Abstract
This thesis presents FishNet Monitor, a resource-efficient framework for detecting "shing net deployment and retrieval actions in long, continuous video streams from commercial fishing vessels. Maritime monitoring systems face significant challenges in edge deployment scenarios where computational resources are limited, yet accurate "shing action detection remains critical for regulatory compliance and sustainable resource management. We address these challenges through a systematic approach that combines feature engineering, algorithmic innovation, and sampling optimization tailored specifically for maritime environments. First, we demonstrate that Average Detections Per Frame (ADPF) provides a more effective feature than bounding box aspect ratios for distinguishing between consecutive "shing actions in challenging sea conditions. Second, we develop an adaptive peak detection algorithm that robustly identifies "shing events despite variable time intervals between them. Third, we implement and evaluate two sampling strategies—fixed and dynamic—across 152 experimental configurations to optimize the inference cost versus accuracy trade off for extended maritime operations. Our results demonstrate that a fixed sampling approach with frame_skip=4 achieves 100%detection accuracy while reducing inference costs to 25% of full-frame processing, enabling continuous monitoring throughout extended "shing expeditions. Further analysis reveals that while both sampling strategies deliver comparable inference cost savings, the fixed approach provides significant runtime advantages for edge deployment on "shing vessels. Performance projections indicate the framework can process maritime surveillance footage in real-time on the Hailo 8 edge computing platform, offering a practical solution for automated fishing action detection that supports compliance monitoring, sustainable "shing practices, and $eetoperation optimization.