There is a need for efficient fish detection and a tracking device to count, track and determine abnormal fish behaviour. This would potentially help fish farms for a non-invasive, less stressful, continuous monitoring of fish during quarantine and growth, for diagnostic behaviour and early intervention of disease with appropriate treatment
The team has developed a video analytics software that uses YOLOv2 (you only look once) model and the Darkflow neural network Image or live fed video is used as a media input to track the fish. The count results were stored in JavaScript Object Notation (JSON) file format and the live detection were stored in MP4 file format. The format is chosen for ease for future data manipulation. The neural network algorithm developed has successfully differentiated multiple species in tanks and able to track fish movements to identify sick fish. Improvements to be made includes expanding the training data set to increase accuracy.
The following picture depicts the overall process (A) and the coding that interacts with the neural network (B).