Neuromorphic perception for greenhouse technology using event-based sensors

  • Sami El Arja

Western Sydney University thesis: Doctoral thesis

Abstract

Event-Based Cameras (EBCs), unlike conventional cameras, feature independent pixels that asynchronously generate outputs upon detecting changes in their field of view. Short calculations are performed on each event to mimic the brain. The output is a sparse sequence of events with high temporal precision. Conventional computer vision algorithms do not leverage these properties. Thus a new paradigm has been devised. While event cameras are very efficient in representing sparse sequences of events with high temporal precision, many approaches are challenged in applications where a large amount of spatially-temporally rich information must be processed in real-time. In reality, most tasks in everyday life take place in complex and uncontrollable environments, which require sophisticated models and intelligent reasoning. Typical hard problems in real-world scenes are detecting various non-uniform objects or navigation in an unknown and complex environment. In addition, colour perception is an essential fundamental property in distinguishing objects in natural scenes. Colour is a new aspect of event-based sensors, which work fundamentally differently from standard cameras, measuring per-pixel brightness changes per colour filter asynchronously rather than measuring "absolute" brightness at a constant rate. This thesis explores neuromorphic event-based processing methods for high-noise and cluttered environments with imbalanced classes. A fully event-driven processing pipeline was developed for agricultural applications to perform fruits detection and classification to unlock the outstanding properties of event cameras. The nature of features in such data was explored, and methods to represent and detect features were demonstrated. A framework for detecting and classifying features was developed and evaluated on the N-MNIST and Dynamic Vision Sensor (DVS) gesture datasets. The same network was evaluated on laboratory recorded and real-world data with various internal variations for fruits detection such as overlap, variation in size and appearance. In addition, a method to handle highly imbalanced data was developed. We examined the characteristics of spatio-temporal patterns for each colour filter to help expand our understanding of this novel data and explored their applications in classification tasks where colours were more relevant features than shapes and appearances. The results presented in this thesis demonstrate the potential and efficacy of event- based systems by demonstrating the applicability of colour event data and the viability of event-driven classification.
Date of Award2022
Original languageEnglish

Keywords

  • neuromorphics
  • computer vision
  • pattern recognition systems
  • neural networks (computer science)

Cite this

'