Robust and stable detection of visual features from objects with depth-axis motion against variable scenes: locust-inspired visual neural pathway modeling
Year: 2026
Type: article
Abstract: Currently, Locust-inspired visual models based on the Lobula Giant Movement Detector 1 (LGMD1) have achieved significant advances in the field of intelligent obstacle avoidance. However, the decision-making reliability of obstacle avoidance systems is highly dependent on the robust detection of visual features of moving objects. Specifically, substantial excitation from translating moving objects and variable scenes pose challenges for robust (i.e., resilience under variable low-contrast scenes) and stable (i.e., consistency despite spatial-contrast variations) detection of visual features from moving objects in the LGMD1-based model. To address this challenge, we present a Locust-inspired motion-object depth-axis vision-feature-detection (MODAVFD) model that integrates ON and OFF pathway competition and contrast neural computation, which shows a preference for objects with depth-axis motion (i.e., motion along the depth axis) and enables robust and stable detection of visual features against variable scenes. The MODAVFD model integrates successive neural computational layers from the retina to the lobula complex. First, changes in luminance are captured by the photoreceptor cells and divided into parallel ON and OFF pathways. Second, the lateral inhibition (LI) mechanism is employed as early visual preprocessing to enhance the selectivity of motion boundaries, and the non-linear instantaneous feedback divisive contrast normalization mechanism is adopted to reduce sensitivity to local contrast. Third, the parallel contrast pathways are activated to inhibit the parallel motion pathways, suppressing high-spatial-contrast optical flow. Finally, neural competition between the ON and OFF pathways is conducted and the winner is transmitted to the LGMD1. Notably, the MODAVFD model does not discriminate between approaching and receding motion, and a downstream mechanism is therefore required before its use in collision avoidance. Through extensive comparative experiments with state-of-the-art (SOTA) bio-inspired models, the MODAVFD model demonstrates a preference for objects with depth-axis motion and responds robustly to low-spatial-contrast inputs. Moreover, the performance of the MODAVFD model was evaluated under three object velocities. First, regarding the sum of the coefficient of variation, it decreased by 40.91%, 74.03%, and 92.36% on the synthetic pure dataset; by 78.56%, 74.95%, and 85.26% on the semi-synthetic cluttered dataset; and by 12.11%, 8.77%, and 17.96% on the semi-synthetic real-world dataset, respectively. Notably, the improvement of the MODAVFD model on the semi-synthetic real-world dataset is smaller than that on the other datasets, which is primarily attributed to the multi-class objects in the backgrounds of this dataset, introducing inter-class interference. Second, regarding the inter-quartile range (IQR) of the coefficient of variation, it changed by − 46.36%, 26.75%, and 44.65% on the semi-synthetic cluttered dataset, and by 7.99%, 5.19%, and 15.38% on the semi-synthetic real-world dataset, respectively. Importantly, the aforementioned improvement is not consistent across object motion velocities and baselines, but the response stability is improved overall. However, the model exhibits a known limitation when processing bisected black and white rectangular stimuli: because such stimuli carry both contrast polarities in balanced proportion, the ON and OFF responses offset at the integration stage, producing no usable output.
Source: Scientific Reports
Authors: Sheng Zhang(Conceptualization, Investigation, Methodology, Writing – original draft), Mengxi Xu(Formal analysis, Funding acquisition, Software, Supervision, Validation), Shengnan Zheng(Data curation, Project administration, Resources, Software, Visualization), Yinglan Ruan(Investigation, Supervision, Writing – original draft, Writing – review & editing)
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: gold
Estimated APC paid: $2,690
Awards: G2129496021, G6086936143, G6605999450
OpenAlex ID: W7234878626
