A Comparative Analysis of the Yolo Models for Intelligent Lobster Surveillance Camera

In lobster farming, monitoring of lobsters is essential and must be performed. Aruna, one of the lobster trading startups in Indonesia, is still using manual monitoring by sending divers underwater to observe the progress of the lobsters. This way is considered less effective because it is done intermittently, only at certain times and in good weather conditions. In this study, we propose to investigate object detection for underwater surveillance cameras as a means of real-time monitoring and counting of lobsters. Underwater video data is fed into the proposed system based on a deep learning-based model. This system consists of three processes: image enhancement with WaterNet as noise reduction, object counting with the YOLOv6/7/8 model, and object tracking using StrongSORT. For details, we annotated the lobster data by head only and full body. Our test results show that YOLOv7, with head annotation, has stable mean Average Precision (mAP), making it suitable for the proposed system. Then, we implemented a combination of WaterNet, YOLOv7 and StrongSORT to the proposed system architecture for the Intelligent surveillance camera.