Nile Tilapia Size Estimation and Tracking Using YOLOv5 and DeepSORT

Abstract

The growing industry of Nile tilapia aquaculture evolved along with the current developments in technology. The manual grading or measurement of fish remained an inconvenience in terms of time consumption and labor. This paper compared the performances of three of the known state-of-art one-stage object detectors; You Only Look Once (YOLO) v5, RetinaNet, and EfficientDet, by training them on Nile tilapia dataset. The fish detection results show that YOLO with 88.1% mean average precision (mAP) and 83% F1-score at 80 epochs, outperforms the other two algorithms. The YOLO algorithm was then deployed and used as a detector of Nile tilapia and integrated with DeepSORT for real-time fish identification and tracking using a single web camera in an experimental controlled environment. The resulting system measurement produced accuracies of 70.06% and 52.10% for length and height measurements, respectively. Unfortunately, DeepSORT shows inconsistent and frequent ID switching due to occlusion. Even so, the fish detection was done successfully and can be instrumental in improving aquaculture for Nile Tilapia monitoring.