Integration of a Hardware Accelerator for On-Device Localization in an FPGA-Based Video Capsule Prototype
Master’s Thesis
Abstract
In video capsule endoscopy, most of the capsule’s limited energy budget is spent on illumination and the wireless transmission of every captured image, even though many frames show organs irrelevant to the examination. This thesis integrates the resource-efficient UltraTrail AI accelerator into the FPGA of an Ovesco capsule prototype, using a Lattice CrossLink-NX-33 and a SystemVerilog design, so that a compact neural network can localize the capsule within the GI tract on-device and restrict transmission to the organ of interest. The design is verified in simulation and tested on real hardware via a USB-to-SPI adapter and a hardware-in-the-loop camera model, and is extended with improved memory organization, a Hidden Markov Model for more accurate localization, and pipelining of image capture with model execution to further reduce energy consumption.