A directionally selective collision-sensing visual neural network based on fractional-order differential operator

Area of Science:

• Computational Neuroscience
• Computer Vision
• Biologically Inspired AI

Background:

• Lobular giant motion detectors (LGMDs) are crucial for collision avoidance in biological systems.
• Existing LGMD models primarily focus on collision detection, lacking directional information.
• There is a need for advanced visual systems that mimic biological motion perception.

Purpose of the Study:

• To propose a directionally selective fractional-order LGMD visual neural network.
• To enable the model to detect collision threats and determine the motion direction of objects.
• To enhance the understanding of biological motion detection mechanisms.

Main Methods:

• Simulating neuronal membrane potential responses using fractional-order differential operators to generate collision spikes.
• Developing a novel correlation mechanism utilizing temporal signal delays between pixels to determine object motion direction.
• Incorporating ON/OFF visual channels to model bipolar neuronal responses to brightness changes.

Main Results:

• The proposed model successfully senses collision threats and identifies the motion direction of objects.
• The system's response characteristics align with biological LGMDs and direction-selective neurons.
• Experimental validation confirms the model's stable and reliable performance.

Conclusions:

• The developed fractional-order LGMD visual neural network offers enhanced capabilities for motion detection and directionality.
• This biologically inspired approach provides a robust framework for advanced visual processing systems.
• The study contributes to the field of artificial intelligence by mimicking complex neural functions for practical applications.