Motion perception and the scene statistics of motion
Motion coding in the brain undoubtedly reflects the statistics of retinal image motion occurring in the natural environment. Measuring the statistics of motion in natural scenes is an important tool for building our understanding of how the brain works. Unfortunately, there are statistics that are either impossible or prohibitively difficult to measure. For this reason, it is useful to measure scene statistics in artificial movies derived from simulated environments. This is a novel and important methodological approach that allows us to ask questions about optimal coding that are impossible otherwise. This dissertation describes a course of research that develops this research methodology, the simulated scene statistical approach. This dissertation applied the artificial scene statistical approach to understanding the visual statistics of motion during navigation through forest environments. An environmental model of forest scenes was developed based on previously measured range and surface texture statistics. Spatiotemporal power spectra were measured in both simulated and natural scenes for the task of first person motion through a forest environment. These image statistics measurements helped validate the environmental model. Next, the environmental model was used to simulate across-domain statistics to study the ideal aperture size of motion sensors. It was found that across a variety of different scene conditions, the optimal aperture size of motion sensors increases with the speed to which the sensor is tuned. This is an important constraint for understanding both how the brain encodes motion as well as for designing computer motion detectors. This theoretical research inspired a psychophysical experiment estimating the receptive-field size of human foveal motion discrimination. It was found that for narrow-band stimuli the ideal aperture size increases with spatial frequency, but is unchanging with respect to velocity or temporal frequency. This dissertation shows an approach to the study of vision that has applications in psychophysics, neuroscience and computer vision. The emphasis on accurate and validated environmental models for simulating scene statistics can help improve our understanding of the structure and function of the human visual system and also help us build more accurate and robust computer vision systems.