Behaviour Prediction of Shrimp using trajectory analysis and their validation with deployed IoT Sensor

Abstract

This study presents a comprehensive end-to-end pipeline for real-time monitoring and behaviour prediction of shrimp locomotion in variable environmental conditions, integrating state-of-the-art deep learning, computer vision, and signal processing methodologies. The framework combines an enhanced YOLOv8 detection system with Deep SORT tracking algorithms and implements a sophisticated hybrid trajectory denoising approach utilising Savitzky-Golay filtering, cubic spline interpolation, and Gaussian smoothing. Applied to a custom-annotated dataset of 789 underwater shrimp images, the detection model achieved 0.74 precision, 0.856 recall, and 0.794 mAP@0.5. Advanced trajectory analysis techniques enabled 3D visualisation and directional behaviour quantification under four distinct environmental regimes combining pH variations (5.4.6.8) and temperature conditions (33°C, 35°C). A comprehensive comparative analysis with recent advances in underwater object detection, particularly YOLOv8-CPG architectures incorporating Compact Inverted Blocks (CIB), Partial Self-Attention (PSA), and Gold-YOLO feature fusion mechanisms, demonstrates potential performance improvements of 1-3% mAP through architectural optimisation. The methodology's integration with precision aquaculture monitoring systems, IoT sensor networks, and real-time behavioural alerting mechanisms positions it as a critical tool for sustainable aquaculture management, environmental stress detection, and automated welfare assessment in intensive farming operations.