Decoding the Dance: How Video Multitracking Reveals the Hidden Lives of Fish

The secret world of fish behavior is being unlocked, one pixel at a time.

Imagine trying to follow dozens of identical, rapidly moving fish as they weave through water, constantly crossing paths and changing direction. For scientists studying fish behavior, this isn't just a challenge—it's an everyday obstacle. Video multitracking technology has revolutionized this field, transforming hours of video footage into precise, quantitative data about how fish interact, respond to their environment, and make group decisions. This article explores how computer vision and artificial intelligence are helping researchers decode the complex dynamics of fish behavior, with profound implications for marine conservation, aquaculture, and our understanding of animal collectives.

From Manual Counts to Machine Learning: The Evolution of Fish Tracking

Simple Observation Era

The journey to modern fish tracking began with simple observation and manual notation. Early automated methods relied on basic computer vision techniques like background subtraction and frame differencing to detect movement. While these approaches worked in controlled laboratory settings, they struggled with real-world complexities like changing lighting, occlusions, and similar-looking individuals2 4 .

Deep Learning Revolution

The true transformation began with the adoption of deep learning and sophisticated detection algorithms. Convolutional Neural Networks (CNNs) enabled systems to learn the visual characteristics of fish directly from data, dramatically improving detection accuracy even in challenging conditions4 . Today's state-of-the-art systems can simultaneously track dozens of individual fish across thousands of frames, maintaining their identities through complex interactions.

Early Methods

Manual observation and basic computer vision techniques with limited accuracy in complex environments.

Modern AI Approaches

Deep learning models like CNNs that learn directly from data, enabling high-accuracy tracking even with occlusions and similar appearances.

Why Tracking Fish is Uniquely Challenging

Fish present distinctive challenges that set them apart from terrestrial tracking scenarios:

Non-rigid bodies

Fish constantly deform their bodies while swimming, unlike the relatively predictable movements of humans or vehicles1 7

Similar appearances

Individuals of the same species often look nearly identical to automated systems1

Erratic movement patterns

Fish can change direction and speed instantaneously, defying simple motion prediction models1

Frequent occlusions

In dense shoals, fish constantly cross paths and temporarily block each other from view3 7

Complex underwater environments

Factors like water turbidity, light refraction, and floating particles complicate detection1

Key Insight

The combination of these challenges makes fish tracking one of the most difficult computer vision problems, requiring specialized algorithms that can handle non-linear motion, visual similarity, and complex environmental factors.

Inside a Groundbreaking Experiment: Tracking Zebrafish Shoaling Behavior

To understand how modern multitracking works in practice, let's examine a comprehensive 2025 study that developed a sophisticated framework for analyzing zebrafish shoaling behavior2 .

The Experimental Setup

Researchers constructed a specialized behavioral monitoring system with high-resolution cameras (1920 × 1080 at 30 fps) and controlled lighting to ensure consistent image quality. They used wild-type male red zebrafish, divided into groups of five—a group size known to reliably induce shoaling behavior. To test how environmental changes affect behavior, the team exposed different groups to varying ethanol concentrations (0%, 0.50%, and 1% v/v), tracking their behavioral responses2 .

Experimental Parameters
  • Camera Resolution: 1920 × 1080 at 30 fps
  • Group Size: 5 zebrafish per trial
  • Ethanol Concentrations: 0%, 0.50%, 1% v/v
  • Species: Wild-type male red zebrafish

The Tracking Framework: A Two-Stage Approach

The methodology featured a cascaded detection-tracking pipeline:

1. Enhanced Detection with ZebraYOLO

The team started with YOLOv8s, a state-of-the-art object detection model, and made key improvements specifically for fish tracking:

  • Extended feature pyramid hierarchy: Added an additional P2 detection layer (160 × 160 pixels) to better capture small fish, increasing the receptive field to 9 pixels2
  • Global Attention Mechanism (GAM): Enhanced the network's ability to focus on relevant features while ignoring noise, particularly useful for dealing with motion blur2
  • Optimized loss function: Reduced computational overhead and accelerated training2
2. Advanced Tracking Components

The detection results fed into a sophisticated tracking system:

  • Interactive Multiple Model Kalman Filter (IMM-KF): Combined constant velocity, constant acceleration, and turning motion models to adaptively handle erratic zebrafish movement2
  • Posture-aware appearance feature network: Retrained the Re-Identification (Re-ID) network on zebrafish data to better distinguish between individuals2

