A Computationally Efficient Neuronal Model for Collision Detection with Contrast Polarity-Specific Feed-Forward...

Area of Science:

• Neuroscience and Robotics
• Computational Vision
• Bio-inspired Engineering

Background:

• Animals use dynamic vision for collision avoidance, inspiring artificial systems.
• Lobula giant movement detectors (LGMDs) are key to bio-inspired collision sensors.
• Existing models lack efficient separation of ON/OFF contrast signals in feed-forward inhibition (FFI).

Purpose of the Study:

• To develop an optimized computational LGMD model with polarity-specific FFI.
• To enhance energy efficiency and processing speed in artificial vision systems.
• To improve collision detection selectivity and robustness.

Main Methods:

• Proposed a neuronal model with FFI divided into distinct ON/OFF channels.
• Introduced an energy-efficient activation function for neural computation.
• Conducted systematic comparative testing with state-of-the-art models and real-world robot implementation.

Main Results:

• Achieved significant in silico time savings while preserving collision selectivity.
• Demonstrated robust detection of ON/OFF-contrast looming stimuli and suppression of translational motion.
• Implemented model on a micro-mobile robot, reaching 97.51% collision avoidance success with halved processing time.

Conclusions:

• The optimized LGMD model offers a robust and energy-efficient collision-sensing solution.
• Separating FFI into ON/OFF channels significantly accelerates visual processing.
• This approach effectively addresses real-world collision avoidance challenges in robotics.