Academic Report: Research on a Biomimetic Visual Detection Model Based on Fractional Spiking Neuron Circuits

In the fields of intelligent autonomous driving and unmanned aerial vehicles, quickly and effectively predicting collisions and triggering avoidance behaviors have significant application value. The Lobula Giant Movement Detectors (LGMDs) in locusts can effectively predict collisions before they occur and trigger avoidance behaviors. This ability makes LGMDs an ideal model for designing collision avoidance artificial visual systems. Unlike traditional CMOS cameras, event cameras (DVS) can simulate photoreceptors in biological visual systems, emulating the differentiation at the LGMD system’s base layer. They offer high time resolution, high dynamic range, and minimal motion blur advantages.

Background and Significance

The authors of this study are Yabin Deng, Haojie Ruan, Shan He, Tao Yang, and Donghui Guo from Xiamen University and Fuzhou University. They published a research paper titled “A Biomimetic Visual Detection Model: Event-Driven LGMDs Implemented with Fractional Spiking Neuron Circuits” in the IEEE Transactions on Biomedical Engineering journal.

The motivation for this study is that current research on LGMDs falls into two major camps: one emphasizes the presynaptic visual pathway; the other emphasizes the characteristics of LGMDs neurons themselves. Current computational models often neglect the biophysical behavior description of individual LGMD neurons, severely limiting their biological interpretability and practical applicability. By introducing fractional spiking neuron (FSN) circuits, this study aims to construct a compatible biomimetic visual model based on both characteristics to explain and mimic LGMDs behaviors.

Research Methods and Process

Experimental Design and Implementation:

The study mainly comprises the following steps:

1. Event Camera Input Layer: Use event cameras instead of traditional CMOS cameras to preprocess images and generate event streams. These event streams reflect the luminance changes of each pixel, similar to pulse dynamics.

2. ON/OFF Visual Pathways: Define ON pathways and OFF pathways to process events transitioning from bright to dark and dark to bright, respectively. By simulating locust visual pathways, the model can respond to targets in complex scenes.

3. Fractional Spiking Neuron Circuits (FSN): Introduce FSN circuits to simulate the adaptive pulse frequency of LGMD neurons on multiple scales. FSN changes dendritic morphology parameters to mimic the multi-scale spike frequency adaptation (SFA) in LGMDs.

4. Implementation of Event-Driven Model: Combine presynaptic inhibition and postsynaptic inhibition to accomplish functions such as collision detection and approaching object selection in complex scenes.

Data and Algorithm Analysis:

1. Calculating ON/OFF Pathways: For each pixel at any spatial location, calculate the positive signals generated on the ON pathway and negative signals generated on the OFF pathway, representing excitatory and inhibitory signals, respectively.

2. LGMDs Neuron Circuit Model: By introducing multi-scale SFA and dendritic morphology parameters, verify the response characteristics of a single LGMD neuron under different electrical stimulation in experimental settings.

3. System Testing: Conduct experiments simulating complex scenes and real physical videos to test the model’s collision detection and selection capabilities in fast-approaching objects and noise-complex scenes.

Experimental Results

1. Simulation of Single LGMD Neuron Behavior:

   • Under different electrical pulse stimuli, FSN circuits can mimic burst responses and exhibit multi-scale adaptive features under various injected current conditions over different periods.
   • Experiments demonstrated that the pulse frequency adaptation and time constant ranges of FSN circuits align with biological LGMDs, validating their biological interpretability.

2. System Testing:

   • In collision selection tests for multiple object motion patterns, FSN circuits can effectively select approaching objects and produce peak responses before collisions. The model’s responses closely match experimental data from locusts.
   • In physical stimuli experiments in real scenes, the model showed excellent stability and robustness in complex backgrounds and high-dynamic scenes.

3. Testing in Low-Contrast Scenes:

   • Experiments confirmed that FSN circuits could replicate the adaptive characteristics of LGMD neurons, with strong inhibitory effects in low-contrast and complex backgrounds consistent with observed LGMD behavior.

Research Summary and Application Value

This study proposes a biomimetic visual detection model by combining event-driven cameras and FSN circuits through low-level modeling. This model not only demonstrates high robustness and flexibility under visual uncertainty but also has a rapid response capability, providing potential application value for detecting and navigating rapidly moving targets in complex scenes in the future. Additionally, this study expands a novel biomimetic computing method based on the integration of multi-characteristics full-feature simulation, laying the foundation for future developments in neuromorphic computing and intelligent robotics.