A computationally efficient and robust looming perception model based on dynamic neural field.

Area of Science:

• Computational Neuroscience
• Robotics
• Computer Vision

Background:

• Existing bio-inspired looming perception systems face challenges in selectivity and encoding movement attributes.
• Hierarchical neural networks can be fragile in distinguishing approaching movements.
• Nonlinear mapping models struggle to encode physical attributes like angular velocity.

Purpose of the Study:

• To propose a novel, computationally efficient looming perception model inspired by dynamic neural fields (DNF).
• To enhance looming selectivity and robustness compared to existing bio-inspired systems.
• To validate the model's effectiveness in real-world applications like robotic collision detection.

Main Methods:

• Developed a single-field Dynamic Neural Field (DNF) model with adaptive lateral interactions and a dynamic activation threshold.
• Implemented lateral excitation and inhibition for motion discrimination.
• Utilized mathematical analysis and ablation studies to support model effectiveness.

Main Results:

• The DNF model demonstrated superior looming selectivity, responding strongly to approaching objects.
• Achieved a 93% success rate in on-line collision detection tasks for micro-mobile robots.
• Showcased computational efficiency and robustness compared to state-of-the-art model-based methods.

Conclusions:

• The proposed DNF model offers a robust and efficient solution for looming perception.
• This brain-inspired framework effectively addresses limitations of previous bio-inspired systems.
• The model shows significant promise for applications in robotics and autonomous systems.