Optimizing Fire Detection in Edge Devices: Integrating Early Exits in Compact Models
Document Type
Conference Proceeding
Publication Date
2024
Abstract
Efficient fire detection is vital for addressing the downstream ecological and economic impacts caused by rapidly progressing wildfires. This paper proposes using machine learning with sample-wise dynamic models, such as early exits, in fire image processing, to ensure quick identification of flames and smoke. Tailored for edge devices, this approach offers an efficient solution for resource-constrained embedded platforms deployed to fire-prone regions (e.g., aerial drones, orbiting satellites). Our investigation demonstrated accuracy ranging from 87% to 93%, with smaller models like a ResNet-18 backbone consistently outperforming larger architectures. Adopting an early exit algorithm showed a reduction inference latency on edge devices by more than 50% when compared to baseline machine learning models. This work emphasizes the value of compact models for robust fire detection, with dynamic backbone architectures further enhancing performance. © 2024 Elsevier B.V., All rights reserved.
Recommended Citation
Vincent, Grace M.; Desantis, Laura; Wilkerson, Matthew; Couch, Nathan; Bhattacharya, Sambit; Hasnain, Zaki; and Ingham, Michel D., "Optimizing Fire Detection in Edge Devices: Integrating Early Exits in Compact Models" (2024). College of Health, Science, and Technology. 1155.
https://digitalcommons.uncfsu.edu/college_health_science_technology/1155