[Paper Review] A neural model of the locust visual system for detection of object approaches with real-world scenes

Matthias S. Keil, E. Roca+1 more|arXiv (Cornell University)|Jan 24, 2018

Neural dynamics and brain function15 references16 citations

TL;DR

This paper presents a biologically inspired neural model of the locust's LGMD neuron for robust collision detection in real-world scenes. By integrating ON/OFF pathways, feedback lateral inhibition, and diffusion-based signal propagation, the model effectively suppresses responses to background motion and low-contrast objects, achieving reliable detection of approaching objects even under adverse visual conditions.

ABSTRACT

In the central nervous systems of animals like pigeons and locusts, neurons were identified which signal objects approaching the animal on a direct collision course. Unraveling the neural circuitry for collision avoidance, and identifying the underlying computational principles, is promising for building vision-based neuromorphic architectures, which in the near future could find applications in cars or planes. At the present there is no published model available for robust detection of approaching objects under real-world conditions. Here we present a computational architecture for signalling impending collisions, based on known anatomical data of the locust \emph{lobula giant movement detector} (LGMD) neuron. Our model shows robust performance even in adverse situations, such as with approaching low-contrast objects, or with highly textured and moving backgrounds. We furthermore discuss which components need to be added to our model to convert it into a full-fledged real-world-environment collision detector. KEYWORDS: Locust, LGMD, collision detection, lateral inhibition, diffusion, ON-OFF-pathways, neuronal dynamics, computer vision, image processing

Motivation & Objective

• To develop a biologically plausible neural model of the locust LGMD neuron for detecting object approaches in real-world visual environments.
• To address the challenge of robust collision detection under adverse conditions such as low-contrast objects and moving backgrounds.
• To improve upon prior models by introducing feedback lateral inhibition to prevent saturation of inhibition pathways.
• To enable practical application in neuromorphic vision systems for autonomous vehicles by incorporating key neurophysiological principles.

Proposed method

• The model uses two parallel pathways—ON and OFF—sensitive to luminance increments and decrements, respectively, to extract motion signals from video sequences.
• Movement detectors (MDs) are modeled using a differential equation with leaky integration, capturing local luminance changes over time.
• Summing units (SUs) integrate MD outputs and feed them into a diffusion layer that models lateral inhibition via a reaction-diffusion process.
• Feedback inhibition is implemented such that inhibition is decoupled from its own excitation, allowing it to dissipate and avoid persistent saturation.
• The LGMD neuron integrates activity from SUs, with its output representing the final collision signal.
• The model uses logarithmic encoding and nonlinear integration to approximate the η-function response, consistent with neurophysiological data.

Experimental results

Research questions

• RQ1 How can a neural model of the locust LGMD neuron detect approaching objects robustly in real-world scenes with complex backgrounds?
• RQ2 What role does feedback lateral inhibition play in suppressing responses to non-collisional motion such as background movement?
• RQ3 How do ON and OFF pathways contribute to improved detection of approaching objects compared to single-pathway models?
• RQ4 Can the model maintain sensitivity to approaching objects under low-contrast or high-texture conditions?
• RQ5 What additional mechanisms are required to make the model suitable for real-world applications like autonomous navigation?

Key findings

• The model successfully detects approaching objects in real-world videos, including complex scenes like pedestrian crossings and highway traffic.
• Feedback lateral inhibition prevents saturation of the inhibition layer, allowing the model to remain responsive even under strong background motion.
• The ON and OFF pathways reduce overall activity in the diffusion layer, minimizing false suppression and improving signal-to-noise ratio.
• Secondary response peaks due to object entry or exit are suppressed by lateral inhibition, with recovery occurring after inhibition dissipates.
• The model's output correlates well with the η-function, indicating that the LGMD's response dynamics are effectively emulated through logarithmic encoding and nonlinear integration.
• Despite robust performance, the model still responds to non-colliding large objects, indicating a need for additional decision logic to distinguish true collisions.