Abstract

Real-time fire detection on edge devices presents significant computational challenges. Existing solutions struggle to balance detection accuracy with efficiency in resource-constrained environments. This paper introduces Adaptive Conditional Model Activation (ACMA), a novel conditional execution framework that optimizes deep learning deployment through dynamic model gating. The proposed approach employs multi-color space analysis and scene-aware adaptive thresholding to selectively activate YOLO model only when preliminary fire indicators exceed dynamically calculated thresholds. Experimental results demonstrate that ACMA achieves 77% filtering accuracy with only a 3.2% system-level accuracy reduction compared to continuous YOLO. While CPU usage reduction appears modest (5%) on severely constrained hardware like Raspberry Pi 4B where baseline utilization is already saturated it enables a transformative throughput to increase from 0.14 to 38 FPS, a 270× improvement. On a desktop i5 CPU, ACMA reduces usage by 80% and increases FPS by 25 times.

References

• Celik, T.: Fast and efficient method for fire detection using image processing. ETRI J. 32(6), (2010).
• Celik, T.: Fast and efficient method for fire detection using image processing. ETRI J. 32(6), (2010).
• Ostroukh, E.N. et al.: Color detection algorithm for early fire diagnostics. J. Phys.: Conf. Ser. 2131, (2021).
• Mazur, M.: Analysis of methods for reducing the number of false alarms in video-based fire detection systems. J. Autom. Electr. Eng. 5(2), (2023).
• Poojary, R., Pai, A.: Comparative study of model optimization techniques in fine-tuned CNN models. In: Proc. International Conference on Electrical and Computing Technologies and Applications (2019).
• Liang, D. et al.: Fire and smoke detection method based on improved YOLOv5s. In: Proc. International Conference on Communication, Image and Signal Processing (2023).
• Hoang, V.H. et al.: Enhancing fire detection with YOLO models: a Bayesian hyperparameter tuning approach. Comput. Mater. Contin. 83(3), (2025).
• He, Y. et al.: DCGC-YOLO: the efficient dual-channel bottleneck structure YOLO detection algorithm for fire detection. IEEE Access 12, (2024).
• Wang, H. et al.: YOLO-LFD: a lightweight and fast model for forest fire detection. Comput. Mater. Contin. 82(2), (2025).
• Mahmoudi, S. et al.: Edge AI system for real-time and explainable forest fire detection using compressed deep learning models. In: Proc. 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics (2025).
• Li, C. et al.: Intelligent monitoring of tunnel fire smoke based on improved YOLOX and edge computing. Appl. Sci. 15(4), (2025).
• Xiao, L. et al.: EMG-YOLO: an efficient fire detection model for embedded devices. Digit. Signal Process. 156, (2025).
• Vazquez, G. et al.: Detecting wildfire flame and smoke through edge computing using transfer learning enhanced deep learning models. arXiv preprint arXiv:2501.08639 (2025).
• Ryu, J., Kwak, D.: A method of detecting candidate regions and flames based on deep learning using color-based pre-processing. Fire 5(6), (2022).
• Vincent, G. et al.: Optimizing fire detection in edge devices: integrating early exits in compact models. In: AIAA Aviation Forum (2024).
• Gong, L. et al.: Explainable semantic federated learning enabled industrial edge network for fire surveillance. IEEE Trans. Ind. Inform. (2024).