PaperlayerView exact revision on arXiv

Source-linked AI summary

A Bio-Plausible Visual Neural Network for Locust-Inspired Collision Perception

Qinbing Fu, Jiani Li, Jiajun Huang, Jigen Peng

arXiv:2609.10183v1cs.NEq-bio.NC

TL;DR

Looming perception supports collision-related safety decisions, while traditional visual approaches can be costly and data-intensive for resource-constrained real-time systems. The paper proposes a locust-inspired network combining isotropic spatial sampling, population voting, inhibition, and LIF dynamics. Experiments report collision-cue detection and looming selectivity across controlled stimuli and real-world driving scenarios.

  • Problem

    Traditional visual approaches can require substantial computation, memory, and training data, limiting their use in resource-constrained real-time systems where looming detection supports safety decisions.

    Source: §I · p. 1

  • Method

    The network combines hexagonal ommatidial sampling, population voting, inhibitory processing, and leaky integrate-and-fire neuronal dynamics for looming detection.

    Source: §I · pp. 1–2, A. Isotropic Spatial Sampling · p. 2

  • Results

    Experiments showed collision-cue detection in crash sequences and low activity in non-collision driving scenarios, alongside looming selectivity across controlled and real-world stimuli.

    Source: A. Basic Functionality and Selectivity · p. 5

  • Takeaways & Limitations

    The experiments demonstrate collision-perception capability across controlled visual stimuli and complex real-world environments.

    Source: A. Basic Functionality and Selectivity · p. 5

    Vehicle-data precision was 47.30%, with false positives associated with complex motion, background motion, and camera-induced disturbances.

    Source: C. Robustness in Vehicle Data Set · p. 6

Abstract

from arXiv · show

Locust visual systems have long served as an important biological paradigm for studying looming perception and collision avoidance. Numerous computational models have successfully reproduced the selective responses of Lobula Giant Movement Detector (LGMD) neurons to approaching objects, thereby emulating the fundamental functionality of the biological system. However, existing models remain limited in biological plausibility and robustness when operating in complex and dynamic visual environments. To address these limitations, we propose a biologically plausible neural network for locust-inspired looming detection. The proposed framework incorporates a spatially isotropic sampling strategy that mimics the ommatidial organization of the locust compound eye, a population-voting mechanism inspired by population coding in biological neural systems, and leaky integrate-and-fire neuronal dynamics to replace conventional sigmoid-based membrane activation. Systematic experiments on synthetic stimuli, laboratory sequences, and real-world driving scenarios demonstrate that the proposed model improves robustness under challenging visual conditions while preserving computational efficiency and enhancing biological fidelity. These results highlight the potential of biologically grounded neural computation for robust and efficient collision perception.

I. INTRODUCTION

Existing LGMD-based models face robustness and biological-plausibility limitations in complex scenes; this work proposes a locust-inspired network with ommatidial sampling, population voting, inhibition, and LIF dynamics. Its pipeline processes spatially sampled visual input through motion and inhibitory stages before population integration and LIF output.

I. INTRODUCTION

  • Existing LGMD models rely on single-sigmoid decisions, lack biologically plausible integration of distributed responses, and often retain pixel-wise receptive fields unlike locust ommatidia.Source: §I · p. 1
  • The proposed network combines ommatidial spatial sampling, population voting, self- and lateral inhibition, and leaky integrate-and-fire output dynamics.Source: §I · pp. 1–2
  • Experiments span synthetic stimuli, laboratory sequences, and ground-vehicle scenes, reporting improved biological plausibility, robustness, and stability with low computational complexity.Source: §I · p. 2

II. METHOD

  • The method encodes mean intensity on hexagonal units, then processes ON/OFF signals through excitatory and inhibitory interactions before voting and LIF output.Source: §II · p. 2, §I · p. 2

A. Isotropic Spatial Sampling

The model samples images on a hexagonal lattice of local receptive fields, then processes ON/OFF motion signals through excitation and inhibitory pathways. The sampling preserves isotropic spatial relationships, while the processing combines local and global inhibition.

A. Isotropic Spatial Sampling

  • Hexagonal receptive fields output mean pixel intensity, preserving isotropic spatial relationships and reducing directional bias relative to rectangular grids.Odd rows are horizontally offset by half a grid spacing relative to even rows.Source: A. Isotropic Spatial Sampling · p. 2

B. Multi-Layer Processing of ON/OFF-Contrast

  • After spatial sampling, temporal brightness changes form ON/OFF contrast signals, with the OFF pathway following the same computational structure as the ON pathway.The remaining computation is described for the ON pathway; OFF follows an identical structure.Source: B. Multi-Layer Processing of… · pp. 2–3
  • Feed-forward inhibition suppresses excessive global activation when many photoreceptors are excited simultaneously, using their spatially averaged activation.Source: B. Multi-Layer Processing of… · p. 2
  • ON/OFF responses are aggregated across the spatial field to model population-level activity.Source: B. Multi-Layer Processing of… · p. 3
  • Gaussian excitation is shaped by lateral inhibition between neighboring neurons and self-inhibition based on the local visual field.Lateral inhibition makes one neuron's activation suppress surrounding neurons; self-inhibition reduces excessive local excitation.Source: B. Multi-Layer Processing of… · pp. 2–3

C. Population-Voting Mechanism

The population-voting mechanism combines local neural responses across neighboring receptive fields into a global response. Binary activation and whole-field integration carry this population activity downstream.

