Evaluating the Feasibility of a Real-Time Fire Detection System on i.MX RT1064

Abstract

With the increasing frequency of wildfires due to climate change, early and accurate fire detection systems are critical to minimizing damage and ensuring public safety. This thesis explores the feasibility of deploying a real-time fire detection system on a resource-constrained Micro Controller Unit (MCU) integrated into power line monitoring units.

In collaboration with AVJU, a Norwegian power line monitoring company, this research focuses on developing a prototype system that integrates a USB camera with an i.MX RT1064 MCU that they are deploying in their power line monitoring units. To process images through a pre-trained fire detection model. The system utilizes FreeRTOS for task scheduling and synchronization, SDRAM for handling large image and tensor data, and interpolation methods to optimize image preprocessing.

Evaluation of the system focused on measuring the performance of three interpolation algorithms on the i.MX RT1064. We also evaluate the performance of each task, including runtime, resource usage (in terms of execution time), and system efficiency, to pinpoint areas for optimization. These evaluations help us discuss and provide a more precise answer to the feasibility of the proposed system.

The thesis examines critical bottlenecks identified during the initial working implementation through experiments, discusses solutions to these bottlenecks, and presents a faster and more efficient version of the system. The findings offer insights into the design and implementation of edge AI systems for environmental monitoring and provide a starting point for future deployment of real-time hazard detection systems in monitoring units.