machine learning framework for non-injurious classification of aggressiveness that could serve as components of a prospective high-throughput behavioral phenotyping system for selective breeding. An infrared depth camera is employed to mitigate the influence of substrate color or material in the behavioral screening arena, and to minimize the disturbance caused by visible light, which is critical given the nocturnal nature of the species. The contributions of this paper are twofold. First, we present a computer vision algorithm optimized for infrared depth imaging that enables real-time extraction of quantitative behavioral measurements from single or pairwise screening trials. The algorithm was evaluated using ten adult European lobsters of wild origin to investigate its performance and feasibility. Second, we propose an automated module for the assessment of lobster aggressiveness that utilizes single-animal trials to avoid injurious encounters and employs a support vector machine classifier. This concept study demonstrates technical feasibility by achieving a cross-validated error of approximately 20%, and highlights the potential for scalable and automated aggressiveness assessment using larger datasets.
Yan et al. (Wed,) studied this question.