Boston University
Neuromorphic models and principles of the human visual system for sense and avoid capabilities in autonomous robotic systems
Abstract
dc:description.abstractVisually-guided navigation in real-world scenarios is a very challenging problem that humans and other species across the animal kingdom handle with ease. Nevertheless, many basic, fundamental visual mechanisms have proven very difficult to expound for use in autonomous robotic systems to date. Unmanned aerial vehicles (UAVs), specifically, have become increasing common. For safety and economic reasons it will be critically important that these aircraft become fully autonomous, including the capability to sense and avoid obstacles in the air without human intervention. This thesis introduces two new models that seek to address how humans represent motion in the early visual system and how human-inspired navigation strategies can be combined with classical computer vision approaches to successfully navigate the world while avoiding obstacles. The first model introduced is the Space-variant Cortical Representation for Apparent Motion (SCRAM) model, which proposes an explanation for how aspects of apparent motion appear linear despite the human early visual system containing many non-linear mechanisms. The model combines cortical magnification, receptive field overlap and scatter, and spatial and temporal response characteristics present in retinal ganglion cells into a single cortical representation for motion based on Gaussian activity profiles. The model output is consistent with psychophysics data, finding that the maximum flash distance that is perceived as apparent motion increases linearly as a function of eccentricity. The second model introduced is a model for FLying Independently, Guided by a Human-inspired Tree-cascade detector for Radially Expanding Flow and Salience, aka the FLIGHT REFS model. The FLIGHT REFS model is composed of two systems that work in concert towards a complete sense and avoid capibility for UAVs: (1) a parafoveal system that seeks to detect targets of interest based on goal-directed, learned saliency and (2) a foveal system that ties context learning to a classical computer vision framework to identify objects on a collision course by detecting radially expanding flow fields. Results from testing the model on both simulated and real-world collision avoidance datasets demonstrate that classical computer vision approaches for collision avoidance can be improved with techniques modeled after the human visual system.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Cognitive & Neural Systems
- Grantor
- Boston University
- Year dc:date.issued
- 2016
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Wurbs, Jeremy
- Advisors dc:contributor.advisor
-
- Versace, Massimiliano
- Yazdanbakhsh, Arash
Subjects
dc:subject × 2Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53562