Neuromorphic Vision Processing

Student thesis: Doctoral Thesis

Abstract

In recent years there has been a paradigm shift in the field of computer vision with the emergence of event-based vision. Event-based vision is founded on concepts derived from understanding of biological vision systems and their hardware implementations. It is a new approach to visual sensing, capture and communication which allows for transmission of asynchronous data and is becoming an increasingly important area of research in robotics and perception. This thesis introduces a number of contributions to event-based vision processing. The sequence of the research contributions begins with the development of a low-level descriptor and framework for event-based vision data that are evaluated in terms of motion recognition problems. The descriptor consists of feature vectors that describe motion and event data patterns respectively. The research into the event data pattern vector forms its own contribution as several bio-inspired neural coding strategies. Vector representations of the encoded pattern of event data are proposed and implemented; these strategies are evaluated with respect to performance accuracy and compared with the state-of-the-art event vision approaches. Finally, a binary tree data structure is introduced which is specifically designed to represent and encode event-based data enabling rapid access to recent correlated events through efficient in-memory data management. This data structure is self-balancing and self-pruning. A framework is developed that exploits characteristics of the novel data structure for the identification of corners and edge features within event-based vision data. This framework is shown to achieve near real-time processing while maintaining accurate feature extraction compared with the state-of-the-art.
Date of AwardDec 2020
Original languageEnglish
SupervisorPratheepan Yogarajah (Supervisor), Sonya Coleman (Supervisor) & Dermot Kerr (Supervisor)

Keywords

  • Event-based vision
  • Motion description
  • Neuromorphic processing
  • Motion recognition
  • Object classification
  • Low-level feature extraction

Cite this

'