Low-Light Fire and Smoke Detection in Electric Vehicles via a Lightweight Dual-Branch Architecture with Adaptive Reflectance Enhancement

Abstract

Fire and smoke detection in electric vehicle environments is essential for safety monitoring, yet existing methods often perform poorly under low-light conditions because of weak feature representation, flame overexposure, and limited nighttime data. In addition, many detection models cannot effectively balance accuracy and real-time efficiency, which restricts their deployment in practical applications. To address these issues, this study proposes a lightweight dual-branch detection framework for low-light fire and smoke recognition in electric vehicle environments. The proposed method integrates reflectance enhancement, adaptive feature weighting, and multi-scale feature learning within a unified architecture. Specifically, the auxiliary branch enhances weak fire-related features under low illumination through the Adaptive Reflectance Enhancement Module (AREM) and the Element-wise Adaptive Weight Convolution (EAConv), while the main branch employs a Self-Adaptive Feature Convolution (SAConv) module to strengthen multi-scale representation with limited computational overhead. Moreover, a self-constructed and publicly released dataset containing 3,512 images is introduced, including 1,910 daytime and 1,602 nighttime samples, to provide a more realistic benchmark for fire detection in electric vehicle environments under diverse lighting conditions. Experimental results show that the proposed method achieves 85.3% AP50 and 52.2% AP on the full dataset, and 52.9% AP50 and 26.5% AP on the nighttime subset. Compared with YOLOv11-n, the proposed model improves AP50 by 1.2% and AP by 1.7%, while demonstrating more stable performance in challenging low-light environments. These results indicate that the proposed framework provides a practical and reliable solution for real-time fire detection in electric vehicle environments.