Fish Detection and Tracking in Pisciculture Environment using Deep Instance Segmentation

This study presents a novel approach in detecting and tracking of fish in pisciculture. Pisciculture in general involves challenging tasks of counting and monitoring fish in natural or nature like, man-made habitats such as inland fisheries for breeding, feeding and sorting purposes. These are presently achieved using conventional methods that are inefficient when implemented in large-scale commercial productions. To overcome such difficulties and improve the efficiency of the processes, images of fish and fish seeds are captured in natural murky water habitats through a vision sensor on board an unmanned aerial vehicle (UAV). In this research paper, a deep instance segmentation algorithm called Mask R-CNN along with GOTURN tracking algorithm is employed for real time fish detection and tracking. A comparison study is also carried out (i) fish detection on high resolution images (ii) fish detection on high resolution image multi-region parallel processing (iii) fish detection on high resolution image multi-region parallel processing with tracking. The results are found to be accurate with image multi-region parallel processing along with tracking, with an F1 score of 0.91 at 16 frames per seconds on in-land fishes environment.