Automated Vitality Evaluation of Shrimp Postlarvae via Machine Vision: A Multi-Object Tracking and Behavioral Analytics Approach

Abstract

Activity of shrimp postlarvae is a critical proxy for quality and survival in aquaculture, yet current assessments rely on manual observation of swimming behavior, which is subjective and difficult to standardize. We introduce an automated, objective, and scalable video-analytics pipeline for evaluating activity in Litopenaeus vannamei postlarvae. The system first applies YOLOv8-Pose to extract keypoint-based positional data and body lengths from individual postlarvae across video frames. These detections are linked over time using the BoTSORT multi-object tracker to derive individual-level motion trajectories. From these trajectories, we construct an evaluation framework grounded in four population-level metrics—trajectory irregularity, uniformity, surface skimming ratio, and polarization—that jointly capture swarm dynamics and vertical distribution. We quantify the relative importance of these metrics via a hybrid weighting scheme combining the Analytic Hierarchy Process (AHP) with cluster-based collective decision-making, and fuse them into a comprehensive activity score through weighted aggregation of the machine-vision outputs. Finally, we stratify the resulting activity scores into three operational levels—high, moderate, and low—using K-means clustering. This approach replaces subjective inspection with a reproducible, data-driven assessment, enabling standardized monitoring and decision support in hatchery practice.