Dynamic Neural Field Modeling of Visual Contrast for Perceiving Incoherent Looming

TL;DR

The paper addresses robust looming perception in noisy, incoherent visual scenes by extending single-field DNFs to a three-field C-DNF that separately encodes ON/OFF visual contrast and uses lateral Gaussian excitation. The Summation field integrates these signals with a DoG-based interaction to produce a collision indicator, triggering alerts when $I_v$ exceeds a threshold. Empirical results show strong looming selectivity across coherence levels and resilience to synthetic rain, outperforming SDNF and matching or exceeding LGMD2 and FLGMD ONn in many cases, with real-world accuracy around 77.14%. This approach offers a low-energy, brain-inspired solution for reliable collision detection in complex environments and could integrate with broader cognitive DNFs for navigation and decision-making in robots.

Abstract

Amari's Dynamic Neural Field (DNF) framework provides a brain-inspired approach to modeling the average activation of neuronal groups. Leveraging a single field, DNF has become a promising foundation for low-energy looming perception module in robotic applications. However, the previous DNF methods face significant challenges in detecting incoherent or inconsistent looming features--conditions commonly encountered in real-world scenarios, such as collision detection in rainy weather. Insights from the visual systems of fruit flies and locusts reveal encoding ON/OFF visual contrast plays a critical role in enhancing looming selectivity. Additionally, lateral excitation mechanism potentially refines the responses of loom-sensitive neurons to both coherent and incoherent stimuli. Together, these offer valuable guidance for improving looming perception models. Building on these biological evidence, we extend the previous single-field DNF framework by incorporating the modeling of ON/OFF visual contrast, each governed by a dedicated DNF. Lateral excitation within each ON/OFF-contrast field is formulated using a normalized Gaussian kernel, and their outputs are integrated in the Summation field to generate collision alerts. Experimental evaluations show that the proposed model effectively addresses incoherent looming detection challenges and significantly outperforms state-of-the-art locust-inspired models. It demonstrates robust performance across diverse stimuli, including synthetic rain effects, underscoring its potential for reliable looming perception in complex, noisy environments with inconsistent visual cues.

Figures (5)

• Figure 1: The diagram of the proposed C-DNF model for looming perception. The proposed framework consists of three distinct DNF fields: ON-contrast, OFF-contrast, and Summation field. The model input is the luminance change between successive video frames, with each neuron corresponding to a pixel in the input video (e.g., the yellow pixel in the input corresponds to the red neuron in the OFF field for demonstration). For a resolution of $m\times n$, each field comprises $m\times n$ neurons. Luminance increments and decrements are separated serving as input to the respective fields. Neurons within the ON/OFF-contrast fields receive excitatory signals from their $8$ immediate neighbors (as shown by red and pink neurons in the inset), modulated a normalized Gaussian kernel (3D inset). In contrast, the lateral interaction within the Summation field is regulated by a Difference of Gaussian (DoG) function (3D inset), where proximal neurons provide excitation (depicted as red neurons in the inset) and distal neurons provide inhibition (depicted as blue neurons in the inset). The processed ON and OFF signals, scaled by contrast coefficient $\alpha_{on}$ and $\alpha_{off}$ flow into the Summation field. Neuron activations in the Summation field are integrated into an overall signal $I_v$ via a sigmoid-like function, triggering collision alert when $I_v$ exceeds a predefined threshold. Notably, dashed yellow lines in the diagram indicate pixel-wise computations, while solid yellow lines represent integration of the entire neuron field into a single signal.
• Figure 2: The model responses of C-DNF to $6$ synthetic stimuli. Each sub-figure comprises three rows: the first row displays snapshots of the tested synthetic stimuli, the second row presents the model's integrated signal $I_v$ along with the fixed threshold $I_{Thre} = 0.506$ and the third row indicates the occurrence of collision alerts. The proposed model demonstrates a robust response to approaching objects irrespective of the contrast, exhibits symmetric responses to receding objects, and no response to translating objects.
• Figure 3: Tested samples of synthetic incoherent looming stimuli and real-world stimuli with or without the synthetic rain effect (a) Samples of incoherent looming stimuli with varying coherence levels, including $75\%$, $50\%$, $20\%$ and $5\%$. (b) Two recorded real-world collision scenarios are shown, both with and without synthetic rain effects. The first and third rows display snapshots without rain, while the second and fourth rows show the corresponding scenarios with synthetic rain added.
• Figure 4: The response characteristics of the comparative models to different motion stimuli with varying coherence degrees. The first and the second rows of each sub-figure illustrate the spatial-temporal moving pattern of the looming, receding and translating object under the coherence degrees of $5\%$ and $100\%$, respectively. The FLGMD ONn is abbreviated as ONn in the figure. The responses of the models are shown as horizontal colored bars beneath the stimulus representation. In general, most of the tested models retained their looming selectivity, as previously reported in Table \ref{['tab:exp result syn']}, with the exception of FLGMD ONn.
• Figure 5: Comparative experiment results on recorded real-world stimuli with or without synthetic rain effect. The first row of the figures illustrates snapshots of the original tested real-world stimuli, while the second row depicts the same stimuli with synthetic rain effects applied. The manually identified collision period is highlighted with a pink shade in each sub-figure. (a) Comparative experiment results when tested with a stimulus depicting a bus collision. (b) Comparative experiment results when tested with a stimulus depicting a balloon hits the car.