Performance Comparison of Tracking Methods

Method HOTA IDF1 Key Advantages
SU-T (MFT25) 34.1 44.6 Optimized for non-linear fish motion1
ZebraYOLO + IMM-KF Not specified Not specified Adapts to erratic movement patterns2
Traditional Motion Tracking Lower Lower Simple implementation

Results: Revealing the Biphasic Effects of Ethanol

The experiment yielded fascinating insights into how chemical stressors affect fish behavior. The data revealed a biphasic response to ethanol exposure:

Low Concentrations

0.50% v/v

Increased global motion intensity, creating hyperactivity in the shoal2

Control Group

0%

Baseline behavior with normal shoal cohesion and locomotor activity2

Higher Concentrations

1% v/v

Reduced locomotor activity and disrupted shoal cohesion2

Zebrafish Behavioral Response to Ethanol Exposure
Ethanol Concentration Shoal Cohesion Locomotor Activity Overall Effect
0% (Control) Normal Normal Baseline behavior
0.50% v/v Slight disruption Increased Hyperactivity
1% v/v Significant disruption Decreased Lethargy, disorganization
Research Insight

These findings demonstrate how multitracking can detect subtle behavioral changes that might be invisible to the human eye, providing early warning signs of environmental stress.

The Researcher's Toolkit: Essential Solutions for Fish Behavior Studies

Modern fish behavior labs rely on a sophisticated array of technical solutions:

Detection Models

YOLOv8, ZebraYOLO, GAB-YOLO2 5

Identify fish in video
Tracking Algorithms

Unscented Kalman Filter, Interactive Multiple Model KF1 2

Predict motion
Feature Extraction

FishIoU, Posture-aware Re-ID networks1 2

Distinguish individuals
Data Management

MFT25 dataset, BrackishMOT1

Standardized benchmarks

Beyond the Lab: Real-World Applications

The implications of advanced fish tracking extend far beyond basic research:

Conservation and Ecology

By tracking wild fish populations, researchers can monitor species responses to environmental changes, pollution, and climate change. For instance, one study tracked pulse-type weakly electric fish in their natural habitat using arrays of electrodes, revealing their nocturnal movement patterns and territorial behaviors6 .

Sustainable Aquaculture

In fish farming, tracking enables early detection of health issues and stress responses. Systems like GAB-YOLO can identify abnormal behaviors in valuable species like Greater Amberjack, allowing farmers to intervene before problems spread5 . Another study demonstrated how tracking can detect ammonia nitrogen stress in fish, providing early warning of water quality issues9 .

Scientific Discovery

Video multitracking has enabled fundamental discoveries about collective behavior, revealing how fish make group decisions, transfer information, and coordinate movements without centralized leadership.

The Future of Fish Tracking

As the field advances, researchers are working to overcome remaining challenges, including long-term identity preservation, 3D tracking in complex environments, and applications in open-water settings. The development of comprehensive benchmark datasets like MFT25—containing 408,578 annotated bounding boxes across 48,066 frames—is accelerating progress by providing standardized testing grounds for new algorithms1 .

Emerging Technologies

The integration of increasingly sophisticated AI, including transformer architectures and enhanced appearance models, promises to further improve tracking accuracy while reducing computational demands. As these technologies mature, we can expect even deeper insights into the fascinating underwater world of fish behavior.

Future Research Directions

  • Long-term identity preservation In Progress
  • 3D tracking in complex environments Experimental
  • Open-water applications Challenging
  • Transformer architectures Emerging

Conclusion

From delicate zebrafish in laboratory tanks to great amberjack in commercial aquaculture, video multitracking is transforming how we understand and interact with aquatic life. By converting the elegant dance of fish shoals into precise data, this technology bridges the gap between intuition and evidence, revealing patterns and connections that have remained hidden for millennia.

Article Highlights
  • Evolution from manual to AI tracking
  • Unique challenges in fish tracking
  • Groundbreaking zebrafish experiment
  • Essential research tools
  • Real-world applications
Key Statistics
MFT25 Dataset Size 408,578 boxes
Frames Annotated 48,066
Camera Resolution 1080p
SU-T HOTA Score 34.1
Share This Article

References