Event-Based Computation of Motion Flow on a Neuromorphic Analog Neural Platform

Area of Science:

• Neuroscience
• Robotics
• Computer Vision

Background:

• Estimating object motion is vital for agents in dynamic environments.
• Biological systems use retinal ganglion cells for motion detection.
• Artificial systems struggle to match biological efficiency in optical flow extraction.

Purpose of the Study:

• To develop a robust optical flow extraction architecture using an analog neuromorphic multi-chip system.
• To mimic the direction-selective ganglion cells of the rabbit retina, as proposed by Barlow and Levick.
• To demonstrate reliable optical flow extraction in artificial systems inspired by biological principles.

Main Methods:

• Designed a feed-forward network of analog integrate-and-fire neurons.
• Utilized contrast-sensitive photoreceptors as input.
• Employed precise spike timing for computation and time-lag analysis between neurons.
• Modeled the architecture after Barlow and Levick's 1965 findings on rabbit retinal ganglion cells.

Main Results:

• The neuromorphic detectors successfully encoded apparent visual motion's amplitude and direction in their spiking patterns.
• Demonstrated reliable optical flow extraction from a simple visual scene using mismatched analog components.
• Analyzed the system's latency, scalability, and robustness.

Conclusions:

• Precise spike timing, biological inspiration, and neuromorphic engineering can be integrated to solve complex visual tasks.
• The proposed analog neuromorphic system offers a robust and efficient method for optical flow extraction.
• This approach advances artificial systems' ability to perceive and react to motion in real-time.