Dendritic arithmetic and dynamics for neuromorphic temporal pattern detection

Abstract

This thesis investigates the utilization of event-based camera technology to detect motion within a grid of pixels, emphasizing the development of models readily adaptable to neuromorphic circuits. Event-based cameras, renowned for their asynchronous sensing mechanism, offer distinct advantages over conventional frame-based cameras, boasting high temporal resolution and minimal latency. Leveraging this technology, the thesis introduces novel algorithms and methodologies to process the stream of events generated by the camera, with a focus on spatio-temporal filtering to exploit the event-based camera's advantages. The research begins with an in-depth examination of event-based camera principles, emphasizing the generation of events based on brightness changes, while bypassing detailed hardware functioning as this thesis focuses on the motion detection model's mathematical design. Subsequently, existing movement detection algorithms are reviewed comprehensively, identifying limitations and areas for enhancement within the event-based camera framework. Building upon this groundwork, the thesis presents innovative approaches for movement detection, tailored to exploit the event-based camera's characteristics and inspired by dendritic spike generation and the non-linear properties of dendrites. Two distinct methodologies are investigated, structured following the organization of dendritic branches, where each dendritic compartment receives events from a specific pixel. The first approach leverages dendritic arithmetic with additive operations, while the second integrates both additive and multiplicative arithmetic operations, mimicking the excitation and inhibition between dendritic compartments. Equilibrium and nullcline analyses are performed to optimize model parameters. Results reveal the first approach's susceptibility to event shape dependency, while the second approach exhibits robustness with minimal dependency on event shape. Further exploration includes the study of consecutive movement detection and the development of a bidirectional movement detection model using the second approach. Additionally, the extension of the working model to a 2-dimensional setup is conducted to detect 2-dimensional movements effectively. Validation of proposed methodologies is achieved through extensive simulations, where the limitations of the designed models are investigated, and the range of parameter values under which the models function optimally is determined. These simulations provide insights into the robustness and adaptability of the developed approaches across various environmental conditions and scenarios. The thesis concludes with discussions on potential applications and implications of the developed movement detection system in domains such as targets of interest tracking or moving obstacles detection. By harnessing event-based camera technology and drawing inspiration from neural computation, this research contributes to advancing movement detection capabilities, laying the foundation for future developments in visual sensing and perception.