C. Population-Voting Mechanism

D. Neuronal Dynamics

A leaky integrate-and-fire neuron generates the final collision decision by integrating activity over time. Reaching its firing threshold produces a spike and resets the membrane potential.

D. Neuronal Dynamics

  • A leaky integrate-and-fire neuron generates the final collision decision by modeling temporal integration dynamics of biological LGMD neurons.Source: D. Neuronal Dynamics · p. 4
  • When membrane potential reaches the firing threshold, the neuron generates a spike and resets its potential.Source: D. Neuronal Dynamics · p. 4
  • The input current is based on population activity dynamics, with video sampled in frames per second; the mechanism includes temporal integration, threshold-triggered spikes, and post-spike reset.Source: D. Neuronal Dynamics · p. 4

III. RESULTS AND ANALYSIS

The experiments cover controlled laboratory sequences and real-world dashboard-camera recordings, including collision and non-collision driving scenarios. The section also identifies the computing hardware used.

III. RESULTS AND ANALYSIS

  • Evaluation covers controlled laboratory video sequences and real-world dashboard-camera recordings with collision and non-collision driving scenarios.Source: §III · p. 4
  • The section's stimuli include looming and translational motion of an indoor ball and an outdoor ground vehicle.Source: Figure 4 · p. 4
  • Experiments were conducted on a workstation with an Intel Core i5-10500 CPU and 16 GB of RAM.Source: §III · p. 4

A. Basic Functionality and Selectivity

The model responds selectively to approaching motion across synthetic, laboratory, and real-world stimuli. It generates action potentials for synthetic looming, responds to approaching balls across target speeds, and detects collision cues in crash sequences.

A. Basic Functionality and Selectivity

1) Spatial Sampling:

The model uses a hexagonal sampling grid to mimic locust ommatidia and preserve more isotropic spatial relationships. Compared with bilinear interpolation and Gaussian downsampling, it better preserves edges and structural consistency while maintaining comparable information entropy.

1) Spatial Sampling:

  • A hexagonal grid mimics the locust compound eye’s ommatidial arrangement and preserves more isotropic relationships between neighboring samples.Each sample aggregates photoreceptor responses within a surrounding local receptive field.Source: 1) Spatial Sampling · p. 5
  • The hexagonal scheme provides better edge preservation and structural consistency than bilinear interpolation and Gaussian downsampling, with comparable information entropy.The authors suggest this encoding is better suited to edge-sensitive motion processing and improves spatial coverage uniformity while reducing directional bias.Source: 1) Spatial Sampling · p. 5

2) Visualization of Spatiotemporal Signal Patterns:

The visualizations show how the model’s intermediate responses differ between looming and translational motion. Looming drives increasingly strong excitation and rising summation activity, whereas inhibition suppresses the weak responses to translation.

2) Visualization of Spatiotemporal Signal Patterns:

  • As the approaching square draws closer, excitation strengthens and concentrates spatially, while self-inhibition and lateral inhibition lag behind it.The bubble maps represent signal intensity by bubble area across the stimulus, excitation, self-inhibition, lateral inhibition, and residual excitation after inhibition.Source: 2) Visualization of Spatiotemporal… · p. 5, Figure 5 · p. 5
  • Under translation, excitation remains weak and localized after a brief onset transient, while stable self-inhibition and lateral inhibition help keep output low.The combined inhibition suppresses weak excitatory inputs, preventing substantial response accumulation.Source: 2) Visualization of Spatiotemporal… · pp. 5–6, Figure 6 · p. 5
  • Fig. 7 evaluates model robustness under Gaussian noise for both looming and translating motion.The first row shows looming-motion results and the second row translating-motion results.Source: Figure 7 · p. 6

3) Robustness Against Noise:

On real-world dashcam sequences, the model combines high collision recall with lower precision. The paper also reports stable responses under Gaussian noise, LIF-based spiking dynamics, and a unified framework combining spatial sampling, population voting, and spiking dynamics.

3) Robustness Against Noise:

  • Under Gaussian noise perturbations, the model maintained stable response characteristics, which the authors attribute to spatial averaging and population redundancy.The redundancy helps prevent isolated noise-induced excitations from dominating the final response.Source: 3) Robustness Against Noise · p. 6

4) Neuronal dynamics:

  • The global signal is fed directly to an LIF neuron, preserving temporal accumulation and producing event-driven spiking responses consistent with biological neurons.The design omits additional nonlinear transformations before the LIF neuron.Source: 4) Neuronal dynamics · p. 6

C. Robustness in Vehicle Data Set

  • On 194 dashcam sequences, the model achieved 76.29% accuracy, 83.33% recall, and 60.35% F1-score.The set contained 42 collision and 152 non-collision cases; precision was 47.30%, with false positives attributed mainly to complex motion, background motion, and camera-induced disturbances.Source: C. Robustness in Vehicle Data Set · p. 6, Table III · p. 6

IV. SUMMARY

  • The framework integrates spatially isotropic sampling, population voting, and spiking neural dynamics for collision detection.The summary describes distributed population coding as enhancing robustness against complex visual disturbances while preserving looming selectivity and improving biological fidelity.Source: §IV · p. 6
Loading 2609.10183v1…