A Directionally Selective Neural Network with Separated ON and OFF Pathways for Translational Motion Perception in a Visually Cluttered Environment

TL;DR

This paper introduces a biologically inspired neural network model that separates ON and OFF pathways to enhance translational motion perception in cluttered environments, improving speed response and robustness over previous models.

Contribution

The model mimics fly's physiology, separates ON/OFF pathways, and improves motion detection speed and clutter robustness compared to prior bio-plausible detectors.

Findings

Outperforms previous bio-plausible motion detectors

Enhances speed response to dark/light features

Reduces impact of irrelevant background motion

Abstract

With respect to biological findings underlying fly's physiology in the past decade, we present a directionally selective neural network, with a feed-forward structure and entirely low-level visual processing, so as to implement direction selective neurons in the fly's visual system, which are mainly sensitive to wide-field translational movements in four cardinal directions. In this research, we highlight the functionality of ON and OFF pathways, separating motion information for parallel computation corresponding to light-on and light-off selectivity. Through this modeling study, we demonstrate several achievements compared with former bio-plausible translational motion detectors, like the elementary motion detectors. First, we thoroughly mimic the fly's preliminary motion-detecting pathways with newly revealed fly's physiology. Second, we improve the speed response to moving dark/light features via the design of ensembles of same polarity cells in the dual-pathways. Moreover, we alleviate the impact of irrelevant motion in a visually cluttered environment like the shifting of background and windblown vegetation, via the modeling of spatiotemporal dynamics. We systematically tested the DSNN against stimuli ranging from synthetic and real-world scenes, to notably a visual modality of a ground micro robot. The results demonstrated that the DSNN outperforms former bio-plausible translational motion detectors. Importantly, we verified its computational simplicity and effectiveness benefiting the building of neuromorphic vision sensor for robots.