Enhancing Fire and Smoke Recognition in Electric Vehicles Through a Lightweight Dual Branch Architecture

Abstract

Fire and smoke detection in electric vehicles (EVs) is crucial for safety; however, existing methods struggle in low-light conditions due to dataset limitations and suboptimal feature extraction. Generation detectors fail to balance real-time efficiency and accuracy, limiting their practical application. This study presents a lightweight dual-branch detection framework for fire and smoke recognition, with a focus on low-light conditions. The proposed model integrates a dual-branch architecture for enhanced feature extraction. The auxiliary branch incorporates the Adaptive Reflectance Enhancement Module (AREM) to reduce flame overexposure and improve contrast, while the Element-wise Adaptive Weight Convolution (EAConv) module refines fire feature representation by dynamically adjusting feature importance. The main branch employs the Self-Adaptive Feature Convolution (SAConv) module, which dynamically adjusts the receptive field to enhance multi-scale feature detection. Additionally, a new public dataset is introduced, containing 3,512 fire and smoke images from EV environments, including 1,910 daytime images and 1,602 nighttime images, addressing dataset gaps in nighttime fire detection. Experiments conducted on the self-constructed dataset demonstrate that the proposed framework outperforms state-of-the-art models, achieving a 1.2% improvement in mean Average Precision (mAP) over baseline approaches. The model attains 85.3% AP50 and 52.2% AP on the full dataset and 52.9% AP50 and 26.5% AP on nighttime fire detection, surpassing YOLOv11 and other contemporary detectors. These results validate the model’s effectiveness in low-light fire detection, demonstrating higher accuracy and improved real-time performance, making it a reliable solution for EV safety monitoring and fire prevention applications. The dataset will be released on https://github.com/Tx1101/FireSmoke-Monitor.

Keywords: Electric Vehicle Safety, Electric vehicles (EVs), Fire Detection, Low-Light Optimization, Self-Adaptive Feature Convolution (SAConv)