BehaveAI enables rapid detection and classification of objects and behavior from motion

Area of Science:

• Computer Vision
• Bio-inspired AI
• Machine Learning

Background:

• Traditional video analysis often struggles with complex natural scenes and subtle motion patterns.
• Integrating static and motion information is crucial for robust object detection and behavior classification.
• Current methods can be computationally intensive and require large annotated datasets.

Purpose of the Study:

• Introduce BehaveAI, a novel video analysis framework.
• Enhance object detection and behavior classification using biologically inspired techniques.
• Improve the efficiency and accessibility of AI model development for video analysis.

Main Methods:

• Developed a color-from-motion encoding strategy to translate movement into color gradients.
• Integrated static and motion information for enhanced feature representation.
• Utilized a deep learning architecture (YOLO11) with a semi-supervised annotation workflow.
• Implemented flexible hierarchical model structures for separate detection and classification tasks.

Main Results:

• Demonstrated robust detection of challenging objects in complex natural scenes.
• Achieved reliable classification of behaviors in animals and single-celled organisms.
• Showcased reduced dataset annotation effort, enabling model creation within an hour.
• Confirmed real-time performance on low-end edge devices like Raspberry Pi.

Conclusions:

• BehaveAI offers a computationally lightweight and efficient solution for video analysis.
• The framework significantly lowers the barrier to entry for AI model development in this field.
• Open-source release promotes accessibility and further research in bio-inspired AI for behavior analysis.