A fast and dynamic tracking-based Micropterus salmoides fry counting method in highly occluded scenarios

Rapid and precise fry counting is crucial in aquaculture engineering, impacting breeding quality and marketing efficiency. Previous methods based on static image analysis are constrained by occlusion and limited generalization. Dynamic counting methods based on multi-object tracking (MOT) hold promise in addressing these challenges, but they still have counting accuracy and speed limitations due to algorithmic constraints. To address these limitations, we propose a dynamic Micropterus salmoides (largemouth bass) fry tracking and counting method tailored for densely occluded environments. Specifically, we employed the anchor-free You Only Look Once version 8 (YOLOv8) architecture, replacing its backbone with the lightweight and truncating the detection head at the highest level of YOLOv8 to enhance detection accuracy and efficiency. Furthermore, we adopted the Tracking-By-Detection architecture, optimized Kalman filtering's parameter updating for motion prediction, and enhanced the data association algorithm. This approach facilitates rapid and stable fry trajectory prediction and tracking while dramatically reducing identification switches, improving counting accuracy. Additionally, we constructed a dedicated dataset for fry tracking and counting, classifying videos by the density of fry in a single frame and the total number of fry across all frames. Experiment results have demonstrated that our method achieves up to 97 % counting accuracy in densely occluded scenarios (up to 62 fry per frame) and 100 % counting accuracy in scenarios with minimal occlusion (up to 19 fry per frame). Moreover, it can run at 19 frames per second on edge devices, thereby meeting the requisite speed and accuracy criteria.