Peripheral Processing Facilitates Optic Flow-Based Depth Perception

Abstract

Flying insects, such as flies or bees, rely on consistent information regarding the depth structure of the environment when performing their flight maneuvers in cluttered natural environments. These behaviors include avoiding collisions, approaching targets or spatial navigation. Insects are thought to obtain depth information visually from the retinal image displacements (“optic flow”) during translational ego-motion. Optic flow in the insect visual system is processed by a mechanism that can be modeled by correlation-type elementary motion detectors (EMDs). However, it is still an open question how spatial information can be extracted reliably from the responses of the highly contrast- and pattern-dependent EMD responses, especially if the vast range of light intensities encountered in natural environments is taken into account. This question will be addressed here by systematically modeling the peripheral visual system of flies, including various adaptive mechanisms. Different model variants of the peripheral visual system were stimulated with image sequences that mimic the panoramic visual input during translational ego-motion in various natural environments, and the resulting peripheral signals were fed into an array of EMDs. We characterized the influence of each peripheral computational unit on the representation of spatial information in the EMD responses. Our model simulations reveal that information about the overall light level needs to be eliminated from the EMD input as is accomplished under light-adapted conditions in the insect peripheral visual system. The response characteristics of large monopolar cells (LMCs) resemble that of a band-pass filter, which reduces the contrast dependency of EMDs strongly, effectively enhancing the representation of the nearness of objects and, especially, of their contours. We furthermore show that local brightness adaptation of photoreceptors allows for spatial vision under a wide range of dynamic light conditions.

Local motion adaptation enhances the representation of spatial structure at EMD arrays

“…With genetic tools, more and more details about the neuronal basis of the motion detector circuits are being unraveled [ 12 – 20 ]. It has been shown in modeling studies that signals represented at the output of EMD arrays correlate well with the contrast-weighted nearness during behaviorally shaped translational self-motion [ 21 , 22 ].…”

“…First, we developed an adaptive model of the visual motion pathway of insects that captures benchmark features of motion adaptation as analyzed in previous electrophysiological studies on LPTCs [ 26 – 28 ]. Our adaptive EMD model is based on an adaptation mechanism similar to the mechanisms previously proposed for light adaptation by photoreceptors [ 22 ], here however, operating on the output of EMDs and with much larger time constants. Based on this adaptive model of the visual motion pathway, our intention was to understand how motion adaptation affects the signal representation at the output of arrays of motion detectors and, in particular, the representation of the spatial layout of the environment during translational self-motion in 3D environments.…”

Resource-efficient bio-inspired visual processing on the hexapod walking robot HECTOR

“…Due to the large panoramic visual field and the low spatial resolution of the visual system of flies [30] the computational requirements of processing optic flow can be reduced, since optic flow induced on the retina by selfmotion or motion in the environment can be processed with a relatively small number of computational units. Hence, the preprocessing of camera images is based on a) remapping and scaling of the visual input to emulate the relatively low spatial resolution of the fly's compound eye, as well as, b) brightness adaptation to varying light intensities, such as performed by photoreceptors and peripheral visual interneurons [31,32], which allows flies to navigate their environment even under dynamic brightness conditions [26].…”

“…The array of the absolute values of these local motion vectors μ r due to translatory selfmotion resembles a map of contrast-weighted relative nearness to objects in the environment, providing information on the contours of nearby objects [25,26].…”

“…EMDs can be used to model the neuronal mechanisms for estimating optic flow in flying insects [21,22], avian [23] and other vertebrate species including man (for review, see [24]). The responses of EMDs to pure translational optic flow have been concluded to resemble a representation of the relative contrast-weighted nearness to objects in the environment, or, in other words, of the contours of nearby objects [16,[25][26][27][28]. The collision avoidance model (a) extracts nearness information from optic flow by EMDs, (b) determines a collision avoidance direction from the map of nearness estimates and (c) computes a collision avoidance necessity, i.e.…”