Deep Learning Based Object Detection and Tracking Method for Flatworm Behavior Analysis

• Published in: OCEANS 2023 - MTS/IEEE U.S. Gulf Coast pp. 1 - 8
• Main Authors: Zhao, Yanguo, Chen, Leon, Wong, Tim, Xu, Song
• Format: Conference Proceeding
• Language: English
• Published: The Marine Technology Society (MTS) 25.09.2023
• Subjects: Behavioral sciences; Biological system modeling; Data refinement; Deep learning; Filtering algorithms; Flatworms; Kalman Filter; Nanoscale devices; Strong-Sort; Target tracking; Trajectory; YOLOV5
• DOI: 10.23919/OCEANS52994.2023.10337404

In marine ecosystems, flatworm species feed on economically important species such as oysters and could increase the mortality and thereby lower the yield. Understanding the behavior of flatworms in response to environmental factors and, more importantly, knowing the distribution patterns of the flatworms in the field is crucial and has important implications for fishery management. Currently, the dominant methods for flatworm observation are all Kalman Filter-based. However, the traditional Kalman Filter-based tracking algorithms were mostly based only on one single frame as the reference template for the tracking of the target while flatworms are highly deformable targets that can have a great degree of shape-shifting. Here, we proposed a workflow of flatworm tracking using deep learning module (YOLOV5 and StrongSort) to track flatworm movement with high accuracy. We conducted a series of food choice essay experiments and verified our workflow by tracking movement and reporting the trajectory of the subject flatworm species. We use data refinement algorithm based on the similarity of images to greatly reduce the dataset size to allow for faster training speed. We used the workflow developed in this study to track the flatworm specimen in different experiment videos and achieved a 100% successful tracking rate (STR) in the tracking of the target. In addition, the tracking method was deployed on edge computing device Nvidia Jetson Nano with the achievement of 7 frames per second, indicating the capability of our algorithm in real-time tracking. More importantly, our workflow successfully detected and then tracked the movement of target flatworm specimen in all experiment videos. Based on the tracking data, we computed the trajectory and the moving speed of the flatworm. It is expected that this algorithm could be further applied to field video data for the tracking of flatworm, particularly the tracking of S. eliptica, and thereby assisting the study of benthic ecology in oyster farms.