FishMonitorAI: Methodology and Annotated Video Dataset for Fish Detection and Multi-Object Tracking in Recirculating Aquaculture System

The transition to precision aquaculture systems requires the introduction of automated methods for continuous monitoring of the physiological state and behavior of aquatic organisms. The practical implementation of video analytics systems in recirculating aquaculture systems (RAS) is constrained by the non-representativeness of existing image datasets, the shortage of annotated video sequences for individual fish instance tracking tasks, and the lack of systematized engineering recommendations for organizing video recording in an artificial aquatic environment with high concentrations of suspended solids, variable lighting, and hydrodynamic interference. This study proposes methodologies for video recording and creating an annotated video dataset required for training a video analytics system for RAS. An analytical review of modern approaches to recognizing feeding behavior, fish detection and tracking, and morphometric analysis of fish was conducted. The identified limitations of existing open datasets relate to the predominance of static RGB images, the insufficient number of annotated video sequences, and limited species representation. Requirements are formulated for video recording equipment and its installation, as well as for lighting systems, auxiliary sensors, and recording schedules, with the aim of reducing the impact of common image artifacts such as low contrast, blur, glare, uneven illumination, and color distortion. To address the tasks of detecting and tracking a group of objects, a methodology for frame-by-frame annotation of video sequences using a modified TrackMe tool was developed, allowing the consistent formation of trajectories of individual fish while preserving their identity under conditions of occlusion and high stocking density in RAS. The developed FishMonitorAI Dataset corpus includes 8017 frames obtained from 12 video recordings of Lena sturgeon (Acipenser baerii) and grouped into four subsets corresponding to behavioral states (rest, feeding) and external stimuli (light stimulus, strike on the tank wall). In the current version, 900 frames of the calm-state subset (three 30-s segments) have been manually annotated with bounding boxes and individual track identifiers (16,543 boxes) following an explicitly defined visibility criterion. Baseline experiments with leave-one-sequence-out cross-validation showed that four detectors (YOLOv8n, YOLO11n, YOLO11s, RT-DETR-l) reach [email protected] of 91.6–92.2% but only 44.6–47.3% [email protected]:0.95, while the best tracking configuration achieves HOTA = 42.6% and IDF1 = 47.0%; even with ground-truth boxes, IDF1 does not exceed 73.4%, indicating that preserving the identity of visually similar sturgeon is the main challenge of the dataset. The engineering recommendations, annotation methodology, and annotated corpus with baselines are intended to facilitate the development of fish detection and tracking models for recirculating aquaculture systems.

Authors

Institutions

Publication Details

Journal
Fishes
Published
2026-11-18
DOI
https://doi.org/10.3390/fishes11110673
Primary Topic
Water Quality Monitoring Technologies
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

FishMonitorAI: Methodology and Annotated Video Dataset for Fish Detection and Multi-Object Tracking in Recirculating Aquaculture System

Mikhail Uzdiaev, Andrey Leonidovich Ronzhin, V. V. Volodina, Александра Фигурек et al.
5 links
Fishes
Water Quality Monitoring Technologies
article

FishMonitorAI: Methodology and Annotated Video Dataset for Fish Detection and Multi-Object Tracking in Recirculating Aquaculture System

Mikhail Uzdiaev, Andrey Leonidovich Ronzhin, V. V. Volodina, Александра Фигурек, Ekaterina Cherskikh, Marina Alekseevna Astapova, Le Van Nghia
article en

Abstract

The transition to precision aquaculture systems requires the introduction of automated methods for continuous monitoring of the physiological state and behavior of aquatic organisms. The practical implementation of video analytics systems in recirculating aquaculture systems (RAS) is constrained by the non-representativeness of existing image datasets, the shortage of annotated video sequences for individual fish instance tracking tasks, and the lack of systematized engineering recommendations for organizing video recording in an artificial aquatic environment with high concentrations of suspended solids, variable lighting, and hydrodynamic interference. This study proposes methodologies for video recording and creating an annotated video dataset required for training a video analytics system for RAS. An analytical review of modern approaches to recognizing feeding behavior, fish detection and tracking, and morphometric analysis of fish was conducted. The identified limitations of existing open datasets relate to the predominance of static RGB images, the insufficient number of annotated video sequences, and limited species representation. Requirements are formulated for video recording equipment and its installation, as well as for lighting systems, auxiliary sensors, and recording schedules, with the aim of reducing the impact of common image artifacts such as low contrast, blur, glare, uneven illumination, and color distortion. To address the tasks of detecting and tracking a group of objects, a methodology for frame-by-frame annotation of video sequences using a modified TrackMe tool was developed, allowing the consistent formation of trajectories of individual fish while preserving their identity under conditions of occlusion and high stocking density in RAS. The developed FishMonitorAI Dataset corpus includes 8017 frames obtained from 12 video recordings of Lena sturgeon (Acipenser baerii) and grouped into four subsets corresponding to behavioral states (rest, feeding) and external stimuli (light stimulus, strike on the tank wall). In the current version, 900 frames of the calm-state subset (three 30-s segments) have been manually annotated with bounding boxes and individual track identifiers (16,543 boxes) following an explicitly defined visibility criterion. Baseline experiments with leave-one-sequence-out cross-validation showed that four detectors (YOLOv8n, YOLO11n, YOLO11s, RT-DETR-l) reach [email protected] of 91.6–92.2% but only 44.6–47.3% [email protected]:0.95, while the best tracking configuration achieves HOTA = 42.6% and IDF1 = 47.0%; even with ground-truth boxes, IDF1 does not exceed 73.4%, indicating that preserving the identity of visually similar sturgeon is the main challenge of the dataset. The engineering recommendations, annotation methodology, and annotated corpus with baselines are intended to facilitate the development of fish detection and tracking models for recirculating aquaculture systems.

FishesVol. 11(11)
Le Quy Don Technical University (VN), University of Nicosia (CY), St. Petersburg Institute for Informatics and Automation (RU), State Marine Technical University of St. Petersburg (RU), Kaliningrad State Technical University (RU)
Russian Science Foundation
Zero hunger
Water Quality Monitoring Technologies
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon