Fully Memristive Elementary Motion Detectors for a Maneuver Prediction

Area of Science:

• Neuroscience
• Materials Science
• Computer Science

Background:

• Insects utilize elementary motion detectors (EMDs) for efficient motion detection.
• Conventional machine vision systems demand substantial computational power for dynamic motion processing.

Purpose of the Study:

• To develop a fully memristive elementary motion detector (M-EMD) mimicking biological EMDs.
• To implement the Hassenstein-Reichardt (HR) correlator using memristive components.
• To evaluate the M-EMD's performance in a neuromorphic system for vehicle behavior prediction.

Main Methods:

• Designed a memristive EMD (M-EMD) using a Wye (Y) configuration with a static resistor, dynamic memristor, and Mott memristor.
• Integrated M-EMDs into a neuromorphic system for lane-changing maneuver prediction.
• Utilized the Next Generation Simulation (NGSIM) dataset for system validation.

Main Results:

• The M-EMD successfully achieved direction-selective responses through spatio-temporal integration.
• The neuromorphic system demonstrated high accuracy (> 87%) in predicting vehicle lane-changing maneuvers.
• Computational cost was reduced by 92.9% compared to conventional systems.

Conclusions:

• The developed M-EMD offers an efficient, bio-inspired approach to motion detection.
• The M-EMD-based neuromorphic system shows significant potential for edge-level computing applications, particularly in autonomous driving.
• This research bridges neuroscience and materials science for advanced AI systems.